Add additional tests for async error cases

Also fixes one bug that was found when an async function throws an error
after being scheduled on a task.

.github/workflows/release-webhook.yml

@@ -0,0 +1,108 @@
+name: Release Webhook
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  send-webhook:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Send release webhook
+        env:
+          WEBHOOK_URL: ${{ secrets.RELEASE_GITHUB_WEBHOOK_URL }}
+          WEBHOOK_SECRET: ${{ secrets.RELEASE_GITHUB_WEBHOOK_SECRET }}
+        run: |
+          # Generate UUID for delivery ID
+          DELIVERY_ID=$(uuidgen)
+          HOOK_ID="release-webhook-$(date +%s)"
+          
+          # Create webhook payload matching GitHub release webhook format
+          PAYLOAD=$(cat <<EOF
+          {
+            "action": "published",
+            "release": {
+              "id": ${{ github.event.release.id }},
+              "node_id": "${{ github.event.release.node_id }}",
+              "url": "${{ github.event.release.url }}",
+              "html_url": "${{ github.event.release.html_url }}",
+              "assets_url": "${{ github.event.release.assets_url }}",
+              "upload_url": "${{ github.event.release.upload_url }}",
+              "tag_name": "${{ github.event.release.tag_name }}",
+              "target_commitish": "${{ github.event.release.target_commitish }}",
+              "name": ${{ toJSON(github.event.release.name) }},
+              "body": ${{ toJSON(github.event.release.body) }},
+              "draft": ${{ github.event.release.draft }},
+              "prerelease": ${{ github.event.release.prerelease }},
+              "created_at": "${{ github.event.release.created_at }}",
+              "published_at": "${{ github.event.release.published_at }}",
+              "author": {
+                "login": "${{ github.event.release.author.login }}",
+                "id": ${{ github.event.release.author.id }},
+                "node_id": "${{ github.event.release.author.node_id }}",
+                "avatar_url": "${{ github.event.release.author.avatar_url }}",
+                "url": "${{ github.event.release.author.url }}",
+                "html_url": "${{ github.event.release.author.html_url }}",
+                "type": "${{ github.event.release.author.type }}",
+                "site_admin": ${{ github.event.release.author.site_admin }}
+              },
+              "tarball_url": "${{ github.event.release.tarball_url }}",
+              "zipball_url": "${{ github.event.release.zipball_url }}",
+              "assets": ${{ toJSON(github.event.release.assets) }}
+            },
+            "repository": {
+              "id": ${{ github.event.repository.id }},
+              "node_id": "${{ github.event.repository.node_id }}",
+              "name": "${{ github.event.repository.name }}",
+              "full_name": "${{ github.event.repository.full_name }}",
+              "private": ${{ github.event.repository.private }},
+              "owner": {
+                "login": "${{ github.event.repository.owner.login }}",
+                "id": ${{ github.event.repository.owner.id }},
+                "node_id": "${{ github.event.repository.owner.node_id }}",
+                "avatar_url": "${{ github.event.repository.owner.avatar_url }}",
+                "url": "${{ github.event.repository.owner.url }}",
+                "html_url": "${{ github.event.repository.owner.html_url }}",
+                "type": "${{ github.event.repository.owner.type }}",
+                "site_admin": ${{ github.event.repository.owner.site_admin }}
+              },
+              "html_url": "${{ github.event.repository.html_url }}",
+              "clone_url": "${{ github.event.repository.clone_url }}",
+              "git_url": "${{ github.event.repository.git_url }}",
+              "ssh_url": "${{ github.event.repository.ssh_url }}",
+              "url": "${{ github.event.repository.url }}",
+              "created_at": "${{ github.event.repository.created_at }}",
+              "updated_at": "${{ github.event.repository.updated_at }}",
+              "pushed_at": "${{ github.event.repository.pushed_at }}",
+              "default_branch": "${{ github.event.repository.default_branch }}",
+              "fork": ${{ github.event.repository.fork }}
+            },
+            "sender": {
+              "login": "${{ github.event.sender.login }}",
+              "id": ${{ github.event.sender.id }},
+              "node_id": "${{ github.event.sender.node_id }}",
+              "avatar_url": "${{ github.event.sender.avatar_url }}",
+              "url": "${{ github.event.sender.url }}",
+              "html_url": "${{ github.event.sender.html_url }}",
+              "type": "${{ github.event.sender.type }}",
+              "site_admin": ${{ github.event.sender.site_admin }}
+            }
+          }
+          EOF
+          )
+          
+          # Generate HMAC-SHA256 signature
+          SIGNATURE=$(echo -n "$PAYLOAD" | openssl dgst -sha256 -hmac "$WEBHOOK_SECRET" -hex | cut -d' ' -f2)
+          
+          # Send webhook with required headers
+          curl -X POST "$WEBHOOK_URL" \
+            -H "Content-Type: application/json" \
+            -H "X-GitHub-Event: release" \
+            -H "X-GitHub-Delivery: $DELIVERY_ID" \
+            -H "X-GitHub-Hook-ID: $HOOK_ID" \
+            -H "X-Hub-Signature-256: sha256=$SIGNATURE" \
+            -H "User-Agent: GitHub-Actions-Webhook/1.0" \
+            -d "$PAYLOAD" \
+            --fail --silent --show-error
+          
+          echo "✅ Release webhook sent successfully"

.github/workflows/stable-release.yml

@@ -102,5 +102,4 @@ jobs:
           file: ComfyUI_windows_portable_nvidia.7z
           tag: ${{ inputs.git_tag }}
           overwrite: true
-          prerelease: true
-          make_latest: false
+          draft: true

.github/workflows/test-unit.yml

@@ -28,3 +28,7 @@ jobs:
       run: |
         pip install -r tests-unit/requirements.txt
         python -m pytest tests-unit
+    - name: Run Execution Model Tests
+      run: |
+        python -m pytest tests/inference/test_execution.py
+

README.md

@@ -65,12 +65,16 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/)
    - [Lumina Image 2.0](https://comfyanonymous.github.io/ComfyUI_examples/lumina2/)
    - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/)
+   - [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
+- Image Editing Models
+   - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
+   - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
 - Video Models
    - [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/)
    - [Mochi](https://comfyanonymous.github.io/ComfyUI_examples/mochi/)
    - [LTX-Video](https://comfyanonymous.github.io/ComfyUI_examples/ltxv/)
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
-   - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/)
+   - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/) and [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
 - Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
@@ -272,6 +276,8 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve
 
 #### DirectML (AMD Cards on Windows)
 
+This is very badly supported and is not recommended. There are some unofficial builds of pytorch ROCm on windows that exist that will give you a much better experience than this. This readme will be updated once official pytorch ROCm builds for windows come out.
+
 ```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```
 
 #### Ascend NPUs

comfy/cli_args.py

@@ -151,6 +151,7 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win
 
 parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
 parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
+parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.")
 parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")

comfy/k_diffusion/sampling.py

@@ -710,6 +710,7 @@ def sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args=None, callback=Non
         # logged_x = torch.cat((logged_x, x.unsqueeze(0)), dim=0)
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     """DPM-Solver++ (stochastic)."""
@@ -721,38 +722,49 @@ def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=N
     seed = extra_args.get("seed", None)
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
     s_in = x.new_ones([x.shape[0]])
-    sigma_fn = lambda t: t.neg().exp()
-    t_fn = lambda sigma: sigma.log().neg()
+
+    model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling')
+    sigma_fn = partial(half_log_snr_to_sigma, model_sampling=model_sampling)
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
 
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
         if sigmas[i + 1] == 0:
-            # Euler method
-            d = to_d(x, sigmas[i], denoised)
-            dt = sigmas[i + 1] - sigmas[i]
-            x = x + d * dt
+            # Denoising step
+            x = denoised
         else:
             # DPM-Solver++
-            t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1])
-            h = t_next - t
-            s = t + h * r
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
+            lambda_s_1 = lambda_s + r * h
             fac = 1 / (2 * r)
 
+            sigma_s_1 = sigma_fn(lambda_s_1)
+
+            alpha_s = sigmas[i] * lambda_s.exp()
+            alpha_s_1 = sigma_s_1 * lambda_s_1.exp()
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
             # Step 1
-            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta)
-            s_ = t_fn(sd)
-            x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised
-            x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su
-            denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
+            sd, su = get_ancestral_step(lambda_s.neg().exp(), lambda_s_1.neg().exp(), eta)
+            lambda_s_1_ = sd.log().neg()
+            h_ = lambda_s_1_ - lambda_s
+            x_2 = (alpha_s_1 / alpha_s) * (-h_).exp() * x - alpha_s_1 * (-h_).expm1() * denoised
+            if eta > 0 and s_noise > 0:
+                x_2 = x_2 + alpha_s_1 * noise_sampler(sigmas[i], sigma_s_1) * s_noise * su
+            denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
             # Step 2
-            sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta)
-            t_next_ = t_fn(sd)
+            sd, su = get_ancestral_step(lambda_s.neg().exp(), lambda_t.neg().exp(), eta)
+            lambda_t_ = sd.log().neg()
+            h_ = lambda_t_ - lambda_s
             denoised_d = (1 - fac) * denoised + fac * denoised_2
-            x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d
-            x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su
+            x = (alpha_t / alpha_s) * (-h_).exp() * x - alpha_t * (-h_).expm1() * denoised_d
+            if eta > 0 and s_noise > 0:
+                x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * su
     return x
 
 
@@ -1435,14 +1447,15 @@ def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None,
         old_d = d
     return x
 
+
 @torch.no_grad()
 def sample_gradient_estimation_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):
     return sample_gradient_estimation(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, ge_gamma=ge_gamma, cfg_pp=True)
 
+
 @torch.no_grad()
-def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3):
-    """
-    Extended Reverse-Time SDE solver (VE ER-SDE-Solver-3). Arxiv: https://arxiv.org/abs/2309.06169.
+def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1.0, noise_sampler=None, noise_scaler=None, max_stage=3):
+    """Extended Reverse-Time SDE solver (VP ER-SDE-Solver-3). arXiv: https://arxiv.org/abs/2309.06169.
     Code reference: https://github.com/QinpengCui/ER-SDE-Solver/blob/main/er_sde_solver.py.
     """
     extra_args = {} if extra_args is None else extra_args
@@ -1450,12 +1463,18 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
     s_in = x.new_ones([x.shape[0]])
 
-    def default_noise_scaler(sigma):
-        return sigma * ((sigma ** 0.3).exp() + 10.0)
-    noise_scaler = default_noise_scaler if noise_scaler is None else noise_scaler
+    def default_er_sde_noise_scaler(x):
+        return x * ((x ** 0.3).exp() + 10.0)
+
+    noise_scaler = default_er_sde_noise_scaler if noise_scaler is None else noise_scaler
     num_integration_points = 200.0
     point_indice = torch.arange(0, num_integration_points, dtype=torch.float32, device=x.device)
 
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+    half_log_snrs = sigma_to_half_log_snr(sigmas, model_sampling)
+    er_lambdas = half_log_snrs.neg().exp()  # er_lambda_t = sigma_t / alpha_t
+
     old_denoised = None
     old_denoised_d = None
 
@@ -1466,32 +1485,36 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
         stage_used = min(max_stage, i + 1)
         if sigmas[i + 1] == 0:
             x = denoised
-        elif stage_used == 1:
-            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
-            x = r * x + (1 - r) * denoised
         else:
-            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
-            x = r * x + (1 - r) * denoised
+            er_lambda_s, er_lambda_t = er_lambdas[i], er_lambdas[i + 1]
+            alpha_s = sigmas[i] / er_lambda_s
+            alpha_t = sigmas[i + 1] / er_lambda_t
+            r_alpha = alpha_t / alpha_s
+            r = noise_scaler(er_lambda_t) / noise_scaler(er_lambda_s)
 
-            dt = sigmas[i + 1] - sigmas[i]
-            sigma_step_size = -dt / num_integration_points
-            sigma_pos = sigmas[i + 1] + point_indice * sigma_step_size
-            scaled_pos = noise_scaler(sigma_pos)
+            # Stage 1 Euler
+            x = r_alpha * r * x + alpha_t * (1 - r) * denoised
 
-            # Stage 2
-            s = torch.sum(1 / scaled_pos) * sigma_step_size
-            denoised_d = (denoised - old_denoised) / (sigmas[i] - sigmas[i - 1])
-            x = x + (dt + s * noise_scaler(sigmas[i + 1])) * denoised_d
+            if stage_used >= 2:
+                dt = er_lambda_t - er_lambda_s
+                lambda_step_size = -dt / num_integration_points
+                lambda_pos = er_lambda_t + point_indice * lambda_step_size
+                scaled_pos = noise_scaler(lambda_pos)
 
-            if stage_used >= 3:
-                # Stage 3
-                s_u = torch.sum((sigma_pos - sigmas[i]) / scaled_pos) * sigma_step_size
-                denoised_u = (denoised_d - old_denoised_d) / ((sigmas[i] - sigmas[i - 2]) / 2)
-                x = x + ((dt ** 2) / 2 + s_u * noise_scaler(sigmas[i + 1])) * denoised_u
-            old_denoised_d = denoised_d
+                # Stage 2
+                s = torch.sum(1 / scaled_pos) * lambda_step_size
+                denoised_d = (denoised - old_denoised) / (er_lambda_s - er_lambdas[i - 1])
+                x = x + alpha_t * (dt + s * noise_scaler(er_lambda_t)) * denoised_d
 
-        if s_noise != 0 and sigmas[i + 1] > 0:
-            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * (sigmas[i + 1] ** 2 - sigmas[i] ** 2 * r ** 2).sqrt().nan_to_num(nan=0.0)
+                if stage_used >= 3:
+                    # Stage 3
+                    s_u = torch.sum((lambda_pos - er_lambda_s) / scaled_pos) * lambda_step_size
+                    denoised_u = (denoised_d - old_denoised_d) / ((er_lambda_s - er_lambdas[i - 2]) / 2)
+                    x = x + alpha_t * ((dt ** 2) / 2 + s_u * noise_scaler(er_lambda_t)) * denoised_u
+                old_denoised_d = denoised_d
+
+            if s_noise > 0:
+                x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * (er_lambda_t ** 2 - er_lambda_s ** 2 * r ** 2).sqrt().nan_to_num(nan=0.0)
         old_denoised = denoised
     return x
 

comfy/ldm/flux/controlnet.py

@@ -123,6 +123,8 @@ class ControlNetFlux(Flux):
 
         if y is None:
             y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+        else:
+            y = y[:, :self.params.vec_in_dim]
 
         # running on sequences img
         img = self.img_in(img)

comfy/ldm/flux/layers.py

@@ -118,7 +118,7 @@ class Modulation(nn.Module):
 def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
     if modulation_dims is None:
         if m_add is not None:
-            return tensor * m_mult + m_add
+            return torch.addcmul(m_add, tensor, m_mult)
         else:
             return tensor * m_mult
     else:

comfy/ldm/flux/model.py

@@ -195,20 +195,50 @@ class Flux(nn.Module):
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def forward(self, x, timestep, context, y=None, guidance=None, control=None, transformer_options={}, **kwargs):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
         bs, c, h, w = x.shape
         patch_size = self.patch_size
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
 
         img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
-
         h_len = ((h + (patch_size // 2)) // patch_size)
         w_len = ((w + (patch_size // 2)) // patch_size)
+
+        h_offset = ((h_offset + (patch_size // 2)) // patch_size)
+        w_offset = ((w_offset + (patch_size // 2)) // patch_size)
+
         img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
-        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        return img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
+        bs, c, h_orig, w_orig = x.shape
+        patch_size = self.patch_size
+
+        h_len = ((h_orig + (patch_size // 2)) // patch_size)
+        w_len = ((w_orig + (patch_size // 2)) // patch_size)
+        img, img_ids = self.process_img(x)
+        img_tokens = img.shape[1]
+        if ref_latents is not None:
+            h = 0
+            w = 0
+            for ref in ref_latents:
+                h_offset = 0
+                w_offset = 0
+                if ref.shape[-2] + h > ref.shape[-1] + w:
+                    w_offset = w
+                else:
+                    h_offset = h
+
+                kontext, kontext_ids = self.process_img(ref, index=1, h_offset=h_offset, w_offset=w_offset)
+                img = torch.cat([img, kontext], dim=1)
+                img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+                h = max(h, ref.shape[-2] + h_offset)
+                w = max(w, ref.shape[-1] + w_offset)
 
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
-        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]
+        out = out[:, :img_tokens]
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h_orig,:w_orig]

comfy/ldm/models/autoencoder.py

@@ -11,7 +11,7 @@ from comfy.ldm.modules.ema import LitEma
 import comfy.ops
 
 class DiagonalGaussianRegularizer(torch.nn.Module):
-    def __init__(self, sample: bool = True):
+    def __init__(self, sample: bool = False):
         super().__init__()
         self.sample = sample
 
@@ -19,16 +19,12 @@ class DiagonalGaussianRegularizer(torch.nn.Module):
         yield from ()
 
     def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
-        log = dict()
         posterior = DiagonalGaussianDistribution(z)
         if self.sample:
             z = posterior.sample()
         else:
             z = posterior.mode()
-        kl_loss = posterior.kl()
-        kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
-        log["kl_loss"] = kl_loss
-        return z, log
+        return z, None
 
 
 class AbstractAutoencoder(torch.nn.Module):

comfy/ldm/modules/sub_quadratic_attention.py

@@ -31,7 +31,7 @@ def dynamic_slice(
     starts: List[int],
     sizes: List[int],
 ) -> Tensor:
-    slicing = [slice(start, start + size) for start, size in zip(starts, sizes)]
+    slicing = tuple(slice(start, start + size) for start, size in zip(starts, sizes))
     return x[slicing]
 
 class AttnChunk(NamedTuple):

comfy/ldm/omnigen/omnigen2.py

@@ -0,0 +1,469 @@
+# Original code: https://github.com/VectorSpaceLab/OmniGen2
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from comfy.ldm.lightricks.model import Timesteps
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.modules.attention import optimized_attention_masked
+import comfy.model_management
+import comfy.ldm.common_dit
+
+
+def apply_rotary_emb(x, freqs_cis):
+    if x.shape[1] == 0:
+        return x
+
+    t_ = x.reshape(*x.shape[:-1], -1, 1, 2)
+    t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
+    return t_out.reshape(*x.shape).to(dtype=x.dtype)
+
+
+def swiglu(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+    return F.silu(x) * y
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels: int, time_embed_dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear_1 = operations.Linear(in_channels, time_embed_dim, dtype=dtype, device=device)
+        self.act = nn.SiLU()
+        self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+
+class LuminaRMSNormZero(nn.Module):
+    def __init__(self, embedding_dim: int, norm_eps: float = 1e-5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(min(embedding_dim, 1024), 4 * embedding_dim, dtype=dtype, device=device)
+        self.norm = operations.RMSNorm(embedding_dim, eps=norm_eps, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor, emb: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)
+        x = self.norm(x) * (1 + scale_msa[:, None])
+        return x, gate_msa, scale_mlp, gate_mlp
+
+
+class LuminaLayerNormContinuous(nn.Module):
+    def __init__(self, embedding_dim: int, conditioning_embedding_dim: int, elementwise_affine: bool = False, eps: float = 1e-6, out_dim: Optional[int] = None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear_1 = operations.Linear(conditioning_embedding_dim, embedding_dim, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(embedding_dim, eps, elementwise_affine, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(embedding_dim, out_dim, bias=True, dtype=dtype, device=device) if out_dim is not None else None
+
+    def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
+        emb = self.linear_1(self.silu(conditioning_embedding).to(x.dtype))
+        x = self.norm(x) * (1 + emb)[:, None, :]
+        if self.linear_2 is not None:
+            x = self.linear_2(x)
+        return x
+
+
+class LuminaFeedForward(nn.Module):
+    def __init__(self, dim: int, inner_dim: int, multiple_of: int = 256, dtype=None, device=None, operations=None):
+        super().__init__()
+        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)
+        self.linear_1 = operations.Linear(dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(inner_dim, dim, bias=False, dtype=dtype, device=device)
+        self.linear_3 = operations.Linear(dim, inner_dim, bias=False, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h1, h2 = self.linear_1(x), self.linear_3(x)
+        return self.linear_2(swiglu(h1, h2))
+
+
+class Lumina2CombinedTimestepCaptionEmbedding(nn.Module):
+    def __init__(self, hidden_size: int = 4096, text_feat_dim: int = 2048, frequency_embedding_size: int = 256, norm_eps: float = 1e-5, timestep_scale: float = 1.0, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0.0, scale=timestep_scale)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=min(hidden_size, 1024), dtype=dtype, device=device, operations=operations)
+        self.caption_embedder = nn.Sequential(
+            operations.RMSNorm(text_feat_dim, eps=norm_eps, dtype=dtype, device=device),
+            operations.Linear(text_feat_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+    def forward(self, timestep: torch.Tensor, text_hidden_states: torch.Tensor, dtype: torch.dtype) -> Tuple[torch.Tensor, torch.Tensor]:
+        timestep_proj = self.time_proj(timestep).to(dtype=dtype)
+        time_embed = self.timestep_embedder(timestep_proj)
+        caption_embed = self.caption_embedder(text_hidden_states)
+        return time_embed, caption_embed
+
+
+class Attention(nn.Module):
+    def __init__(self, query_dim: int, dim_head: int, heads: int, kv_heads: int, eps: float = 1e-5, bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.heads = heads
+        self.kv_heads = kv_heads
+        self.dim_head = dim_head
+        self.scale = dim_head ** -0.5
+
+        self.to_q = operations.Linear(query_dim, heads * dim_head, bias=bias, dtype=dtype, device=device)
+        self.to_k = operations.Linear(query_dim, kv_heads * dim_head, bias=bias, dtype=dtype, device=device)
+        self.to_v = operations.Linear(query_dim, kv_heads * dim_head, bias=bias, dtype=dtype, device=device)
+
+        self.norm_q = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+        self.norm_k = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+
+        self.to_out = nn.Sequential(
+            operations.Linear(heads * dim_head, query_dim, bias=bias, dtype=dtype, device=device),
+            nn.Dropout(0.0)
+        )
+
+    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, image_rotary_emb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        query = self.to_q(hidden_states)
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        query = query.view(batch_size, -1, self.heads, self.dim_head)
+        key = key.view(batch_size, -1, self.kv_heads, self.dim_head)
+        value = value.view(batch_size, -1, self.kv_heads, self.dim_head)
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        if self.kv_heads < self.heads:
+            key = key.repeat_interleave(self.heads // self.kv_heads, dim=1)
+            value = value.repeat_interleave(self.heads // self.kv_heads, dim=1)
+
+        hidden_states = optimized_attention_masked(query, key, value, self.heads, attention_mask, skip_reshape=True)
+        hidden_states = self.to_out[0](hidden_states)
+        return hidden_states
+
+
+class OmniGen2TransformerBlock(nn.Module):
+    def __init__(self, dim: int, num_attention_heads: int, num_kv_heads: int, multiple_of: int, ffn_dim_multiplier: float, norm_eps: float, modulation: bool = True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.modulation = modulation
+
+        self.attn = Attention(
+            query_dim=dim,
+            dim_head=dim // num_attention_heads,
+            heads=num_attention_heads,
+            kv_heads=num_kv_heads,
+            eps=1e-5,
+            bias=False,
+            dtype=dtype, device=device, operations=operations,
+        )
+
+        self.feed_forward = LuminaFeedForward(
+            dim=dim,
+            inner_dim=4 * dim,
+            multiple_of=multiple_of,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        if modulation:
+            self.norm1 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations)
+        else:
+            self.norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+
+        self.ffn_norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+        self.norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+        self.ffn_norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, image_rotary_emb: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        if self.modulation:
+            norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            hidden_states = hidden_states + gate_msa.unsqueeze(1).tanh() * self.norm2(attn_output)
+            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states) * (1 + scale_mlp.unsqueeze(1)))
+            hidden_states = hidden_states + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(mlp_output)
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            hidden_states = hidden_states + self.norm2(attn_output)
+            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states))
+            hidden_states = hidden_states + self.ffn_norm2(mlp_output)
+        return hidden_states
+
+
+class OmniGen2RotaryPosEmbed(nn.Module):
+    def __init__(self, theta: int, axes_dim: Tuple[int, int, int], axes_lens: Tuple[int, int, int] = (300, 512, 512), patch_size: int = 2):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        self.axes_lens = axes_lens
+        self.patch_size = patch_size
+        self.rope_embedder = EmbedND(dim=sum(axes_dim), theta=self.theta, axes_dim=axes_dim)
+
+    def forward(self, batch_size, encoder_seq_len, l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len, ref_img_sizes, img_sizes, device):
+        p = self.patch_size
+
+        seq_lengths = [cap_len + sum(ref_img_len) + img_len for cap_len, ref_img_len, img_len in zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len)]
+
+        max_seq_len = max(seq_lengths)
+        max_ref_img_len = max([sum(ref_img_len) for ref_img_len in l_effective_ref_img_len])
+        max_img_len = max(l_effective_img_len)
+
+        position_ids = torch.zeros(batch_size, max_seq_len, 3, dtype=torch.int32, device=device)
+
+        for i, (cap_seq_len, seq_len) in enumerate(zip(l_effective_cap_len, seq_lengths)):
+            position_ids[i, :cap_seq_len] = repeat(torch.arange(cap_seq_len, dtype=torch.int32, device=device), "l -> l 3")
+
+            pe_shift = cap_seq_len
+            pe_shift_len = cap_seq_len
+
+            if ref_img_sizes[i] is not None:
+                for ref_img_size, ref_img_len in zip(ref_img_sizes[i], l_effective_ref_img_len[i]):
+                    H, W = ref_img_size
+                    ref_H_tokens, ref_W_tokens = H // p, W // p
+
+                    row_ids = repeat(torch.arange(ref_H_tokens, dtype=torch.int32, device=device), "h -> h w", w=ref_W_tokens).flatten()
+                    col_ids = repeat(torch.arange(ref_W_tokens, dtype=torch.int32, device=device), "w -> h w", h=ref_H_tokens).flatten()
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 0] = pe_shift
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 1] = row_ids
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 2] = col_ids
+
+                    pe_shift += max(ref_H_tokens, ref_W_tokens)
+                    pe_shift_len += ref_img_len
+
+            H, W = img_sizes[i]
+            H_tokens, W_tokens = H // p, W // p
+
+            row_ids = repeat(torch.arange(H_tokens, dtype=torch.int32, device=device), "h -> h w", w=W_tokens).flatten()
+            col_ids = repeat(torch.arange(W_tokens, dtype=torch.int32, device=device), "w -> h w", h=H_tokens).flatten()
+
+            position_ids[i, pe_shift_len: seq_len, 0] = pe_shift
+            position_ids[i, pe_shift_len: seq_len, 1] = row_ids
+            position_ids[i, pe_shift_len: seq_len, 2] = col_ids
+
+        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2)
+
+        cap_freqs_cis_shape = list(freqs_cis.shape)
+        cap_freqs_cis_shape[1] = encoder_seq_len
+        cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        ref_img_freqs_cis_shape = list(freqs_cis.shape)
+        ref_img_freqs_cis_shape[1] = max_ref_img_len
+        ref_img_freqs_cis = torch.zeros(*ref_img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        img_freqs_cis_shape = list(freqs_cis.shape)
+        img_freqs_cis_shape[1] = max_img_len
+        img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        for i, (cap_seq_len, ref_img_len, img_len, seq_len) in enumerate(zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len, seq_lengths)):
+            cap_freqs_cis[i, :cap_seq_len] = freqs_cis[i, :cap_seq_len]
+            ref_img_freqs_cis[i, :sum(ref_img_len)] = freqs_cis[i, cap_seq_len:cap_seq_len + sum(ref_img_len)]
+            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_seq_len + sum(ref_img_len):cap_seq_len + sum(ref_img_len) + img_len]
+
+        return cap_freqs_cis, ref_img_freqs_cis, img_freqs_cis, freqs_cis, l_effective_cap_len, seq_lengths
+
+
+class OmniGen2Transformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        out_channels: Optional[int] = None,
+        hidden_size: int = 2304,
+        num_layers: int = 26,
+        num_refiner_layers: int = 2,
+        num_attention_heads: int = 24,
+        num_kv_heads: int = 8,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        norm_eps: float = 1e-5,
+        axes_dim_rope: Tuple[int, int, int] = (32, 32, 32),
+        axes_lens: Tuple[int, int, int] = (300, 512, 512),
+        text_feat_dim: int = 1024,
+        timestep_scale: float = 1.0,
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+
+        self.patch_size = patch_size
+        self.out_channels = out_channels or in_channels
+        self.hidden_size = hidden_size
+        self.dtype = dtype
+
+        self.rope_embedder = OmniGen2RotaryPosEmbed(
+            theta=10000,
+            axes_dim=axes_dim_rope,
+            axes_lens=axes_lens,
+            patch_size=patch_size,
+        )
+
+        self.x_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device)
+        self.ref_image_patch_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device)
+
+        self.time_caption_embed = Lumina2CombinedTimestepCaptionEmbedding(
+            hidden_size=hidden_size,
+            text_feat_dim=text_feat_dim,
+            norm_eps=norm_eps,
+            timestep_scale=timestep_scale, dtype=dtype, device=device, operations=operations
+        )
+
+        self.noise_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.ref_image_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.context_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=False, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.layers = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_layers)
+        ])
+
+        self.norm_out = LuminaLayerNormContinuous(
+            embedding_dim=hidden_size,
+            conditioning_embedding_dim=min(hidden_size, 1024),
+            elementwise_affine=False,
+            eps=1e-6,
+            out_dim=patch_size * patch_size * self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+        self.image_index_embedding = nn.Parameter(torch.empty(5, hidden_size, device=device, dtype=dtype))
+
+    def flat_and_pad_to_seq(self, hidden_states, ref_image_hidden_states):
+        batch_size = len(hidden_states)
+        p = self.patch_size
+
+        img_sizes = [(img.size(1), img.size(2)) for img in hidden_states]
+        l_effective_img_len = [(H // p) * (W // p) for (H, W) in img_sizes]
+
+        if ref_image_hidden_states is not None:
+            ref_image_hidden_states = list(map(lambda ref: comfy.ldm.common_dit.pad_to_patch_size(ref, (p, p)), ref_image_hidden_states))
+            ref_img_sizes = [[(imgs.size(2), imgs.size(3)) if imgs is not None else None for imgs in ref_image_hidden_states]] * batch_size
+            l_effective_ref_img_len = [[(ref_img_size[0] // p) * (ref_img_size[1] // p) for ref_img_size in _ref_img_sizes] if _ref_img_sizes is not None else [0] for _ref_img_sizes in ref_img_sizes]
+        else:
+            ref_img_sizes = [None for _ in range(batch_size)]
+            l_effective_ref_img_len = [[0] for _ in range(batch_size)]
+
+        flat_ref_img_hidden_states = None
+        if ref_image_hidden_states is not None:
+            imgs = []
+            for ref_img in ref_image_hidden_states:
+                B, C, H, W = ref_img.size()
+                ref_img = rearrange(ref_img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
+                imgs.append(ref_img)
+            flat_ref_img_hidden_states = torch.cat(imgs, dim=1)
+
+        img = hidden_states
+        B, C, H, W = img.size()
+        flat_hidden_states = rearrange(img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
+
+        return (
+            flat_hidden_states, flat_ref_img_hidden_states,
+            None, None,
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes,
+        )
+
+    def img_patch_embed_and_refine(self, hidden_states, ref_image_hidden_states, padded_img_mask, padded_ref_img_mask, noise_rotary_emb, ref_img_rotary_emb, l_effective_ref_img_len, l_effective_img_len, temb):
+        batch_size = len(hidden_states)
+
+        hidden_states = self.x_embedder(hidden_states)
+        if ref_image_hidden_states is not None:
+            ref_image_hidden_states = self.ref_image_patch_embedder(ref_image_hidden_states)
+            image_index_embedding = comfy.model_management.cast_to(self.image_index_embedding, dtype=hidden_states.dtype, device=hidden_states.device)
+
+            for i in range(batch_size):
+                shift = 0
+                for j, ref_img_len in enumerate(l_effective_ref_img_len[i]):
+                    ref_image_hidden_states[i, shift:shift + ref_img_len, :] = ref_image_hidden_states[i, shift:shift + ref_img_len, :] + image_index_embedding[j]
+                    shift += ref_img_len
+
+        for layer in self.noise_refiner:
+            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)
+
+        if ref_image_hidden_states is not None:
+            for layer in self.ref_image_refiner:
+                ref_image_hidden_states = layer(ref_image_hidden_states, padded_ref_img_mask, ref_img_rotary_emb, temb)
+
+            hidden_states = torch.cat([ref_image_hidden_states, hidden_states], dim=1)
+
+        return hidden_states
+
+    def forward(self, x, timesteps, context, num_tokens, ref_latents=None, attention_mask=None, **kwargs):
+        B, C, H, W = x.shape
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        _, _, H_padded, W_padded = hidden_states.shape
+        timestep = 1.0 - timesteps
+        text_hidden_states = context
+        text_attention_mask = attention_mask
+        ref_image_hidden_states = ref_latents
+        device = hidden_states.device
+
+        temb, text_hidden_states = self.time_caption_embed(timestep, text_hidden_states, hidden_states[0].dtype)
+
+        (
+            hidden_states, ref_image_hidden_states,
+            img_mask, ref_img_mask,
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes,
+        ) = self.flat_and_pad_to_seq(hidden_states, ref_image_hidden_states)
+
+        (
+            context_rotary_emb, ref_img_rotary_emb, noise_rotary_emb,
+            rotary_emb, encoder_seq_lengths, seq_lengths,
+        ) = self.rope_embedder(
+            hidden_states.shape[0], text_hidden_states.shape[1], [num_tokens] * text_hidden_states.shape[0],
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes, device,
+        )
+
+        for layer in self.context_refiner:
+            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb)
+
+        img_len = hidden_states.shape[1]
+        combined_img_hidden_states = self.img_patch_embed_and_refine(
+            hidden_states, ref_image_hidden_states,
+            img_mask, ref_img_mask,
+            noise_rotary_emb, ref_img_rotary_emb,
+            l_effective_ref_img_len, l_effective_img_len,
+            temb,
+        )
+
+        hidden_states = torch.cat([text_hidden_states, combined_img_hidden_states], dim=1)
+        attention_mask = None
+
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb)
+
+        hidden_states = self.norm_out(hidden_states, temb)
+
+        p = self.patch_size
+        output = rearrange(hidden_states[:, -img_len:], 'b (h w) (p1 p2 c) -> b c (h p1) (w p2)',  h=H_padded // p, w=W_padded// p, p1=p, p2=p)[:, :, :H, :W]
+
+        return -output

comfy/model_base.py

@@ -41,6 +41,7 @@ import comfy.ldm.hunyuan3d.model
 import comfy.ldm.hidream.model
 import comfy.ldm.chroma.model
 import comfy.ldm.ace.model
+import comfy.ldm.omnigen.omnigen2
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -815,6 +816,7 @@ class PixArt(BaseModel):
 class Flux(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.flux.model.Flux):
         super().__init__(model_config, model_type, device=device, unet_model=unet_model)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def concat_cond(self, **kwargs):
         try:
@@ -875,8 +877,23 @@ class Flux(BaseModel):
         guidance = kwargs.get("guidance", 3.5)
         if guidance is not None:
             out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
         return out
 
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+        return out
+
+
 class GenmoMochi(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.genmo.joint_model.asymm_models_joint.AsymmDiTJoint)
@@ -1024,6 +1041,8 @@ class CosmosPredict2(BaseModel):
     def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
         if denoise_mask is None:
             return timestep
+        if denoise_mask.ndim <= 4:
+            return timestep
         condition_video_mask_B_1_T_1_1 = denoise_mask.mean(dim=[1, 3, 4], keepdim=True)
         c_noise_B_1_T_1_1 = 0.0 * (1.0 - condition_video_mask_B_1_T_1_1) + timestep.reshape(timestep.shape[0], 1, 1, 1, 1) * condition_video_mask_B_1_T_1_1
         out = c_noise_B_1_T_1_1.squeeze(dim=[1, 3, 4])
@@ -1228,3 +1247,33 @@ class ACEStep(BaseModel):
         out['speaker_embeds'] = comfy.conds.CONDRegular(torch.zeros(noise.shape[0], 512, device=noise.device, dtype=noise.dtype))
         out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
         return out
+
+class Omnigen2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)
+        self.memory_usage_factor_conds = ("ref_latents",)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            if torch.numel(attention_mask) != attention_mask.sum():
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+        return out

comfy/model_detection.py

@@ -459,6 +459,26 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         return dit_config
 
+    if '{}time_caption_embed.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys:  # Omnigen2
+        dit_config = {}
+        dit_config["image_model"] = "omnigen2"
+        dit_config["axes_dim_rope"] = [40, 40, 40]
+        dit_config["axes_lens"] = [1024, 1664, 1664]
+        dit_config["ffn_dim_multiplier"] = None
+        dit_config["hidden_size"] = 2520
+        dit_config["in_channels"] = 16
+        dit_config["multiple_of"] = 256
+        dit_config["norm_eps"] = 1e-05
+        dit_config["num_attention_heads"] = 21
+        dit_config["num_kv_heads"] = 7
+        dit_config["num_layers"] = 32
+        dit_config["num_refiner_layers"] = 2
+        dit_config["out_channels"] = None
+        dit_config["patch_size"] = 2
+        dit_config["text_feat_dim"] = 2048
+        dit_config["timestep_scale"] = 1000.0
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/model_management.py

@@ -1290,6 +1290,13 @@ def supports_fp8_compute(device=None):
 
     return True
 
+def extended_fp16_support():
+    # TODO: check why some models work with fp16 on newer torch versions but not on older
+    if torch_version_numeric < (2, 7):
+        return False
+
+    return True
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:

comfy/samplers.py

@@ -1039,13 +1039,13 @@ class SchedulerHandler(NamedTuple):
     use_ms: bool = True
 
 SCHEDULER_HANDLERS = {
-    "normal": SchedulerHandler(normal_scheduler),
+    "simple": SchedulerHandler(simple_scheduler),
+    "sgm_uniform": SchedulerHandler(partial(normal_scheduler, sgm=True)),
     "karras": SchedulerHandler(k_diffusion_sampling.get_sigmas_karras, use_ms=False),
     "exponential": SchedulerHandler(k_diffusion_sampling.get_sigmas_exponential, use_ms=False),
-    "sgm_uniform": SchedulerHandler(partial(normal_scheduler, sgm=True)),
-    "simple": SchedulerHandler(simple_scheduler),
     "ddim_uniform": SchedulerHandler(ddim_scheduler),
     "beta": SchedulerHandler(beta_scheduler),
+    "normal": SchedulerHandler(normal_scheduler),
     "linear_quadratic": SchedulerHandler(linear_quadratic_schedule),
     "kl_optimal": SchedulerHandler(kl_optimal_scheduler, use_ms=False),
 }

comfy/sd.py

@@ -44,6 +44,7 @@ import comfy.text_encoders.lumina2
 import comfy.text_encoders.wan
 import comfy.text_encoders.hidream
 import comfy.text_encoders.ace
+import comfy.text_encoders.omnigen2
 
 import comfy.model_patcher
 import comfy.lora
@@ -754,6 +755,7 @@ class CLIPType(Enum):
     HIDREAM = 14
     CHROMA = 15
     ACE = 16
+    OMNIGEN2 = 17
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -773,6 +775,7 @@ class TEModel(Enum):
     LLAMA3_8 = 7
     T5_XXL_OLD = 8
     GEMMA_2_2B = 9
+    QWEN25_3B = 10
 
 def detect_te_model(sd):
     if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -793,6 +796,8 @@ def detect_te_model(sd):
         return TEModel.T5_BASE
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
         return TEModel.GEMMA_2_2B
+    if 'model.layers.0.self_attn.k_proj.bias' in sd:
+        return TEModel.QWEN25_3B
     if "model.layers.0.post_attention_layernorm.weight" in sd:
         return TEModel.LLAMA3_8
     return None
@@ -894,6 +899,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data),
                                                                         clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None)
             clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
+        elif te_model == TEModel.QWEN25_3B:
+            clip_target.clip = comfy.text_encoders.omnigen2.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.omnigen2.Omnigen2Tokenizer
         else:
             # clip_l
             if clip_type == CLIPType.SD3:
@@ -1160,7 +1168,7 @@ def load_diffusion_model_state_dict(sd, model_options={}):
     model.load_model_weights(new_sd, "")
     left_over = sd.keys()
     if len(left_over) > 0:
-        logging.info("left over keys in unet: {}".format(left_over))
+        logging.info("left over keys in diffusion model: {}".format(left_over))
     return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
 
 
@@ -1168,7 +1176,7 @@ def load_diffusion_model(unet_path, model_options={}):
     sd = comfy.utils.load_torch_file(unet_path)
     model = load_diffusion_model_state_dict(sd, model_options=model_options)
     if model is None:
-        logging.error("ERROR UNSUPPORTED UNET {}".format(unet_path))
+        logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".format(unet_path))
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
     return model
 

comfy/sd1_clip.py

@@ -482,7 +482,8 @@ class SDTokenizer:
             if end_token is not None:
                 self.end_token = end_token
             else:
-                self.end_token = empty[0]
+                if has_end_token:
+                    self.end_token = empty[0]
 
         if pad_token is not None:
             self.pad_token = pad_token

comfy/supported_models.py

@@ -18,6 +18,7 @@ import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
 import comfy.text_encoders.wan
 import comfy.text_encoders.ace
+import comfy.text_encoders.omnigen2
 
 from . import supported_models_base
 from . import latent_formats
@@ -1181,6 +1182,41 @@ class ACEStep(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         return supported_models_base.ClipTarget(comfy.text_encoders.ace.AceT5Tokenizer, comfy.text_encoders.ace.AceT5Model)
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep]
+class Omnigen2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "omnigen2",
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 2.6,
+    }
+
+    memory_usage_factor = 1.65 #TODO
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        if comfy.model_management.extended_fp16_support():
+            self.supported_inference_dtypes = [torch.float16] + self.supported_inference_dtypes
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Omnigen2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_3b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.LuminaTokenizer, comfy.text_encoders.omnigen2.te(**hunyuan_detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2]
 
 models += [SVD_img2vid]

comfy/text_encoders/llama.py

@@ -24,6 +24,24 @@ class Llama2Config:
     head_dim = 128
     rms_norm_add = False
     mlp_activation = "silu"
+    qkv_bias = False
+
+@dataclass
+class Qwen25_3BConfig:
+    vocab_size: int = 151936
+    hidden_size: int = 2048
+    intermediate_size: int = 11008
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 2
+    max_position_embeddings: int = 128000
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = True
 
 @dataclass
 class Gemma2_2B_Config:
@@ -40,6 +58,7 @@ class Gemma2_2B_Config:
     head_dim = 256
     rms_norm_add = True
     mlp_activation = "gelu_pytorch_tanh"
+    qkv_bias = False
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -98,9 +117,9 @@ class Attention(nn.Module):
         self.inner_size = self.num_heads * self.head_dim
 
         ops = ops or nn
-        self.q_proj = ops.Linear(config.hidden_size, self.inner_size, bias=False, device=device, dtype=dtype)
-        self.k_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False, device=device, dtype=dtype)
-        self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False, device=device, dtype=dtype)
+        self.q_proj = ops.Linear(config.hidden_size, self.inner_size, bias=config.qkv_bias, device=device, dtype=dtype)
+        self.k_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=config.qkv_bias, device=device, dtype=dtype)
+        self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=config.qkv_bias, device=device, dtype=dtype)
         self.o_proj = ops.Linear(self.inner_size, config.hidden_size, bias=False, device=device, dtype=dtype)
 
     def forward(
@@ -320,6 +339,14 @@ class Llama2(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen25_3B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen25_3BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
 
 class Gemma2_2B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):

comfy/text_encoders/omnigen2.py

@@ -0,0 +1,44 @@
+from transformers import Qwen2Tokenizer
+from comfy import sd1_clip
+import comfy.text_encoders.llama
+import os
+
+
+class Qwen25_3BTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2048, embedding_key='qwen25_3b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+
+
+class Omnigen2Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_3b", tokenizer=Qwen25_3BTokenizer)
+        self.llama_template = '<|im_start|>system\nYou are a helpful assistant that generates high-quality images based on user instructions.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n'
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None,**kwargs):
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+        return super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+
+class Qwen25_3BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen25_3B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class Omnigen2Model(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="qwen25_3b", clip_model=Qwen25_3BModel, model_options=model_options)
+
+
+def te(dtype_llama=None, llama_scaled_fp8=None):
+    class Omnigen2TEModel_(Omnigen2Model):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return Omnigen2TEModel_

comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json

@@ -0,0 +1,241 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<|img|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151666": {
+      "content": "<|endofimg|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151667": {
+      "content": "<|meta|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151668": {
+      "content": "<|endofmeta|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- messages[0]['content'] }}\n    {%- else %}\n        {{- 'You are a helpful assistant.' }}\n    {%- endif %}\n    {{- \"\\n\\n# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0]['role'] == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0]['content'] + '<|im_end|>\\n' }}\n    {%- else %}\n        {{- '<|im_start|>system\\nYou are a helpful assistant.<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) or (message.role == \"assistant\" and not message.tool_calls) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {{- '<|im_start|>' + message.role }}\n        {%- if message.content %}\n            {{- '\\n' + message.content }}\n        {%- endif %}\n        {%- for tool_call in message.tool_calls %}\n            {%- if tool_call.function is defined %}\n                {%- set tool_call = tool_call.function %}\n            {%- endif %}\n            {{- '\\n<tool_call>\\n{\"name\": \"' }}\n            {{- tool_call.name }}\n            {{- '\", \"arguments\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- '}\\n</tool_call>' }}\n        {%- endfor %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n{%- endif %}\n",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen2_5_VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

comfy/text_encoders/t5.py

@@ -146,7 +146,7 @@ class T5Attention(torch.nn.Module):
         )
         values = self.relative_attention_bias(relative_position_bucket, out_dtype=dtype)  # shape (query_length, key_length, num_heads)
         values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, query_length, key_length)
-        return values
+        return values.contiguous()
 
     def forward(self, x, mask=None, past_bias=None, optimized_attention=None):
         q = self.q(x)

comfy/utils.py

@@ -997,11 +997,12 @@ def set_progress_bar_global_hook(function):
     PROGRESS_BAR_HOOK = function
 
 class ProgressBar:
-    def __init__(self, total):
+    def __init__(self, total, node_id=None):
         global PROGRESS_BAR_HOOK
         self.total = total
         self.current = 0
         self.hook = PROGRESS_BAR_HOOK
+        self.node_id = node_id
 
     def update_absolute(self, value, total=None, preview=None):
         if total is not None:
@@ -1010,7 +1011,7 @@ class ProgressBar:
             value = self.total
         self.current = value
         if self.hook is not None:
-            self.hook(self.current, self.total, preview)
+            self.hook(self.current, self.total, preview, node_id=self.node_id)
 
     def update(self, value):
         self.update_absolute(self.current + value)

comfy_config/config_parser.py

@@ -11,6 +11,43 @@ from comfy_config.types import (
     PyProjectSettings
 )
 
+def validate_and_extract_os_classifiers(classifiers: list) -> list:
+    os_classifiers = [c for c in classifiers if c.startswith("Operating System :: ")]
+    if not os_classifiers:
+        return []
+
+    os_values = [c[len("Operating System :: ") :] for c in os_classifiers]
+    valid_os_prefixes = {"Microsoft", "POSIX", "MacOS", "OS Independent"}
+
+    for os_value in os_values:
+        if not any(os_value.startswith(prefix) for prefix in valid_os_prefixes):
+            return []
+
+    return os_values
+
+
+def validate_and_extract_accelerator_classifiers(classifiers: list) -> list:
+    accelerator_classifiers = [c for c in classifiers if c.startswith("Environment ::")]
+    if not accelerator_classifiers:
+        return []
+
+    accelerator_values = [c[len("Environment :: ") :] for c in accelerator_classifiers]
+
+    valid_accelerators = {
+        "GPU :: NVIDIA CUDA",
+        "GPU :: AMD ROCm",
+        "GPU :: Intel Arc",
+        "NPU :: Huawei Ascend",
+        "GPU :: Apple Metal",
+    }
+
+    for accelerator_value in accelerator_values:
+        if accelerator_value not in valid_accelerators:
+            return []
+
+    return accelerator_values
+
+
 """
 Extract configuration from a custom node directory's pyproject.toml file or a Python file.
 
@@ -78,6 +115,24 @@ def extract_node_configuration(path) -> Optional[PyProjectConfig]:
     tool_data = raw_settings.tool
     comfy_data = tool_data.get("comfy", {}) if tool_data else {}
 
+    dependencies = project_data.get("dependencies", [])
+    supported_comfyui_frontend_version = ""
+    for dep in dependencies:
+        if isinstance(dep, str) and dep.startswith("comfyui-frontend-package"):
+            supported_comfyui_frontend_version = dep.removeprefix("comfyui-frontend-package")
+            break
+
+    supported_comfyui_version = comfy_data.get("requires-comfyui", "")
+
+    classifiers = project_data.get('classifiers', [])
+    supported_os = validate_and_extract_os_classifiers(classifiers)
+    supported_accelerators = validate_and_extract_accelerator_classifiers(classifiers)
+
+    project_data['supported_os'] = supported_os
+    project_data['supported_accelerators'] = supported_accelerators
+    project_data['supported_comfyui_frontend_version'] = supported_comfyui_frontend_version
+    project_data['supported_comfyui_version'] = supported_comfyui_version
+
     return PyProjectConfig(project=project_data, tool_comfy=comfy_data)
 
 

comfy_config/types.py

@@ -51,7 +51,7 @@ class ComfyConfig(BaseModel):
     models: List[Model] = Field(default_factory=list, alias="Models")
     includes: List[str] = Field(default_factory=list)
     web: Optional[str] = None
-
+    banner_url: str = ""
 
 class License(BaseModel):
     file: str = ""
@@ -66,6 +66,10 @@ class ProjectConfig(BaseModel):
     dependencies: List[str] = Field(default_factory=list)
     license: License = Field(default_factory=License)
     urls: URLs = Field(default_factory=URLs)
+    supported_os: List[str] = Field(default_factory=list)
+    supported_accelerators: List[str] = Field(default_factory=list)
+    supported_comfyui_version: str = ""
+    supported_comfyui_frontend_version: str = ""
 
     @field_validator('license', mode='before')
     @classmethod

comfy_execution/caching.py

@@ -1,6 +1,7 @@
 import itertools
 from typing import Sequence, Mapping, Dict
 from comfy_execution.graph import DynamicPrompt
+from abc import ABC, abstractmethod
 
 import nodes
 
@@ -16,12 +17,13 @@ def include_unique_id_in_input(class_type: str) -> bool:
     NODE_CLASS_CONTAINS_UNIQUE_ID[class_type] = "UNIQUE_ID" in class_def.INPUT_TYPES().get("hidden", {}).values()
     return NODE_CLASS_CONTAINS_UNIQUE_ID[class_type]
 
-class CacheKeySet:
+class CacheKeySet(ABC):
     def __init__(self, dynprompt, node_ids, is_changed_cache):
         self.keys = {}
         self.subcache_keys = {}
 
-    def add_keys(self, node_ids):
+    @abstractmethod
+    async def add_keys(self, node_ids):
         raise NotImplementedError()
 
     def all_node_ids(self):
@@ -60,9 +62,8 @@ class CacheKeySetID(CacheKeySet):
     def __init__(self, dynprompt, node_ids, is_changed_cache):
         super().__init__(dynprompt, node_ids, is_changed_cache)
         self.dynprompt = dynprompt
-        self.add_keys(node_ids)
 
-    def add_keys(self, node_ids):
+    async def add_keys(self, node_ids):
         for node_id in node_ids:
             if node_id in self.keys:
                 continue
@@ -77,37 +78,36 @@ class CacheKeySetInputSignature(CacheKeySet):
         super().__init__(dynprompt, node_ids, is_changed_cache)
         self.dynprompt = dynprompt
         self.is_changed_cache = is_changed_cache
-        self.add_keys(node_ids)
 
     def include_node_id_in_input(self) -> bool:
         return False
 
-    def add_keys(self, node_ids):
+    async def add_keys(self, node_ids):
         for node_id in node_ids:
             if node_id in self.keys:
                 continue
             if not self.dynprompt.has_node(node_id):
                 continue
             node = self.dynprompt.get_node(node_id)
-            self.keys[node_id] = self.get_node_signature(self.dynprompt, node_id)
+            self.keys[node_id] = await self.get_node_signature(self.dynprompt, node_id)
             self.subcache_keys[node_id] = (node_id, node["class_type"])
 
-    def get_node_signature(self, dynprompt, node_id):
+    async def get_node_signature(self, dynprompt, node_id):
         signature = []
         ancestors, order_mapping = self.get_ordered_ancestry(dynprompt, node_id)
-        signature.append(self.get_immediate_node_signature(dynprompt, node_id, order_mapping))
+        signature.append(await self.get_immediate_node_signature(dynprompt, node_id, order_mapping))
         for ancestor_id in ancestors:
-            signature.append(self.get_immediate_node_signature(dynprompt, ancestor_id, order_mapping))
+            signature.append(await self.get_immediate_node_signature(dynprompt, ancestor_id, order_mapping))
         return to_hashable(signature)
 
-    def get_immediate_node_signature(self, dynprompt, node_id, ancestor_order_mapping):
+    async def get_immediate_node_signature(self, dynprompt, node_id, ancestor_order_mapping):
         if not dynprompt.has_node(node_id):
             # This node doesn't exist -- we can't cache it.
             return [float("NaN")]
         node = dynprompt.get_node(node_id)
         class_type = node["class_type"]
         class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
-        signature = [class_type, self.is_changed_cache.get(node_id)]
+        signature = [class_type, await self.is_changed_cache.get(node_id)]
         if self.include_node_id_in_input() or (hasattr(class_def, "NOT_IDEMPOTENT") and class_def.NOT_IDEMPOTENT) or include_unique_id_in_input(class_type):
             signature.append(node_id)
         inputs = node["inputs"]
@@ -150,9 +150,10 @@ class BasicCache:
         self.cache = {}
         self.subcaches = {}
 
-    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+    async def set_prompt(self, dynprompt, node_ids, is_changed_cache):
         self.dynprompt = dynprompt
         self.cache_key_set = self.key_class(dynprompt, node_ids, is_changed_cache)
+        await self.cache_key_set.add_keys(node_ids)
         self.is_changed_cache = is_changed_cache
         self.initialized = True
 
@@ -201,13 +202,13 @@ class BasicCache:
         else:
             return None
 
-    def _ensure_subcache(self, node_id, children_ids):
+    async def _ensure_subcache(self, node_id, children_ids):
         subcache_key = self.cache_key_set.get_subcache_key(node_id)
         subcache = self.subcaches.get(subcache_key, None)
         if subcache is None:
             subcache = BasicCache(self.key_class)
             self.subcaches[subcache_key] = subcache
-        subcache.set_prompt(self.dynprompt, children_ids, self.is_changed_cache)
+        await subcache.set_prompt(self.dynprompt, children_ids, self.is_changed_cache)
         return subcache
 
     def _get_subcache(self, node_id):
@@ -259,10 +260,10 @@ class HierarchicalCache(BasicCache):
         assert cache is not None
         cache._set_immediate(node_id, value)
 
-    def ensure_subcache_for(self, node_id, children_ids):
+    async def ensure_subcache_for(self, node_id, children_ids):
         cache = self._get_cache_for(node_id)
         assert cache is not None
-        return cache._ensure_subcache(node_id, children_ids)
+        return await cache._ensure_subcache(node_id, children_ids)
 
 class LRUCache(BasicCache):
     def __init__(self, key_class, max_size=100):
@@ -273,8 +274,8 @@ class LRUCache(BasicCache):
         self.used_generation = {}
         self.children = {}
 
-    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
-        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+    async def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+        await super().set_prompt(dynprompt, node_ids, is_changed_cache)
         self.generation += 1
         for node_id in node_ids:
             self._mark_used(node_id)
@@ -303,11 +304,11 @@ class LRUCache(BasicCache):
         self._mark_used(node_id)
         return self._set_immediate(node_id, value)
 
-    def ensure_subcache_for(self, node_id, children_ids):
+    async def ensure_subcache_for(self, node_id, children_ids):
         # Just uses subcaches for tracking 'live' nodes
-        super()._ensure_subcache(node_id, children_ids)
+        await super()._ensure_subcache(node_id, children_ids)
 
-        self.cache_key_set.add_keys(children_ids)
+        await self.cache_key_set.add_keys(children_ids)
         self._mark_used(node_id)
         cache_key = self.cache_key_set.get_data_key(node_id)
         self.children[cache_key] = []
@@ -337,7 +338,7 @@ class DependencyAwareCache(BasicCache):
         self.ancestors = {}    # Maps node_id -> set of ancestor node_ids
         self.executed_nodes = set()  # Tracks nodes that have been executed
 
-    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+    async def set_prompt(self, dynprompt, node_ids, is_changed_cache):
         """
         Clear the entire cache and rebuild the dependency graph.
 
@@ -354,7 +355,7 @@ class DependencyAwareCache(BasicCache):
         self.executed_nodes.clear()
 
         # Call the parent method to initialize the cache with the new prompt
-        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+        await super().set_prompt(dynprompt, node_ids, is_changed_cache)
 
         # Rebuild the dependency graph
         self._build_dependency_graph(dynprompt, node_ids)
@@ -405,7 +406,7 @@ class DependencyAwareCache(BasicCache):
         """
         return self._get_immediate(node_id)
 
-    def ensure_subcache_for(self, node_id, children_ids):
+    async def ensure_subcache_for(self, node_id, children_ids):
         """
         Ensure a subcache exists for a node and update dependencies.
 
@@ -416,7 +417,7 @@ class DependencyAwareCache(BasicCache):
         Returns:
             The subcache object for the node.
         """
-        subcache = super()._ensure_subcache(node_id, children_ids)
+        subcache = await super()._ensure_subcache(node_id, children_ids)
         for child_id in children_ids:
             self.descendants[node_id].add(child_id)
             self.ancestors[child_id].add(node_id)

comfy_execution/graph.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 from typing import Type, Literal
 
 import nodes
+import asyncio
 from comfy_execution.graph_utils import is_link
 from comfy.comfy_types.node_typing import ComfyNodeABC, InputTypeDict, InputTypeOptions
 
@@ -100,6 +101,8 @@ class TopologicalSort:
         self.pendingNodes = {}
         self.blockCount = {} # Number of nodes this node is directly blocked by
         self.blocking = {} # Which nodes are blocked by this node
+        self.externalBlocks = 0
+        self.unblockedEvent = asyncio.Event()
 
     def get_input_info(self, unique_id, input_name):
         class_type = self.dynprompt.get_node(unique_id)["class_type"]
@@ -153,6 +156,16 @@ class TopologicalSort:
         for link in links:
             self.add_strong_link(*link)
 
+    def add_external_block(self, node_id):
+        assert node_id in self.blockCount, "Can't add external block to a node that isn't pending"
+        self.externalBlocks += 1
+        self.blockCount[node_id] += 1
+        def unblock():
+            self.externalBlocks -= 1
+            self.blockCount[node_id] -= 1
+            self.unblockedEvent.set()
+        return unblock
+
     def is_cached(self, node_id):
         return False
 
@@ -181,11 +194,16 @@ class ExecutionList(TopologicalSort):
     def is_cached(self, node_id):
         return self.output_cache.get(node_id) is not None
 
-    def stage_node_execution(self):
+    async def stage_node_execution(self):
         assert self.staged_node_id is None
         if self.is_empty():
             return None, None, None
         available = self.get_ready_nodes()
+        while len(available) == 0 and self.externalBlocks > 0:
+            # Wait for an external block to be released
+            await self.unblockedEvent.wait()
+            self.unblockedEvent.clear()
+            available = self.get_ready_nodes()
         if len(available) == 0:
             cycled_nodes = self.get_nodes_in_cycle()
             # Because cycles composed entirely of static nodes are caught during initial validation,

comfy_execution/progress.py

@@ -0,0 +1,288 @@
+from typing import TypedDict, Dict, Optional
+from typing_extensions import override
+from PIL import Image
+from enum import Enum
+from abc import ABC
+from tqdm import tqdm
+from comfy_execution.graph import DynamicPrompt
+from protocol import BinaryEventTypes
+
+class NodeState(Enum):
+    Pending = "pending"
+    Running = "running"
+    Finished = "finished"
+    Error = "error"
+
+class NodeProgressState(TypedDict):
+    """
+    A class to represent the state of a node's progress.
+    """
+    state: NodeState
+    value: float
+    max: float
+
+class ProgressHandler(ABC):
+    """
+    Abstract base class for progress handlers.
+    Progress handlers receive progress updates and display them in various ways.
+    """
+    def __init__(self, name: str):
+        self.name = name
+        self.enabled = True
+
+    def set_registry(self, registry: "ProgressRegistry"):
+        pass
+
+    def start_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        """Called when a node starts processing"""
+        pass
+
+    def update_handler(self, node_id: str, value: float, max_value: float,
+                      state: NodeProgressState, prompt_id: str, image: Optional[Image.Image] = None):
+        """Called when a node's progress is updated"""
+        pass
+
+    def finish_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        """Called when a node finishes processing"""
+        pass
+
+    def reset(self):
+        """Called when the progress registry is reset"""
+        pass
+
+    def enable(self):
+        """Enable this handler"""
+        self.enabled = True
+
+    def disable(self):
+        """Disable this handler"""
+        self.enabled = False
+
+class CLIProgressHandler(ProgressHandler):
+    """
+    Handler that displays progress using tqdm progress bars in the CLI.
+    """
+    def __init__(self):
+        super().__init__("cli")
+        self.progress_bars: Dict[str, tqdm] = {}
+
+    @override
+    def start_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Create a new tqdm progress bar
+        if node_id not in self.progress_bars:
+            self.progress_bars[node_id] = tqdm(
+                total=state["max"],
+                desc=f"Node {node_id}",
+                unit="steps",
+                leave=True,
+                position=len(self.progress_bars)
+            )
+
+    @override
+    def update_handler(self, node_id: str, value: float, max_value: float,
+                      state: NodeProgressState, prompt_id: str, image: Optional[Image.Image] = None):
+        # Handle case where start_handler wasn't called
+        if node_id not in self.progress_bars:
+            self.progress_bars[node_id] = tqdm(
+                total=max_value,
+                desc=f"Node {node_id}",
+                unit="steps",
+                leave=True,
+                position=len(self.progress_bars)
+            )
+            self.progress_bars[node_id].update(value)
+        else:
+            # Update existing progress bar
+            if max_value != self.progress_bars[node_id].total:
+                self.progress_bars[node_id].total = max_value
+            # Calculate the update amount (difference from current position)
+            current_position = self.progress_bars[node_id].n
+            update_amount = value - current_position
+            if update_amount > 0:
+                self.progress_bars[node_id].update(update_amount)
+
+    @override
+    def finish_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Complete and close the progress bar if it exists
+        if node_id in self.progress_bars:
+            # Ensure the bar shows 100% completion
+            remaining = state["max"] - self.progress_bars[node_id].n
+            if remaining > 0:
+                self.progress_bars[node_id].update(remaining)
+            self.progress_bars[node_id].close()
+            del self.progress_bars[node_id]
+
+    @override
+    def reset(self):
+        # Close all progress bars
+        for bar in self.progress_bars.values():
+            bar.close()
+        self.progress_bars.clear()
+
+class WebUIProgressHandler(ProgressHandler):
+    """
+    Handler that sends progress updates to the WebUI via WebSockets.
+    """
+    def __init__(self, server_instance):
+        super().__init__("webui")
+        self.server_instance = server_instance
+
+    def set_registry(self, registry: "ProgressRegistry"):
+        self.registry = registry
+
+    def _send_progress_state(self, prompt_id: str, nodes: Dict[str, NodeProgressState]):
+        """Send the current progress state to the client"""
+        if self.server_instance is None:
+            return
+
+        # Only send info for non-pending nodes
+        active_nodes = {
+            node_id: {
+                "value": state["value"],
+                "max": state["max"],
+                "state": state["state"].value,
+                "node_id": node_id,
+                "prompt_id": prompt_id,
+                "display_node_id": self.registry.dynprompt.get_display_node_id(node_id),
+                "parent_node_id": self.registry.dynprompt.get_parent_node_id(node_id),
+                "real_node_id": self.registry.dynprompt.get_real_node_id(node_id)
+            }
+            for node_id, state in nodes.items()
+            if state["state"] != NodeState.Pending
+        }
+
+        # Send a combined progress_state message with all node states
+        self.server_instance.send_sync("progress_state", {
+            "prompt_id": prompt_id,
+            "nodes": active_nodes
+        })
+
+    @override
+    def start_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Send progress state of all nodes
+        if self.registry:
+            self._send_progress_state(prompt_id, self.registry.nodes)
+
+    @override
+    def update_handler(self, node_id: str, value: float, max_value: float,
+                      state: NodeProgressState, prompt_id: str, image: Optional[Image.Image] = None):
+        # Send progress state of all nodes
+        if self.registry:
+            self._send_progress_state(prompt_id, self.registry.nodes)
+        if image:
+            metadata = {
+                "node_id": node_id,
+                "prompt_id": prompt_id,
+                "display_node_id": self.registry.dynprompt.get_display_node_id(node_id),
+                "parent_node_id": self.registry.dynprompt.get_parent_node_id(node_id),
+                "real_node_id": self.registry.dynprompt.get_real_node_id(node_id)
+            }
+            self.server_instance.send_sync(BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA, (image, metadata), self.server_instance.client_id)
+
+
+    @override
+    def finish_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Send progress state of all nodes
+        if self.registry:
+            self._send_progress_state(prompt_id, self.registry.nodes)
+
+class ProgressRegistry:
+    """
+    Registry that maintains node progress state and notifies registered handlers.
+    """
+    def __init__(self, prompt_id: str, dynprompt: DynamicPrompt):
+        self.prompt_id = prompt_id
+        self.dynprompt = dynprompt
+        self.nodes: Dict[str, NodeProgressState] = {}
+        self.handlers: Dict[str, ProgressHandler] = {}
+
+    def register_handler(self, handler: ProgressHandler) -> None:
+        """Register a progress handler"""
+        self.handlers[handler.name] = handler
+
+    def unregister_handler(self, handler_name: str) -> None:
+        """Unregister a progress handler"""
+        if handler_name in self.handlers:
+            # Allow handler to clean up resources
+            self.handlers[handler_name].reset()
+            del self.handlers[handler_name]
+
+    def enable_handler(self, handler_name: str) -> None:
+        """Enable a progress handler"""
+        if handler_name in self.handlers:
+            self.handlers[handler_name].enable()
+
+    def disable_handler(self, handler_name: str) -> None:
+        """Disable a progress handler"""
+        if handler_name in self.handlers:
+            self.handlers[handler_name].disable()
+
+    def ensure_entry(self, node_id: str) -> NodeProgressState:
+        """Ensure a node entry exists"""
+        if node_id not in self.nodes:
+            self.nodes[node_id] = NodeProgressState(
+                state = NodeState.Pending,
+                value = 0,
+                max = 1
+            )
+        return self.nodes[node_id]
+
+    def start_progress(self, node_id: str) -> None:
+        """Start progress tracking for a node"""
+        entry = self.ensure_entry(node_id)
+        entry["state"] = NodeState.Running
+        entry["value"] = 0.0
+        entry["max"] = 1.0
+
+        # Notify all enabled handlers
+        for handler in self.handlers.values():
+            if handler.enabled:
+                handler.start_handler(node_id, entry, self.prompt_id)
+
+    def update_progress(self, node_id: str, value: float, max_value: float, image: Optional[Image.Image]) -> None:
+        """Update progress for a node"""
+        entry = self.ensure_entry(node_id)
+        entry["state"] = NodeState.Running
+        entry["value"] = value
+        entry["max"] = max_value
+
+        # Notify all enabled handlers
+        for handler in self.handlers.values():
+            if handler.enabled:
+                handler.update_handler(node_id, value, max_value, entry, self.prompt_id, image)
+
+    def finish_progress(self, node_id: str) -> None:
+        """Finish progress tracking for a node"""
+        entry = self.ensure_entry(node_id)
+        entry["state"] = NodeState.Finished
+        entry["value"] = entry["max"]
+
+        # Notify all enabled handlers
+        for handler in self.handlers.values():
+            if handler.enabled:
+                handler.finish_handler(node_id, entry, self.prompt_id)
+
+    def reset_handlers(self) -> None:
+        """Reset all handlers"""
+        for handler in self.handlers.values():
+            handler.reset()
+
+# Global registry instance
+global_progress_registry: ProgressRegistry = ProgressRegistry(prompt_id="", dynprompt=DynamicPrompt({}))
+
+def reset_progress_state(prompt_id: str, dynprompt: DynamicPrompt) -> None:
+    global global_progress_registry
+
+    # Reset existing handlers if registry exists
+    if global_progress_registry is not None:
+        global_progress_registry.reset_handlers()
+
+    # Create new registry
+    global_progress_registry = ProgressRegistry(prompt_id, dynprompt)
+
+def add_progress_handler(handler: ProgressHandler) -> None:
+    handler.set_registry(global_progress_registry)
+    global_progress_registry.register_handler(handler)
+
+def get_progress_state() -> ProgressRegistry:
+    return global_progress_registry

comfy_execution/utils.py

@@ -0,0 +1,46 @@
+import contextvars
+from typing import Optional, NamedTuple
+
+class ExecutionContext(NamedTuple):
+    """
+    Context information about the currently executing node.
+
+    Attributes:
+        node_id: The ID of the currently executing node
+        list_index: The index in a list being processed (for operations on batches/lists)
+    """
+    prompt_id: str
+    node_id: str
+    list_index: Optional[int]
+
+current_executing_context: contextvars.ContextVar[Optional[ExecutionContext]] = contextvars.ContextVar("current_executing_context", default=None)
+
+def get_executing_context() -> Optional[ExecutionContext]:
+    return current_executing_context.get(None)
+
+class CurrentNodeContext:
+    """
+    Context manager for setting the current executing node context.
+
+    Sets the current_executing_context on enter and resets it on exit.
+
+    Example:
+        with CurrentNodeContext(node_id="123", list_index=0):
+            # Code that should run with the current node context set
+            process_image()
+    """
+    def __init__(self, prompt_id: str, node_id: str, list_index: Optional[int] = None):
+        self.context = ExecutionContext(
+            prompt_id= prompt_id,
+            node_id= node_id,
+            list_index= list_index
+        )
+        self.token = None
+
+    def __enter__(self):
+        self.token = current_executing_context.set(self.context)
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.token is not None:
+            current_executing_context.reset(self.token)

comfy_extras/nodes_custom_sampler.py

@@ -2,6 +2,7 @@ import math
 import comfy.samplers
 import comfy.sample
 from comfy.k_diffusion import sampling as k_diffusion_sampling
+from comfy.comfy_types import IO, ComfyNodeABC, InputTypeDict
 import latent_preview
 import torch
 import comfy.utils
@@ -480,6 +481,46 @@ class SamplerDPMAdaptative:
                                                               "s_noise":s_noise })
         return (sampler, )
 
+
+class SamplerER_SDE(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {
+                "solver_type": (IO.COMBO, {"options": ["ER-SDE", "Reverse-time SDE", "ODE"]}),
+                "max_stage": (IO.INT, {"default": 3, "min": 1, "max": 3}),
+                "eta": (
+                    IO.FLOAT,
+                    {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False, "tooltip": "Stochastic strength of reverse-time SDE.\nWhen eta=0, it reduces to deterministic ODE. This setting doesn't apply to ER-SDE solver type."},
+                ),
+                "s_noise": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False}),
+            }
+        }
+
+    RETURN_TYPES = (IO.SAMPLER,)
+    CATEGORY = "sampling/custom_sampling/samplers"
+
+    FUNCTION = "get_sampler"
+
+    def get_sampler(self, solver_type, max_stage, eta, s_noise):
+        if solver_type == "ODE" or (solver_type == "Reverse-time SDE" and eta == 0):
+            eta = 0
+            s_noise = 0
+
+        def reverse_time_sde_noise_scaler(x):
+            return x ** (eta + 1)
+
+        if solver_type == "ER-SDE":
+            # Use the default one in sample_er_sde()
+            noise_scaler = None
+        else:
+            noise_scaler = reverse_time_sde_noise_scaler
+
+        sampler_name = "er_sde"
+        sampler = comfy.samplers.ksampler(sampler_name, {"s_noise": s_noise, "noise_scaler": noise_scaler, "max_stage": max_stage})
+        return (sampler,)
+
+
 class Noise_EmptyNoise:
     def __init__(self):
         self.seed = 0
@@ -609,8 +650,14 @@ class Guider_DualCFG(comfy.samplers.CFGGuider):
     def predict_noise(self, x, timestep, model_options={}, seed=None):
         negative_cond = self.conds.get("negative", None)
         middle_cond = self.conds.get("middle", None)
+        positive_cond = self.conds.get("positive", None)
+        if model_options.get("disable_cfg1_optimization", False) == False:
+            if math.isclose(self.cfg2, 1.0):
+                negative_cond = None
+                if math.isclose(self.cfg1, 1.0):
+                    middle_cond = None
 
-        out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, self.conds.get("positive", None)], x, timestep, model_options)
+        out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options)
         return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1
 
 class DualCFGGuider:
@@ -781,6 +828,7 @@ NODE_CLASS_MAPPINGS = {
     "SamplerDPMPP_SDE": SamplerDPMPP_SDE,
     "SamplerDPMPP_2S_Ancestral": SamplerDPMPP_2S_Ancestral,
     "SamplerDPMAdaptative": SamplerDPMAdaptative,
+    "SamplerER_SDE": SamplerER_SDE,
     "SplitSigmas": SplitSigmas,
     "SplitSigmasDenoise": SplitSigmasDenoise,
     "FlipSigmas": FlipSigmas,

comfy_extras/nodes_edit_model.py

@@ -0,0 +1,26 @@
+import node_helpers
+
+
+class ReferenceLatent:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"conditioning": ("CONDITIONING", ),
+                            },
+                "optional": {"latent": ("LATENT", ),}
+               }
+
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "append"
+
+    CATEGORY = "advanced/conditioning/edit_models"
+    DESCRIPTION = "This node sets the guiding latent for an edit model. If the model supports it you can chain multiple to set multiple reference images."
+
+    def append(self, conditioning, latent=None):
+        if latent is not None:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_latents": [latent["samples"]]}, append=True)
+        return (conditioning, )
+
+
+NODE_CLASS_MAPPINGS = {
+    "ReferenceLatent": ReferenceLatent,
+}

comfy_extras/nodes_flux.py

@@ -1,4 +1,5 @@
 import node_helpers
+import comfy.utils
 
 class CLIPTextEncodeFlux:
     @classmethod
@@ -56,8 +57,52 @@ class FluxDisableGuidance:
         return (c, )
 
 
+PREFERED_KONTEXT_RESOLUTIONS = [
+    (672, 1568),
+    (688, 1504),
+    (720, 1456),
+    (752, 1392),
+    (800, 1328),
+    (832, 1248),
+    (880, 1184),
+    (944, 1104),
+    (1024, 1024),
+    (1104, 944),
+    (1184, 880),
+    (1248, 832),
+    (1328, 800),
+    (1392, 752),
+    (1456, 720),
+    (1504, 688),
+    (1568, 672),
+]
+
+
+class FluxKontextImageScale:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"image": ("IMAGE", ),
+                            },
+               }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "scale"
+
+    CATEGORY = "advanced/conditioning/flux"
+    DESCRIPTION = "This node resizes the image to one that is more optimal for flux kontext."
+
+    def scale(self, image):
+        width = image.shape[2]
+        height = image.shape[1]
+        aspect_ratio = width / height
+        _, width, height = min((abs(aspect_ratio - w / h), w, h) for w, h in PREFERED_KONTEXT_RESOLUTIONS)
+        image = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "lanczos", "center").movedim(1, -1)
+        return (image, )
+
+
 NODE_CLASS_MAPPINGS = {
     "CLIPTextEncodeFlux": CLIPTextEncodeFlux,
     "FluxGuidance": FluxGuidance,
     "FluxDisableGuidance": FluxDisableGuidance,
+    "FluxKontextImageScale": FluxKontextImageScale,
 }

comfy_extras/nodes_images.py

@@ -304,10 +304,23 @@ Optional spacing can be added between images.
                 image2.movedim(-1, 1), target_w, target_h, "lanczos", "disabled"
             ).movedim(1, -1)
 
+        color_map = {
+            "white": 1.0,
+            "black": 0.0,
+            "red": (1.0, 0.0, 0.0),
+            "green": (0.0, 1.0, 0.0),
+            "blue": (0.0, 0.0, 1.0),
+        }
+
+        color_val = color_map[spacing_color]
+
         # When not matching sizes, pad to align non-concat dimensions
         if not match_image_size:
             h1, w1 = image1.shape[1:3]
             h2, w2 = image2.shape[1:3]
+            pad_value = 0.0
+            if not isinstance(color_val, tuple):
+                pad_value = color_val
 
             if direction in ["left", "right"]:
                 # For horizontal concat, pad heights to match
@@ -316,11 +329,11 @@ Optional spacing can be added between images.
                     if h1 < target_h:
                         pad_h = target_h - h1
                         pad_top, pad_bottom = pad_h // 2, pad_h - pad_h // 2
-                        image1 = torch.nn.functional.pad(image1, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=0.0)
+                        image1 = torch.nn.functional.pad(image1, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=pad_value)
                     if h2 < target_h:
                         pad_h = target_h - h2
                         pad_top, pad_bottom = pad_h // 2, pad_h - pad_h // 2
-                        image2 = torch.nn.functional.pad(image2, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=0.0)
+                        image2 = torch.nn.functional.pad(image2, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=pad_value)
             else:  # up, down
                 # For vertical concat, pad widths to match
                 if w1 != w2:
@@ -328,11 +341,11 @@ Optional spacing can be added between images.
                     if w1 < target_w:
                         pad_w = target_w - w1
                         pad_left, pad_right = pad_w // 2, pad_w - pad_w // 2
-                        image1 = torch.nn.functional.pad(image1, (0, 0, pad_left, pad_right), mode='constant', value=0.0)
+                        image1 = torch.nn.functional.pad(image1, (0, 0, pad_left, pad_right), mode='constant', value=pad_value)
                     if w2 < target_w:
                         pad_w = target_w - w2
                         pad_left, pad_right = pad_w // 2, pad_w - pad_w // 2
-                        image2 = torch.nn.functional.pad(image2, (0, 0, pad_left, pad_right), mode='constant', value=0.0)
+                        image2 = torch.nn.functional.pad(image2, (0, 0, pad_left, pad_right), mode='constant', value=pad_value)
 
         # Ensure same number of channels
         if image1.shape[-1] != image2.shape[-1]:
@@ -366,15 +379,6 @@ Optional spacing can be added between images.
         if spacing_width > 0:
             spacing_width = spacing_width + (spacing_width % 2)  # Ensure even
 
-            color_map = {
-                "white": 1.0,
-                "black": 0.0,
-                "red": (1.0, 0.0, 0.0),
-                "green": (0.0, 1.0, 0.0),
-                "blue": (0.0, 0.0, 1.0),
-            }
-            color_val = color_map[spacing_color]
-
             if direction in ["left", "right"]:
                 spacing_shape = (
                     image1.shape[0],
@@ -410,6 +414,62 @@ Optional spacing can be added between images.
         concat_dim = 2 if direction in ["left", "right"] else 1
         return (torch.cat(images, dim=concat_dim),)
 
+class ResizeAndPadImage:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "target_width": ("INT", {
+                    "default": 512,
+                    "min": 1,
+                    "max": MAX_RESOLUTION,
+                    "step": 1
+                }),
+                "target_height": ("INT", {
+                    "default": 512,
+                    "min": 1,
+                    "max": MAX_RESOLUTION,
+                    "step": 1
+                }),
+                "padding_color": (["white", "black"],),
+                "interpolation": (["area", "bicubic", "nearest-exact", "bilinear", "lanczos"],),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "resize_and_pad"
+    CATEGORY = "image/transform"
+
+    def resize_and_pad(self, image, target_width, target_height, padding_color, interpolation):
+        batch_size, orig_height, orig_width, channels = image.shape
+
+        scale_w = target_width / orig_width
+        scale_h = target_height / orig_height
+        scale = min(scale_w, scale_h)
+
+        new_width = int(orig_width * scale)
+        new_height = int(orig_height * scale)
+
+        image_permuted = image.permute(0, 3, 1, 2)
+
+        resized = comfy.utils.common_upscale(image_permuted, new_width, new_height, interpolation, "disabled")
+
+        pad_value = 0.0 if padding_color == "black" else 1.0
+        padded = torch.full(
+            (batch_size, channels, target_height, target_width),
+            pad_value,
+            dtype=image.dtype,
+            device=image.device
+        )
+
+        y_offset = (target_height - new_height) // 2
+        x_offset = (target_width - new_width) // 2
+
+        padded[:, :, y_offset:y_offset + new_height, x_offset:x_offset + new_width] = resized
+
+        output = padded.permute(0, 2, 3, 1)
+        return (output,)
 
 class SaveSVGNode:
     """
@@ -532,5 +592,6 @@ NODE_CLASS_MAPPINGS = {
     "SaveAnimatedPNG": SaveAnimatedPNG,
     "SaveSVGNode": SaveSVGNode,
     "ImageStitch": ImageStitch,
+    "ResizeAndPadImage": ResizeAndPadImage,
     "GetImageSize": GetImageSize,
 }

comfy_extras/nodes_model_merging_model_specific.py

@@ -268,6 +268,52 @@ class ModelMergeWAN2_1(comfy_extras.nodes_model_merging.ModelMergeBlocks):
 
         return {"required": arg_dict}
 
+class ModelMergeCosmosPredict2_2B(comfy_extras.nodes_model_merging.ModelMergeBlocks):
+    CATEGORY = "advanced/model_merging/model_specific"
+
+    @classmethod
+    def INPUT_TYPES(s):
+        arg_dict = { "model1": ("MODEL",),
+                              "model2": ("MODEL",)}
+
+        argument = ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01})
+
+        arg_dict["pos_embedder."] = argument
+        arg_dict["x_embedder."] = argument
+        arg_dict["t_embedder."] = argument
+        arg_dict["t_embedding_norm."] = argument
+
+
+        for i in range(28):
+            arg_dict["blocks.{}.".format(i)] = argument
+
+        arg_dict["final_layer."] = argument
+
+        return {"required": arg_dict}
+
+class ModelMergeCosmosPredict2_14B(comfy_extras.nodes_model_merging.ModelMergeBlocks):
+    CATEGORY = "advanced/model_merging/model_specific"
+
+    @classmethod
+    def INPUT_TYPES(s):
+        arg_dict = { "model1": ("MODEL",),
+                              "model2": ("MODEL",)}
+
+        argument = ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01})
+
+        arg_dict["pos_embedder."] = argument
+        arg_dict["x_embedder."] = argument
+        arg_dict["t_embedder."] = argument
+        arg_dict["t_embedding_norm."] = argument
+
+
+        for i in range(36):
+            arg_dict["blocks.{}.".format(i)] = argument
+
+        arg_dict["final_layer."] = argument
+
+        return {"required": arg_dict}
+
 NODE_CLASS_MAPPINGS = {
     "ModelMergeSD1": ModelMergeSD1,
     "ModelMergeSD2": ModelMergeSD1, #SD1 and SD2 have the same blocks
@@ -281,4 +327,6 @@ NODE_CLASS_MAPPINGS = {
     "ModelMergeCosmos7B": ModelMergeCosmos7B,
     "ModelMergeCosmos14B": ModelMergeCosmos14B,
     "ModelMergeWAN2_1": ModelMergeWAN2_1,
+    "ModelMergeCosmosPredict2_2B": ModelMergeCosmosPredict2_2B,
+    "ModelMergeCosmosPredict2_14B": ModelMergeCosmosPredict2_14B,
 }

comfy_extras/nodes_perpneg.py

@@ -4,6 +4,7 @@ import comfy.sampler_helpers
 import comfy.samplers
 import comfy.utils
 import node_helpers
+import math
 
 def perp_neg(x, noise_pred_pos, noise_pred_neg, noise_pred_nocond, neg_scale, cond_scale):
     pos = noise_pred_pos - noise_pred_nocond
@@ -69,8 +70,23 @@ class Guider_PerpNeg(comfy.samplers.CFGGuider):
         negative_cond = self.conds.get("negative", None)
         empty_cond = self.conds.get("empty_negative_prompt", None)
 
-        (noise_pred_pos, noise_pred_neg, noise_pred_empty) = \
-            comfy.samplers.calc_cond_batch(self.inner_model, [positive_cond, negative_cond, empty_cond], x, timestep, model_options)
+        if model_options.get("disable_cfg1_optimization", False) == False:
+            if math.isclose(self.neg_scale, 0.0):
+                negative_cond = None
+                if math.isclose(self.cfg, 1.0):
+                    empty_cond = None
+
+        conds = [positive_cond, negative_cond, empty_cond]
+
+        out = comfy.samplers.calc_cond_batch(self.inner_model, conds, x, timestep, model_options)
+
+        # Apply pre_cfg_functions since sampling_function() is skipped
+        for fn in model_options.get("sampler_pre_cfg_function", []):
+            args = {"conds":conds, "conds_out": out, "cond_scale": self.cfg, "timestep": timestep,
+                    "input": x, "sigma": timestep, "model": self.inner_model, "model_options": model_options}
+            out = fn(args)
+
+        noise_pred_pos, noise_pred_neg, noise_pred_empty = out
         cfg_result = perp_neg(x, noise_pred_pos, noise_pred_neg, noise_pred_empty, self.neg_scale, self.cfg)
 
         # normally this would be done in cfg_function, but we skipped
@@ -82,6 +98,7 @@ class Guider_PerpNeg(comfy.samplers.CFGGuider):
                 "denoised": cfg_result,
                 "cond": positive_cond,
                 "uncond": negative_cond,
+                "cond_scale": self.cfg,
                 "model": self.inner_model,
                 "uncond_denoised": noise_pred_neg,
                 "cond_denoised": noise_pred_pos,

comfy_extras/nodes_tcfg.py

@@ -0,0 +1,71 @@
+# TCFG: Tangential Damping Classifier-free Guidance - (arXiv: https://arxiv.org/abs/2503.18137)
+
+import torch
+
+from comfy.comfy_types import IO, ComfyNodeABC, InputTypeDict
+
+
+def score_tangential_damping(cond_score: torch.Tensor, uncond_score: torch.Tensor) -> torch.Tensor:
+    """Drop tangential components from uncond score to align with cond score."""
+    # (B, 1, ...)
+    batch_num = cond_score.shape[0]
+    cond_score_flat = cond_score.reshape(batch_num, 1, -1).float()
+    uncond_score_flat = uncond_score.reshape(batch_num, 1, -1).float()
+
+    # Score matrix A (B, 2, ...)
+    score_matrix = torch.cat((uncond_score_flat, cond_score_flat), dim=1)
+    try:
+        _, _, Vh = torch.linalg.svd(score_matrix, full_matrices=False)
+    except RuntimeError:
+        # Fallback to CPU
+        _, _, Vh = torch.linalg.svd(score_matrix.cpu(), full_matrices=False)
+
+    # Drop the tangential components
+    v1 = Vh[:, 0:1, :].to(uncond_score_flat.device)  # (B, 1, ...)
+    uncond_score_td = (uncond_score_flat @ v1.transpose(-2, -1)) * v1
+    return uncond_score_td.reshape_as(uncond_score).to(uncond_score.dtype)
+
+
+class TCFG(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {
+                "model": (IO.MODEL, {}),
+            }
+        }
+
+    RETURN_TYPES = (IO.MODEL,)
+    RETURN_NAMES = ("patched_model",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/guidance"
+    DESCRIPTION = "TCFG – Tangential Damping CFG (2503.18137)\n\nRefine the uncond (negative) to align with the cond (positive) for improving quality."
+
+    def patch(self, model):
+        m = model.clone()
+
+        def tangential_damping_cfg(args):
+            #  Assume [cond, uncond, ...]
+            x = args["input"]
+            conds_out = args["conds_out"]
+            if len(conds_out) <= 1 or None in args["conds"][:2]:
+                # Skip when either cond or uncond is None
+                return conds_out
+            cond_pred = conds_out[0]
+            uncond_pred = conds_out[1]
+            uncond_td = score_tangential_damping(x - cond_pred, x - uncond_pred)
+            uncond_pred_td = x - uncond_td
+            return [cond_pred, uncond_pred_td] + conds_out[2:]
+
+        m.set_model_sampler_pre_cfg_function(tangential_damping_cfg)
+        return (m,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "TCFG": TCFG,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "TCFG": "Tangential Damping CFG",
+}

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.41"
+__version__ = "0.3.43"

execution.py

@@ -8,12 +8,14 @@ import time
 import traceback
 from enum import Enum
 from typing import List, Literal, NamedTuple, Optional
+import asyncio
 
 import torch
 
 import comfy.model_management
 import nodes
 from comfy_execution.caching import (
+    BasicCache,
     CacheKeySetID,
     CacheKeySetInputSignature,
     DependencyAwareCache,
@@ -28,6 +30,8 @@ from comfy_execution.graph import (
 )
 from comfy_execution.graph_utils import GraphBuilder, is_link
 from comfy_execution.validation import validate_node_input
+from comfy_execution.progress import get_progress_state, reset_progress_state, add_progress_handler, WebUIProgressHandler
+from comfy_execution.utils import CurrentNodeContext
 
 
 class ExecutionResult(Enum):
@@ -39,12 +43,13 @@ class DuplicateNodeError(Exception):
     pass
 
 class IsChangedCache:
-    def __init__(self, dynprompt, outputs_cache):
+    def __init__(self, prompt_id: str, dynprompt: DynamicPrompt, outputs_cache: BasicCache):
+        self.prompt_id = prompt_id
         self.dynprompt = dynprompt
         self.outputs_cache = outputs_cache
         self.is_changed = {}
 
-    def get(self, node_id):
+    async def get(self, node_id):
         if node_id in self.is_changed:
             return self.is_changed[node_id]
 
@@ -62,7 +67,8 @@ class IsChangedCache:
         # Intentionally do not use cached outputs here. We only want constants in IS_CHANGED
         input_data_all, _ = get_input_data(node["inputs"], class_def, node_id, None)
         try:
-            is_changed = _map_node_over_list(class_def, input_data_all, "IS_CHANGED")
+            is_changed = await _async_map_node_over_list(self.prompt_id, node_id, class_def, input_data_all, "IS_CHANGED")
+            is_changed = await resolve_map_node_over_list_results(is_changed)
             node["is_changed"] = [None if isinstance(x, ExecutionBlocker) else x for x in is_changed]
         except Exception as e:
             logging.warning("WARNING: {}".format(e))
@@ -164,7 +170,19 @@ def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, e
 
 map_node_over_list = None #Don't hook this please
 
-def _map_node_over_list(obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None):
+async def resolve_map_node_over_list_results(results):
+    remaining = [x for x in results if isinstance(x, asyncio.Task) and not x.done()]
+    if len(remaining) == 0:
+        return [x.result() if isinstance(x, asyncio.Task) else x for x in results]
+    else:
+        done, pending = await asyncio.wait(remaining)
+        for task in done:
+            exc = task.exception()
+            if exc is not None:
+                raise exc
+        return [x.result() if isinstance(x, asyncio.Task) else x for x in results]
+
+async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None):
     # check if node wants the lists
     input_is_list = getattr(obj, "INPUT_IS_LIST", False)
 
@@ -178,7 +196,7 @@ def _map_node_over_list(obj, input_data_all, func, allow_interrupt=False, execut
         return {k: v[i if len(v) > i else -1] for k, v in d.items()}
 
     results = []
-    def process_inputs(inputs, index=None, input_is_list=False):
+    async def process_inputs(inputs, index=None, input_is_list=False):
         if allow_interrupt:
             nodes.before_node_execution()
         execution_block = None
@@ -194,20 +212,37 @@ def _map_node_over_list(obj, input_data_all, func, allow_interrupt=False, execut
         if execution_block is None:
             if pre_execute_cb is not None and index is not None:
                 pre_execute_cb(index)
-            results.append(getattr(obj, func)(**inputs))
+            f = getattr(obj, func)
+            if inspect.iscoroutinefunction(f):
+                async def async_wrapper(f, prompt_id, unique_id, list_index, args):
+                    with CurrentNodeContext(prompt_id, unique_id, list_index):
+                        return await f(**args)
+                task = asyncio.create_task(async_wrapper(f, prompt_id, unique_id, index, args=inputs))
+                # Give the task a chance to execute without yielding
+                await asyncio.sleep(0)
+                if task.done():
+                    result = task.result()
+                    results.append(result)
+                else:
+                    results.append(task)
+            else:
+                with CurrentNodeContext(prompt_id, unique_id, index):
+                    result = f(**inputs)
+                results.append(result)
         else:
             results.append(execution_block)
 
     if input_is_list:
-        process_inputs(input_data_all, 0, input_is_list=input_is_list)
+        await process_inputs(input_data_all, 0, input_is_list=input_is_list)
     elif max_len_input == 0:
-        process_inputs({})
+        await process_inputs({})
     else:
         for i in range(max_len_input):
             input_dict = slice_dict(input_data_all, i)
-            process_inputs(input_dict, i)
+            await process_inputs(input_dict, i)
     return results
 
+
 def merge_result_data(results, obj):
     # check which outputs need concatenating
     output = []
@@ -229,11 +264,18 @@ def merge_result_data(results, obj):
             output.append([o[i] for o in results])
     return output
 
-def get_output_data(obj, input_data_all, execution_block_cb=None, pre_execute_cb=None):
+async def get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=None, pre_execute_cb=None):
+    return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
+    has_pending_task = any(isinstance(r, asyncio.Task) and not r.done() for r in return_values)
+    if has_pending_task:
+        return return_values, {}, False, has_pending_task
+    output, ui, has_subgraph = get_output_from_returns(return_values, obj)
+    return output, ui, has_subgraph, False
+
+def get_output_from_returns(return_values, obj):
     results = []
     uis = []
     subgraph_results = []
-    return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
     has_subgraph = False
     for i in range(len(return_values)):
         r = return_values[i]
@@ -267,6 +309,10 @@ def get_output_data(obj, input_data_all, execution_block_cb=None, pre_execute_cb
     else:
         output = []
     ui = dict()
+    # TODO: Think there's an existing bug here
+    # If we're performing a subgraph expansion, we probably shouldn't be returning UI values yet.
+    # They'll get cached without the completed subgraphs. It's an edge case and I'm not aware of
+    # any nodes that use both subgraph expansion and custom UI outputs, but might be a problem in the future.
     if len(uis) > 0:
         ui = {k: [y for x in uis for y in x[k]] for k in uis[0].keys()}
     return output, ui, has_subgraph
@@ -279,7 +325,7 @@ def format_value(x):
     else:
         return str(x)
 
-def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results):
+async def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes):
     unique_id = current_item
     real_node_id = dynprompt.get_real_node_id(unique_id)
     display_node_id = dynprompt.get_display_node_id(unique_id)
@@ -291,11 +337,26 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
         if server.client_id is not None:
             cached_output = caches.ui.get(unique_id) or {}
             server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id)
+        get_progress_state().finish_progress(unique_id)
         return (ExecutionResult.SUCCESS, None, None)
 
     input_data_all = None
     try:
-        if unique_id in pending_subgraph_results:
+        if unique_id in pending_async_nodes:
+            results = []
+            for r in pending_async_nodes[unique_id]:
+                if isinstance(r, asyncio.Task):
+                    try:
+                        results.append(r.result())
+                    except Exception as ex:
+                        # An async task failed - propagate the exception up
+                        del pending_async_nodes[unique_id]
+                        raise ex
+                else:
+                    results.append(r)
+            del pending_async_nodes[unique_id]
+            output_data, output_ui, has_subgraph = get_output_from_returns(results, class_def)
+        elif unique_id in pending_subgraph_results:
             cached_results = pending_subgraph_results[unique_id]
             resolved_outputs = []
             for is_subgraph, result in cached_results:
@@ -317,6 +378,7 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
             output_ui = []
             has_subgraph = False
         else:
+            get_progress_state().start_progress(unique_id)
             input_data_all, missing_keys = get_input_data(inputs, class_def, unique_id, caches.outputs, dynprompt, extra_data)
             if server.client_id is not None:
                 server.last_node_id = display_node_id
@@ -328,7 +390,8 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
                 caches.objects.set(unique_id, obj)
 
             if hasattr(obj, "check_lazy_status"):
-                required_inputs = _map_node_over_list(obj, input_data_all, "check_lazy_status", allow_interrupt=True)
+                required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True)
+                required_inputs = await resolve_map_node_over_list_results(required_inputs)
                 required_inputs = set(sum([r for r in required_inputs if isinstance(r,list)], []))
                 required_inputs = [x for x in required_inputs if isinstance(x,str) and (
                     x not in input_data_all or x in missing_keys
@@ -357,8 +420,18 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
                 else:
                     return block
             def pre_execute_cb(call_index):
+                # TODO - How to handle this with async functions without contextvars (which requires Python 3.12)?
                 GraphBuilder.set_default_prefix(unique_id, call_index, 0)
-            output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
+            output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
+            if has_pending_tasks:
+                pending_async_nodes[unique_id] = output_data
+                unblock = execution_list.add_external_block(unique_id)
+                async def await_completion():
+                    tasks = [x for x in output_data if isinstance(x, asyncio.Task)]
+                    await asyncio.gather(*tasks, return_exceptions=True)
+                    unblock()
+                asyncio.create_task(await_completion())
+                return (ExecutionResult.PENDING, None, None)
         if len(output_ui) > 0:
             caches.ui.set(unique_id, {
                 "meta": {
@@ -401,7 +474,8 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
                     cached_outputs.append((True, node_outputs))
             new_node_ids = set(new_node_ids)
             for cache in caches.all:
-                cache.ensure_subcache_for(unique_id, new_node_ids).clean_unused()
+                subcache = await cache.ensure_subcache_for(unique_id, new_node_ids)
+                subcache.clean_unused()
             for node_id in new_output_ids:
                 execution_list.add_node(node_id)
             for link in new_output_links:
@@ -429,20 +503,24 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
 
         logging.error(f"!!! Exception during processing !!! {ex}")
         logging.error(traceback.format_exc())
+        tips = ""
+
+        if isinstance(ex, comfy.model_management.OOM_EXCEPTION):
+            tips = "This error means you ran out of memory on your GPU.\n\nTIPS: If the workflow worked before you might have accidentally set the batch_size to a large number."
+            logging.error("Got an OOM, unloading all loaded models.")
+            comfy.model_management.unload_all_models()
 
         error_details = {
             "node_id": real_node_id,
-            "exception_message": str(ex),
+            "exception_message": "{}\n{}".format(ex, tips),
             "exception_type": exception_type,
             "traceback": traceback.format_tb(tb),
             "current_inputs": input_data_formatted
         }
-        if isinstance(ex, comfy.model_management.OOM_EXCEPTION):
-            logging.error("Got an OOM, unloading all loaded models.")
-            comfy.model_management.unload_all_models()
 
         return (ExecutionResult.FAILURE, error_details, ex)
 
+    get_progress_state().finish_progress(unique_id)
     executed.add(unique_id)
 
     return (ExecutionResult.SUCCESS, None, None)
@@ -497,6 +575,11 @@ class PromptExecutor:
             self.add_message("execution_error", mes, broadcast=False)
 
     def execute(self, prompt, prompt_id, extra_data={}, execute_outputs=[]):
+        asyncio_loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(asyncio_loop)
+        asyncio.run(self.execute_async(prompt, prompt_id, extra_data, execute_outputs))
+
+    async def execute_async(self, prompt, prompt_id, extra_data={}, execute_outputs=[]):
         nodes.interrupt_processing(False)
 
         if "client_id" in extra_data:
@@ -509,9 +592,11 @@ class PromptExecutor:
 
         with torch.inference_mode():
             dynamic_prompt = DynamicPrompt(prompt)
-            is_changed_cache = IsChangedCache(dynamic_prompt, self.caches.outputs)
+            reset_progress_state(prompt_id, dynamic_prompt)
+            add_progress_handler(WebUIProgressHandler(self.server))
+            is_changed_cache = IsChangedCache(prompt_id, dynamic_prompt, self.caches.outputs)
             for cache in self.caches.all:
-                cache.set_prompt(dynamic_prompt, prompt.keys(), is_changed_cache)
+                await cache.set_prompt(dynamic_prompt, prompt.keys(), is_changed_cache)
                 cache.clean_unused()
 
             cached_nodes = []
@@ -524,6 +609,7 @@ class PromptExecutor:
                           { "nodes": cached_nodes, "prompt_id": prompt_id},
                           broadcast=False)
             pending_subgraph_results = {}
+            pending_async_nodes = {} # TODO - Unify this with pending_subgraph_results
             executed = set()
             execution_list = ExecutionList(dynamic_prompt, self.caches.outputs)
             current_outputs = self.caches.outputs.all_node_ids()
@@ -531,12 +617,13 @@ class PromptExecutor:
                 execution_list.add_node(node_id)
 
             while not execution_list.is_empty():
-                node_id, error, ex = execution_list.stage_node_execution()
+                node_id, error, ex = await execution_list.stage_node_execution()
                 if error is not None:
                     self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
                     break
 
-                result, error, ex = execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results)
+                assert node_id is not None, "Node ID should not be None at this point"
+                result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes)
                 self.success = result != ExecutionResult.FAILURE
                 if result == ExecutionResult.FAILURE:
                     self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
@@ -566,7 +653,7 @@ class PromptExecutor:
                 comfy.model_management.unload_all_models()
 
 
-def validate_inputs(prompt, item, validated):
+async def validate_inputs(prompt_id, prompt, item, validated):
     unique_id = item
     if unique_id in validated:
         return validated[unique_id]
@@ -643,7 +730,7 @@ def validate_inputs(prompt, item, validated):
                 errors.append(error)
                 continue
             try:
-                r = validate_inputs(prompt, o_id, validated)
+                r = await validate_inputs(prompt_id, prompt, o_id, validated)
                 if r[0] is False:
                     # `r` will be set in `validated[o_id]` already
                     valid = False
@@ -768,7 +855,8 @@ def validate_inputs(prompt, item, validated):
             input_filtered['input_types'] = [received_types]
 
         #ret = obj_class.VALIDATE_INPUTS(**input_filtered)
-        ret = _map_node_over_list(obj_class, input_filtered, "VALIDATE_INPUTS")
+        ret = await _async_map_node_over_list(prompt_id, unique_id, obj_class, input_filtered, "VALIDATE_INPUTS")
+        ret = await resolve_map_node_over_list_results(ret)
         for x in input_filtered:
             for i, r in enumerate(ret):
                 if r is not True and not isinstance(r, ExecutionBlocker):
@@ -801,7 +889,7 @@ def full_type_name(klass):
         return klass.__qualname__
     return module + '.' + klass.__qualname__
 
-def validate_prompt(prompt):
+async def validate_prompt(prompt_id, prompt):
     outputs = set()
     for x in prompt:
         if 'class_type' not in prompt[x]:
@@ -844,7 +932,7 @@ def validate_prompt(prompt):
         valid = False
         reasons = []
         try:
-            m = validate_inputs(prompt, o, validated)
+            m = await validate_inputs(prompt_id, prompt, o, validated)
             valid = m[0]
             reasons = m[1]
         except Exception as ex:

main.py

@@ -11,6 +11,8 @@ import itertools
 import utils.extra_config
 import logging
 import sys
+from comfy_execution.progress import get_progress_state
+from comfy_execution.utils import get_executing_context
 
 if __name__ == "__main__":
     #NOTE: These do not do anything on core ComfyUI, they are for custom nodes.
@@ -55,6 +57,9 @@ def apply_custom_paths():
 
 
 def execute_prestartup_script():
+    if args.disable_all_custom_nodes and len(args.whitelist_custom_nodes) == 0:
+        return
+
     def execute_script(script_path):
         module_name = os.path.splitext(script_path)[0]
         try:
@@ -66,9 +71,6 @@ def execute_prestartup_script():
             logging.error(f"Failed to execute startup-script: {script_path} / {e}")
         return False
 
-    if args.disable_all_custom_nodes:
-        return
-
     node_paths = folder_paths.get_folder_paths("custom_nodes")
     for custom_node_path in node_paths:
         possible_modules = os.listdir(custom_node_path)
@@ -81,6 +83,9 @@ def execute_prestartup_script():
 
             script_path = os.path.join(module_path, "prestartup_script.py")
             if os.path.exists(script_path):
+                if args.disable_all_custom_nodes and possible_module not in args.whitelist_custom_nodes:
+                    logging.info(f"Prestartup Skipping {possible_module} due to disable_all_custom_nodes and whitelist_custom_nodes")
+                    continue
                 time_before = time.perf_counter()
                 success = execute_script(script_path)
                 node_prestartup_times.append((time.perf_counter() - time_before, module_path, success))
@@ -128,7 +133,7 @@ import comfy.utils
 
 import execution
 import server
-from server import BinaryEventTypes
+from protocol import BinaryEventTypes
 import nodes
 import comfy.model_management
 import comfyui_version
@@ -185,7 +190,13 @@ def prompt_worker(q, server_instance):
 
             current_time = time.perf_counter()
             execution_time = current_time - execution_start_time
-            logging.info("Prompt executed in {:.2f} seconds".format(execution_time))
+
+            # Log Time in a more readable way after 10 minutes
+            if execution_time > 600:
+                execution_time = time.strftime("%H:%M:%S", time.gmtime(execution_time))
+                logging.info(f"Prompt executed in {execution_time}")
+            else:
+                logging.info("Prompt executed in {:.2f} seconds".format(execution_time))
 
         flags = q.get_flags()
         free_memory = flags.get("free_memory", False)
@@ -218,14 +229,25 @@ async def run(server_instance, address='', port=8188, verbose=True, call_on_star
         server_instance.start_multi_address(addresses, call_on_start, verbose), server_instance.publish_loop()
     )
 
-
 def hijack_progress(server_instance):
-    def hook(value, total, preview_image):
+    def hook(value, total, preview_image, prompt_id=None, node_id=None):
+        executing_context = get_executing_context()
+        if prompt_id is None and executing_context is not None:
+            prompt_id = executing_context.prompt_id
+        if node_id is None and executing_context is not None:
+            node_id = executing_context.node_id
         comfy.model_management.throw_exception_if_processing_interrupted()
-        progress = {"value": value, "max": total, "prompt_id": server_instance.last_prompt_id, "node": server_instance.last_node_id}
+        if prompt_id is None:
+            prompt_id = server_instance.last_prompt_id
+        if node_id is None:
+            node_id = server_instance.last_node_id
+        progress = {"value": value, "max": total, "prompt_id": prompt_id, "node": node_id}
+        get_progress_state().update_progress(node_id, value, total, preview_image)
 
         server_instance.send_sync("progress", progress, server_instance.client_id)
         if preview_image is not None:
+            # Also send old method for backward compatibility
+            # TODO - Remove after this repo is updated to frontend with metadata support
             server_instance.send_sync(BinaryEventTypes.UNENCODED_PREVIEW_IMAGE, preview_image, server_instance.client_id)
 
     comfy.utils.set_progress_bar_global_hook(hook)
@@ -270,7 +292,10 @@ def start_comfyui(asyncio_loop=None):
     prompt_server = server.PromptServer(asyncio_loop)
 
     hook_breaker_ac10a0.save_functions()
-    nodes.init_extra_nodes(init_custom_nodes=not args.disable_all_custom_nodes, init_api_nodes=not args.disable_api_nodes)
+    nodes.init_extra_nodes(
+        init_custom_nodes=(not args.disable_all_custom_nodes) or len(args.whitelist_custom_nodes) > 0,
+        init_api_nodes=not args.disable_api_nodes
+    )
     hook_breaker_ac10a0.restore_functions()
 
     cuda_malloc_warning()

nodes.py

@@ -920,7 +920,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),
@@ -930,7 +930,7 @@ class CLIPLoader:
 
     CATEGORY = "advanced/loaders"
 
-    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5"
+    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl\n hidream: llama-3.1 (Recommend) or t5\nomnigen2: qwen vl 2.5 3B"
 
     def load_clip(self, clip_name, type="stable_diffusion", device="default"):
         clip_type = getattr(comfy.sd.CLIPType, type.upper(), comfy.sd.CLIPType.STABLE_DIFFUSION)
@@ -2187,6 +2187,9 @@ def init_external_custom_nodes():
             module_path = os.path.join(custom_node_path, possible_module)
             if os.path.isfile(module_path) and os.path.splitext(module_path)[1] != ".py": continue
             if module_path.endswith(".disabled"): continue
+            if args.disable_all_custom_nodes and possible_module not in args.whitelist_custom_nodes:
+                logging.info(f"Skipping {possible_module} due to disable_all_custom_nodes and whitelist_custom_nodes")
+                continue
             time_before = time.perf_counter()
             success = load_custom_node(module_path, base_node_names, module_parent="custom_nodes")
             node_import_times.append((time.perf_counter() - time_before, module_path, success))
@@ -2279,6 +2282,8 @@ def init_builtin_extra_nodes():
         "nodes_ace.py",
         "nodes_string.py",
         "nodes_camera_trajectory.py",
+        "nodes_edit_model.py",
+        "nodes_tcfg.py"
     ]
 
     import_failed = []

protocol.py

@@ -0,0 +1,7 @@
+
+class BinaryEventTypes:
+    PREVIEW_IMAGE = 1
+    UNENCODED_PREVIEW_IMAGE = 2
+    TEXT = 3
+    PREVIEW_IMAGE_WITH_METADATA = 4
+

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.41"
+version = "0.3.43"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,13 +1,13 @@
-comfyui-frontend-package==1.21.7
-comfyui-workflow-templates==0.1.28
-comfyui-embedded-docs==0.2.2
+comfyui-frontend-package==1.23.4
+comfyui-workflow-templates==0.1.31
+comfyui-embedded-docs==0.2.3
 torch
 torchsde
 torchvision
 torchaudio
 numpy>=1.25.0
 einops
-transformers>=4.28.1
+transformers>=4.37.2
 tokenizers>=0.13.3
 sentencepiece
 safetensors>=0.4.2

server.py

@@ -35,11 +35,7 @@ from app.model_manager import ModelFileManager
 from app.custom_node_manager import CustomNodeManager
 from typing import Optional, Union
 from api_server.routes.internal.internal_routes import InternalRoutes
-
-class BinaryEventTypes:
-    PREVIEW_IMAGE = 1
-    UNENCODED_PREVIEW_IMAGE = 2
-    TEXT = 3
+from protocol import BinaryEventTypes
 
 async def send_socket_catch_exception(function, message):
     try:
@@ -643,7 +639,8 @@ class PromptServer():
 
             if "prompt" in json_data:
                 prompt = json_data["prompt"]
-                valid = execution.validate_prompt(prompt)
+                prompt_id = str(uuid.uuid4())
+                valid = await execution.validate_prompt(prompt_id, prompt)
                 extra_data = {}
                 if "extra_data" in json_data:
                     extra_data = json_data["extra_data"]
@@ -651,7 +648,6 @@ class PromptServer():
                 if "client_id" in json_data:
                     extra_data["client_id"] = json_data["client_id"]
                 if valid[0]:
-                    prompt_id = str(uuid.uuid4())
                     outputs_to_execute = valid[2]
                     self.prompt_queue.put((number, prompt_id, prompt, extra_data, outputs_to_execute))
                     response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]}
@@ -766,6 +762,10 @@ class PromptServer():
     async def send(self, event, data, sid=None):
         if event == BinaryEventTypes.UNENCODED_PREVIEW_IMAGE:
             await self.send_image(data, sid=sid)
+        elif event == BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA:
+            # data is (preview_image, metadata)
+            preview_image, metadata = data
+            await self.send_image_with_metadata(preview_image, metadata, sid=sid)
         elif isinstance(data, (bytes, bytearray)):
             await self.send_bytes(event, data, sid)
         else:
@@ -804,6 +804,43 @@ class PromptServer():
         preview_bytes = bytesIO.getvalue()
         await self.send_bytes(BinaryEventTypes.PREVIEW_IMAGE, preview_bytes, sid=sid)
 
+    async def send_image_with_metadata(self, image_data, metadata=None, sid=None):
+        image_type = image_data[0]
+        image = image_data[1]
+        max_size = image_data[2]
+        if max_size is not None:
+            if hasattr(Image, 'Resampling'):
+                resampling = Image.Resampling.BILINEAR
+            else:
+                resampling = Image.Resampling.LANCZOS
+
+            image = ImageOps.contain(image, (max_size, max_size), resampling)
+
+        mimetype = "image/png" if image_type == "PNG" else "image/jpeg"
+
+        # Prepare metadata
+        if metadata is None:
+            metadata = {}
+        metadata["image_type"] = mimetype
+
+        # Serialize metadata as JSON
+        import json
+        metadata_json = json.dumps(metadata).encode('utf-8')
+        metadata_length = len(metadata_json)
+
+        # Prepare image data
+        bytesIO = BytesIO()
+        image.save(bytesIO, format=image_type, quality=95, compress_level=1)
+        image_bytes = bytesIO.getvalue()
+
+        # Combine metadata and image
+        combined_data = bytearray()
+        combined_data.extend(struct.pack(">I", metadata_length))
+        combined_data.extend(metadata_json)
+        combined_data.extend(image_bytes)
+
+        await self.send_bytes(BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA, combined_data, sid=sid)
+
     async def send_bytes(self, event, data, sid=None):
         message = self.encode_bytes(event, data)
 

tests-unit/requirements.txt

@@ -1,3 +1,4 @@
 pytest>=7.8.0
 pytest-aiohttp
 pytest-asyncio
+websocket-client

tests/inference/extra_model_paths.yaml

@@ -1,4 +1,4 @@
 # Config for testing nodes
 testing:
-    custom_nodes: tests/inference/testing_nodes
+    custom_nodes: testing_nodes
 

tests/inference/test_async_nodes.py

@@ -0,0 +1,410 @@
+import pytest
+import time
+import torch
+import urllib.error
+import numpy as np
+import subprocess
+
+from pytest import fixture
+from comfy_execution.graph_utils import GraphBuilder
+from tests.inference.test_execution import ComfyClient
+
+
+@pytest.mark.execution
+class TestAsyncNodes:
+    @fixture(scope="class", autouse=True, params=[
+        (False, 0),
+        (True, 0),
+        (True, 100),
+    ])
+    def _server(self, args_pytest, request):
+        pargs = [
+            'python','main.py',
+            '--output-directory', args_pytest["output_dir"],
+            '--listen', args_pytest["listen"],
+            '--port', str(args_pytest["port"]),
+            '--extra-model-paths-config', 'tests/inference/extra_model_paths.yaml',
+        ]
+        use_lru, lru_size = request.param
+        if use_lru:
+            pargs += ['--cache-lru', str(lru_size)]
+        # Running server with args: pargs
+        p = subprocess.Popen(pargs)
+        yield
+        p.kill()
+        torch.cuda.empty_cache()
+
+    @fixture(scope="class", autouse=True)
+    def shared_client(self, args_pytest, _server):
+        client = ComfyClient()
+        n_tries = 5
+        for i in range(n_tries):
+            time.sleep(4)
+            try:
+                client.connect(listen=args_pytest["listen"], port=args_pytest["port"])
+            except ConnectionRefusedError:
+                # Retrying...
+                pass
+            else:
+                break
+        yield client
+        del client
+        torch.cuda.empty_cache()
+
+    @fixture
+    def client(self, shared_client, request):
+        shared_client.set_test_name(f"async_nodes[{request.node.name}]")
+        yield shared_client
+
+    @fixture
+    def builder(self, request):
+        yield GraphBuilder(prefix=request.node.name)
+
+    # Happy Path Tests
+
+    def test_basic_async_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that a basic async node executes correctly."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        sleep_node = g.node("TestSleep", value=image.out(0), seconds=0.1)
+        output = g.node("SaveImage", images=sleep_node.out(0))
+
+        result = client.run(g)
+
+        # Verify execution completed
+        assert result.did_run(sleep_node), "Async sleep node should have executed"
+        assert result.did_run(output), "Output node should have executed"
+
+        # Verify the image passed through correctly
+        result_images = result.get_images(output)
+        assert len(result_images) == 1, "Should have 1 image"
+        assert np.array(result_images[0]).min() == 0 and np.array(result_images[0]).max() == 0, "Image should be black"
+
+    def test_multiple_async_parallel_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that multiple async nodes execute in parallel."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Create multiple async sleep nodes with different durations
+        sleep1 = g.node("TestSleep", value=image.out(0), seconds=0.3)
+        sleep2 = g.node("TestSleep", value=image.out(0), seconds=0.4)
+        sleep3 = g.node("TestSleep", value=image.out(0), seconds=0.5)
+
+        # Add outputs for each
+        _output1 = g.node("PreviewImage", images=sleep1.out(0))
+        _output2 = g.node("PreviewImage", images=sleep2.out(0))
+        _output3 = g.node("PreviewImage", images=sleep3.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Should take ~0.5s (max duration) not 1.2s (sum of durations)
+        assert elapsed_time < 0.8, f"Parallel execution took {elapsed_time}s, expected < 0.8s"
+
+        # Verify all nodes executed
+        assert result.did_run(sleep1) and result.did_run(sleep2) and result.did_run(sleep3)
+
+    def test_async_with_dependencies(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes with proper dependency handling."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Chain of async operations
+        sleep1 = g.node("TestSleep", value=image1.out(0), seconds=0.2)
+        sleep2 = g.node("TestSleep", value=image2.out(0), seconds=0.2)
+
+        # Average depends on both async results
+        average = g.node("TestVariadicAverage", input1=sleep1.out(0), input2=sleep2.out(0))
+        output = g.node("SaveImage", images=average.out(0))
+
+        result = client.run(g)
+
+        # Verify execution order
+        assert result.did_run(sleep1) and result.did_run(sleep2)
+        assert result.did_run(average) and result.did_run(output)
+
+        # Verify averaged result
+        result_images = result.get_images(output)
+        avg_value = np.array(result_images[0]).mean()
+        assert abs(avg_value - 127.5) < 1, f"Average value {avg_value} should be ~127.5"
+
+    def test_async_validate_inputs(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async VALIDATE_INPUTS function."""
+        g = builder
+        # Create a test node with async validation
+        validation_node = g.node("TestAsyncValidation", value=5.0, threshold=10.0)
+        g.node("SaveImage", images=validation_node.out(0))
+
+        # Should pass validation
+        result = client.run(g)
+        assert result.did_run(validation_node)
+
+        # Test validation failure
+        validation_node.inputs['threshold'] = 3.0  # Will fail since value > threshold
+        with pytest.raises(urllib.error.HTTPError):
+            client.run(g)
+
+    def test_async_lazy_evaluation(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes with lazy evaluation."""
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.0, height=512, width=512, batch_size=1)
+
+        # Create async nodes that will be evaluated lazily
+        sleep1 = g.node("TestSleep", value=input1.out(0), seconds=0.3)
+        sleep2 = g.node("TestSleep", value=input2.out(0), seconds=0.3)
+
+        # Use lazy mix that only needs sleep1 (mask=0.0)
+        lazy_mix = g.node("TestLazyMixImages", image1=sleep1.out(0), image2=sleep2.out(0), mask=mask.out(0))
+        g.node("SaveImage", images=lazy_mix.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Should only execute sleep1, not sleep2
+        assert elapsed_time < 0.5, f"Should skip sleep2, took {elapsed_time}s"
+        assert result.did_run(sleep1), "Sleep1 should have executed"
+        assert not result.did_run(sleep2), "Sleep2 should have been skipped"
+
+    def test_async_check_lazy_status(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async check_lazy_status function."""
+        g = builder
+        # Create a node with async check_lazy_status
+        lazy_node = g.node("TestAsyncLazyCheck",
+                          input1="value1",
+                          input2="value2",
+                          condition=True)
+        g.node("SaveImage", images=lazy_node.out(0))
+
+        result = client.run(g)
+        assert result.did_run(lazy_node)
+
+    # Error Handling Tests
+
+    def test_async_execution_error(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that async execution errors are properly handled."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        # Create an async node that will error
+        error_node = g.node("TestAsyncError", value=image.out(0), error_after=0.1)
+        g.node("SaveImage", images=error_node.out(0))
+
+        try:
+            client.run(g)
+            assert False, "Should have raised an error"
+        except Exception as e:
+            assert 'prompt_id' in e.args[0], f"Did not get proper error message: {e}"
+            assert e.args[0]['node_id'] == error_node.id, "Error should be from async error node"
+
+    def test_async_validation_error(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async validation error handling."""
+        g = builder
+        # Node with async validation that will fail
+        validation_node = g.node("TestAsyncValidationError", value=15.0, max_value=10.0)
+        g.node("SaveImage", images=validation_node.out(0))
+
+        with pytest.raises(urllib.error.HTTPError) as exc_info:
+            client.run(g)
+        # Verify it's a validation error
+        assert exc_info.value.code == 400
+
+    def test_async_timeout_handling(self, client: ComfyClient, builder: GraphBuilder):
+        """Test handling of async operations that timeout."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        # Very long sleep that would timeout
+        timeout_node = g.node("TestAsyncTimeout", value=image.out(0), timeout=0.5, operation_time=2.0)
+        g.node("SaveImage", images=timeout_node.out(0))
+
+        try:
+            client.run(g)
+            assert False, "Should have raised a timeout error"
+        except Exception as e:
+            assert 'timeout' in str(e).lower(), f"Expected timeout error, got: {e}"
+
+    def test_concurrent_async_error_recovery(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that workflow can recover after async errors."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # First run with error
+        error_node = g.node("TestAsyncError", value=image.out(0), error_after=0.1)
+        g.node("SaveImage", images=error_node.out(0))
+
+        try:
+            client.run(g)
+        except Exception:
+            pass  # Expected
+
+        # Second run should succeed
+        g2 = GraphBuilder(prefix="recovery_test")
+        image2 = g2.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        sleep_node = g2.node("TestSleep", value=image2.out(0), seconds=0.1)
+        g2.node("SaveImage", images=sleep_node.out(0))
+
+        result = client.run(g2)
+        assert result.did_run(sleep_node), "Should be able to run after error"
+
+    def test_sync_error_during_async_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test handling when sync node errors while async node is executing."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Async node that takes time
+        sleep_node = g.node("TestSleep", value=image.out(0), seconds=0.5)
+
+        # Sync node that will error immediately
+        error_node = g.node("TestSyncError", value=image.out(0))
+
+        # Both feed into output
+        g.node("PreviewImage", images=sleep_node.out(0))
+        g.node("PreviewImage", images=error_node.out(0))
+
+        try:
+            client.run(g)
+            assert False, "Should have raised an error"
+        except Exception as e:
+            # Verify the sync error was caught even though async was running
+            assert 'prompt_id' in e.args[0]
+
+    # Edge Cases
+
+    def test_async_with_execution_blocker(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes with execution blockers."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Async sleep nodes
+        sleep1 = g.node("TestSleep", value=image1.out(0), seconds=0.2)
+        sleep2 = g.node("TestSleep", value=image2.out(0), seconds=0.2)
+
+        # Create list of images
+        image_list = g.node("TestMakeListNode", value1=sleep1.out(0), value2=sleep2.out(0))
+
+        # Create list of blocking conditions - [False, True] to block only the second item
+        int1 = g.node("StubInt", value=1)
+        int2 = g.node("StubInt", value=2)
+        block_list = g.node("TestMakeListNode", value1=int1.out(0), value2=int2.out(0))
+
+        # Compare each value against 2, so first is False (1 != 2) and second is True (2 == 2)
+        compare = g.node("TestIntConditions", a=block_list.out(0), b=2, operation="==")
+
+        # Block based on the comparison results
+        blocker = g.node("TestExecutionBlocker", input=image_list.out(0), block=compare.out(0), verbose=False)
+
+        output = g.node("PreviewImage", images=blocker.out(0))
+
+        result = client.run(g)
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have blocked second image"
+
+    def test_async_caching_behavior(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that async nodes are properly cached."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        sleep_node = g.node("TestSleep", value=image.out(0), seconds=0.2)
+        g.node("SaveImage", images=sleep_node.out(0))
+
+        # First run
+        result1 = client.run(g)
+        assert result1.did_run(sleep_node), "Should run first time"
+
+        # Second run - should be cached
+        start_time = time.time()
+        result2 = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        assert not result2.did_run(sleep_node), "Should be cached"
+        assert elapsed_time < 0.1, f"Cached run took {elapsed_time}s, should be instant"
+
+    def test_async_with_dynamic_prompts(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes within dynamically generated prompts."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Node that generates async nodes dynamically
+        dynamic_async = g.node("TestDynamicAsyncGeneration",
+                              image1=image1.out(0),
+                              image2=image2.out(0),
+                              num_async_nodes=3,
+                              sleep_duration=0.2)
+        g.node("SaveImage", images=dynamic_async.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Should execute async nodes in parallel within dynamic prompt
+        assert elapsed_time < 0.5, f"Dynamic async execution took {elapsed_time}s"
+        assert result.did_run(dynamic_async)
+
+    def test_async_resource_cleanup(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that async resources are properly cleaned up."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Create multiple async nodes that use resources
+        resource_nodes = []
+        for i in range(5):
+            node = g.node("TestAsyncResourceUser",
+                         value=image.out(0),
+                         resource_id=f"resource_{i}",
+                         duration=0.1)
+            resource_nodes.append(node)
+            g.node("PreviewImage", images=node.out(0))
+
+        result = client.run(g)
+
+        # Verify all nodes executed
+        for node in resource_nodes:
+            assert result.did_run(node)
+
+        # Run again to ensure resources were cleaned up
+        result2 = client.run(g)
+        # Should be cached but not error due to resource conflicts
+        for node in resource_nodes:
+            assert not result2.did_run(node), "Should be cached"
+
+    def test_async_cancellation(self, client: ComfyClient, builder: GraphBuilder):
+        """Test cancellation of async operations."""
+        # This would require implementing cancellation in the client
+        # For now, we'll test that long-running async operations can be interrupted
+        pass  # TODO: Implement when cancellation API is available
+
+    def test_mixed_sync_async_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test workflows with both sync and async nodes."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+
+        # Mix of sync and async operations
+        # Sync: lazy mix images
+        sync_op1 = g.node("TestLazyMixImages", image1=image1.out(0), image2=image2.out(0), mask=mask.out(0))
+        # Async: sleep
+        async_op1 = g.node("TestSleep", value=sync_op1.out(0), seconds=0.2)
+        # Sync: custom validation
+        sync_op2 = g.node("TestCustomValidation1", input1=async_op1.out(0), input2=0.5)
+        # Async: sleep again
+        async_op2 = g.node("TestSleep", value=sync_op2.out(0), seconds=0.2)
+
+        output = g.node("SaveImage", images=async_op2.out(0))
+
+        result = client.run(g)
+
+        # Verify all nodes executed in correct order
+        assert result.did_run(sync_op1)
+        assert result.did_run(async_op1)
+        assert result.did_run(sync_op2)
+        assert result.did_run(async_op2)
+
+        # Image should be a mix of black and white (gray)
+        result_images = result.get_images(output)
+        avg_value = np.array(result_images[0]).mean()
+        assert abs(avg_value - 63.75) < 5, f"Average value {avg_value} should be ~63.75"

tests/inference/test_execution.py

@@ -252,7 +252,7 @@ class TestExecution:
 
     @pytest.mark.parametrize("test_type, test_value", [
         ("StubInt", 5),
-        ("StubFloat", 5.0)
+        ("StubMask", 5.0)
     ])
     def test_validation_error_edge1(self, test_type, test_value, client: ComfyClient, builder: GraphBuilder):
         g = builder
@@ -497,6 +497,69 @@ class TestExecution:
         assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25"
         assert not result.did_run(test_node), "The execution should have been cached"
 
+    def test_parallel_sleep_nodes(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Create sleep nodes for each duration
+        sleep_node1 = g.node("TestSleep", value=image.out(0), seconds=2.8)
+        sleep_node2 = g.node("TestSleep", value=image.out(0), seconds=2.9)
+        sleep_node3 = g.node("TestSleep", value=image.out(0), seconds=3.0)
+
+        # Add outputs to verify the execution
+        _output1 = g.node("PreviewImage", images=sleep_node1.out(0))
+        _output2 = g.node("PreviewImage", images=sleep_node2.out(0))
+        _output3 = g.node("PreviewImage", images=sleep_node3.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # The test should take around 0.4 seconds (the longest sleep duration)
+        # plus some overhead, but definitely less than the sum of all sleeps (0.9s)
+        # We'll allow for up to 0.8s total to account for overhead
+        assert elapsed_time < 4.0, f"Parallel execution took {elapsed_time}s, expected less than 0.8s"
+
+        # Verify that all nodes executed
+        assert result.did_run(sleep_node1), "Sleep node 1 should have run"
+        assert result.did_run(sleep_node2), "Sleep node 2 should have run"
+        assert result.did_run(sleep_node3), "Sleep node 3 should have run"
+
+    def test_parallel_sleep_expansion(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        # Create input images with different values
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        image3 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Create a TestParallelSleep node that expands into multiple TestSleep nodes
+        parallel_sleep = g.node("TestParallelSleep",
+                                image1=image1.out(0),
+                                image2=image2.out(0),
+                                image3=image3.out(0),
+                                sleep1=0.4,
+                                sleep2=0.5,
+                                sleep3=0.6)
+        output = g.node("SaveImage", images=parallel_sleep.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Similar to the previous test, expect parallel execution of the sleep nodes
+        # which should complete in less than the sum of all sleeps
+        assert elapsed_time < 0.8, f"Expansion execution took {elapsed_time}s, expected less than 0.8s"
+
+        # Verify the parallel sleep node executed
+        assert result.did_run(parallel_sleep), "ParallelSleep node should have run"
+
+        # Verify we get an image as output (blend of the three input images)
+        result_images = result.get_images(output)
+        assert len(result_images) == 1, "Should have 1 image"
+        # Average pixel value should be around 170 (255 * 2 // 3)
+        avg_value = numpy.array(result_images[0]).mean()
+        assert avg_value == 170, f"Image average value {avg_value} should be 170"
+
     # This tests that nodes with OUTPUT_IS_LIST function correctly when they receive an ExecutionBlocker
     # as input. We also test that when that list (containing an ExecutionBlocker) is passed to a node,
     # only that one entry in the list is blocked.

tests/inference/testing_nodes/testing-pack/__init__.py

@@ -3,6 +3,7 @@ from .flow_control import FLOW_CONTROL_NODE_CLASS_MAPPINGS, FLOW_CONTROL_NODE_DI
 from .util import UTILITY_NODE_CLASS_MAPPINGS, UTILITY_NODE_DISPLAY_NAME_MAPPINGS
 from .conditions import CONDITION_NODE_CLASS_MAPPINGS, CONDITION_NODE_DISPLAY_NAME_MAPPINGS
 from .stubs import TEST_STUB_NODE_CLASS_MAPPINGS, TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS
+from .async_test_nodes import ASYNC_TEST_NODE_CLASS_MAPPINGS, ASYNC_TEST_NODE_DISPLAY_NAME_MAPPINGS
 
 # NODE_CLASS_MAPPINGS = GENERAL_NODE_CLASS_MAPPINGS.update(COMPONENT_NODE_CLASS_MAPPINGS)
 # NODE_DISPLAY_NAME_MAPPINGS = GENERAL_NODE_DISPLAY_NAME_MAPPINGS.update(COMPONENT_NODE_DISPLAY_NAME_MAPPINGS)
@@ -13,6 +14,7 @@ NODE_CLASS_MAPPINGS.update(FLOW_CONTROL_NODE_CLASS_MAPPINGS)
 NODE_CLASS_MAPPINGS.update(UTILITY_NODE_CLASS_MAPPINGS)
 NODE_CLASS_MAPPINGS.update(CONDITION_NODE_CLASS_MAPPINGS)
 NODE_CLASS_MAPPINGS.update(TEST_STUB_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(ASYNC_TEST_NODE_CLASS_MAPPINGS)
 
 NODE_DISPLAY_NAME_MAPPINGS = {}
 NODE_DISPLAY_NAME_MAPPINGS.update(TEST_NODE_DISPLAY_NAME_MAPPINGS)
@@ -20,4 +22,5 @@ NODE_DISPLAY_NAME_MAPPINGS.update(FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS)
 NODE_DISPLAY_NAME_MAPPINGS.update(UTILITY_NODE_DISPLAY_NAME_MAPPINGS)
 NODE_DISPLAY_NAME_MAPPINGS.update(CONDITION_NODE_DISPLAY_NAME_MAPPINGS)
 NODE_DISPLAY_NAME_MAPPINGS.update(TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(ASYNC_TEST_NODE_DISPLAY_NAME_MAPPINGS)
 

tests/inference/testing_nodes/testing-pack/async_test_nodes.py

@@ -0,0 +1,343 @@
+import torch
+import asyncio
+from typing import Dict
+from comfy.utils import ProgressBar
+from comfy_execution.graph_utils import GraphBuilder
+from comfy.comfy_types.node_typing import ComfyNodeABC
+from comfy.comfy_types import IO
+
+
+class TestAsyncValidation(ComfyNodeABC):
+    """Test node with async VALIDATE_INPUTS."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 5.0}),
+                "threshold": ("FLOAT", {"default": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process"
+    CATEGORY = "_for_testing/async"
+
+    @classmethod
+    async def VALIDATE_INPUTS(cls, value, threshold):
+        # Simulate async validation (e.g., checking remote service)
+        await asyncio.sleep(0.05)
+
+        if value > threshold:
+            return f"Value {value} exceeds threshold {threshold}"
+        return True
+
+    def process(self, value, threshold):
+        # Create image based on value
+        intensity = value / 10.0
+        image = torch.ones([1, 512, 512, 3]) * intensity
+        return (image,)
+
+
+class TestAsyncError(ComfyNodeABC):
+    """Test node that errors during async execution."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "error_after": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "error_execution"
+    CATEGORY = "_for_testing/async"
+
+    async def error_execution(self, value, error_after):
+        await asyncio.sleep(error_after)
+        raise RuntimeError("Intentional async execution error for testing")
+
+
+class TestAsyncValidationError(ComfyNodeABC):
+    """Test node with async validation that always fails."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 5.0}),
+                "max_value": ("FLOAT", {"default": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process"
+    CATEGORY = "_for_testing/async"
+
+    @classmethod
+    async def VALIDATE_INPUTS(cls, value, max_value):
+        await asyncio.sleep(0.05)
+        # Always fail validation for values > max_value
+        if value > max_value:
+            return f"Async validation failed: {value} > {max_value}"
+        return True
+
+    def process(self, value, max_value):
+        # This won't be reached if validation fails
+        image = torch.ones([1, 512, 512, 3]) * (value / max_value)
+        return (image,)
+
+
+class TestAsyncTimeout(ComfyNodeABC):
+    """Test node that simulates timeout scenarios."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "timeout": ("FLOAT", {"default": 1.0, "min": 0.1, "max": 10.0}),
+                "operation_time": ("FLOAT", {"default": 2.0, "min": 0.1, "max": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "timeout_execution"
+    CATEGORY = "_for_testing/async"
+
+    async def timeout_execution(self, value, timeout, operation_time):
+        try:
+            # This will timeout if operation_time > timeout
+            await asyncio.wait_for(asyncio.sleep(operation_time), timeout=timeout)
+            return (value,)
+        except asyncio.TimeoutError:
+            raise RuntimeError(f"Operation timed out after {timeout} seconds")
+
+
+class TestSyncError(ComfyNodeABC):
+    """Test node that errors synchronously (for mixed sync/async testing)."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "sync_error"
+    CATEGORY = "_for_testing/async"
+
+    def sync_error(self, value):
+        raise RuntimeError("Intentional sync execution error for testing")
+
+
+class TestAsyncLazyCheck(ComfyNodeABC):
+    """Test node with async check_lazy_status."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": (IO.ANY, {"lazy": True}),
+                "input2": (IO.ANY, {"lazy": True}),
+                "condition": ("BOOLEAN", {"default": True}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process"
+    CATEGORY = "_for_testing/async"
+
+    async def check_lazy_status(self, condition, input1, input2):
+        # Simulate async checking (e.g., querying remote service)
+        await asyncio.sleep(0.05)
+
+        needed = []
+        if condition and input1 is None:
+            needed.append("input1")
+        if not condition and input2 is None:
+            needed.append("input2")
+        return needed
+
+    def process(self, input1, input2, condition):
+        # Return a simple image
+        return (torch.ones([1, 512, 512, 3]),)
+
+
+class TestDynamicAsyncGeneration(ComfyNodeABC):
+    """Test node that dynamically generates async nodes."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image1": ("IMAGE",),
+                "image2": ("IMAGE",),
+                "num_async_nodes": ("INT", {"default": 3, "min": 1, "max": 10}),
+                "sleep_duration": ("FLOAT", {"default": 0.2, "min": 0.1, "max": 1.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "generate_async_workflow"
+    CATEGORY = "_for_testing/async"
+
+    def generate_async_workflow(self, image1, image2, num_async_nodes, sleep_duration):
+        g = GraphBuilder()
+
+        # Create multiple async sleep nodes
+        sleep_nodes = []
+        for i in range(num_async_nodes):
+            image = image1 if i % 2 == 0 else image2
+            sleep_node = g.node("TestSleep", value=image, seconds=sleep_duration)
+            sleep_nodes.append(sleep_node)
+
+        # Average all results
+        if len(sleep_nodes) == 1:
+            final_node = sleep_nodes[0]
+        else:
+            avg_inputs = {"input1": sleep_nodes[0].out(0)}
+            for i, node in enumerate(sleep_nodes[1:], 2):
+                avg_inputs[f"input{i}"] = node.out(0)
+            final_node = g.node("TestVariadicAverage", **avg_inputs)
+
+        return {
+            "result": (final_node.out(0),),
+            "expand": g.finalize(),
+        }
+
+
+class TestAsyncResourceUser(ComfyNodeABC):
+    """Test node that uses resources during async execution."""
+
+    # Class-level resource tracking for testing
+    _active_resources: Dict[str, bool] = {}
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "resource_id": ("STRING", {"default": "resource_0"}),
+                "duration": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 1.0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "use_resource"
+    CATEGORY = "_for_testing/async"
+
+    async def use_resource(self, value, resource_id, duration):
+        # Check if resource is already in use
+        if self._active_resources.get(resource_id, False):
+            raise RuntimeError(f"Resource {resource_id} is already in use!")
+
+        # Mark resource as in use
+        self._active_resources[resource_id] = True
+
+        try:
+            # Simulate resource usage
+            await asyncio.sleep(duration)
+            return (value,)
+        finally:
+            # Always clean up resource
+            self._active_resources[resource_id] = False
+
+
+class TestAsyncBatchProcessing(ComfyNodeABC):
+    """Test async processing of batched inputs."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "process_time_per_item": ("FLOAT", {"default": 0.1, "min": 0.01, "max": 1.0}),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process_batch"
+    CATEGORY = "_for_testing/async"
+
+    async def process_batch(self, images, process_time_per_item, unique_id):
+        batch_size = images.shape[0]
+        pbar = ProgressBar(batch_size, node_id=unique_id)
+
+        # Process each image in the batch
+        processed = []
+        for i in range(batch_size):
+            # Simulate async processing
+            await asyncio.sleep(process_time_per_item)
+
+            # Simple processing: invert the image
+            processed_image = 1.0 - images[i:i+1]
+            processed.append(processed_image)
+
+            pbar.update(1)
+
+        # Stack processed images
+        result = torch.cat(processed, dim=0)
+        return (result,)
+
+
+class TestAsyncConcurrentLimit(ComfyNodeABC):
+    """Test concurrent execution limits for async nodes."""
+
+    _semaphore = asyncio.Semaphore(2)  # Only allow 2 concurrent executions
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "duration": ("FLOAT", {"default": 0.5, "min": 0.1, "max": 2.0}),
+                "node_id": ("INT", {"default": 0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "limited_execution"
+    CATEGORY = "_for_testing/async"
+
+    async def limited_execution(self, value, duration, node_id):
+        async with self._semaphore:
+            # Node {node_id} acquired semaphore
+            await asyncio.sleep(duration)
+            # Node {node_id} releasing semaphore
+            return (value,)
+
+
+# Add node mappings
+ASYNC_TEST_NODE_CLASS_MAPPINGS = {
+    "TestAsyncValidation": TestAsyncValidation,
+    "TestAsyncError": TestAsyncError,
+    "TestAsyncValidationError": TestAsyncValidationError,
+    "TestAsyncTimeout": TestAsyncTimeout,
+    "TestSyncError": TestSyncError,
+    "TestAsyncLazyCheck": TestAsyncLazyCheck,
+    "TestDynamicAsyncGeneration": TestDynamicAsyncGeneration,
+    "TestAsyncResourceUser": TestAsyncResourceUser,
+    "TestAsyncBatchProcessing": TestAsyncBatchProcessing,
+    "TestAsyncConcurrentLimit": TestAsyncConcurrentLimit,
+}
+
+ASYNC_TEST_NODE_DISPLAY_NAME_MAPPINGS = {
+    "TestAsyncValidation": "Test Async Validation",
+    "TestAsyncError": "Test Async Error",
+    "TestAsyncValidationError": "Test Async Validation Error",
+    "TestAsyncTimeout": "Test Async Timeout",
+    "TestSyncError": "Test Sync Error",
+    "TestAsyncLazyCheck": "Test Async Lazy Check",
+    "TestDynamicAsyncGeneration": "Test Dynamic Async Generation",
+    "TestAsyncResourceUser": "Test Async Resource User",
+    "TestAsyncBatchProcessing": "Test Async Batch Processing",
+    "TestAsyncConcurrentLimit": "Test Async Concurrent Limit",
+}

tests/inference/testing_nodes/testing-pack/specific_tests.py

@@ -1,6 +1,11 @@
 import torch
+import time
+import asyncio
+from comfy.utils import ProgressBar
 from .tools import VariantSupport
 from comfy_execution.graph_utils import GraphBuilder
+from comfy.comfy_types.node_typing import ComfyNodeABC
+from comfy.comfy_types import IO
 
 class TestLazyMixImages:
     @classmethod
@@ -333,6 +338,131 @@ class TestMixedExpansionReturns:
                 "expand": g.finalize(),
             }
 
+class TestSamplingInExpansion:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model": ("MODEL",),
+                "clip": ("CLIP",),
+                "vae": ("VAE",),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
+                "steps": ("INT", {"default": 20, "min": 1, "max": 100}),
+                "cfg": ("FLOAT", {"default": 7.0, "min": 0.0, "max": 30.0}),
+                "prompt": ("STRING", {"multiline": True, "default": "a beautiful landscape with mountains and trees"}),
+                "negative_prompt": ("STRING", {"multiline": True, "default": "blurry, bad quality, worst quality"}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "sampling_in_expansion"
+
+    CATEGORY = "Testing/Nodes"
+
+    def sampling_in_expansion(self, model, clip, vae, seed, steps, cfg, prompt, negative_prompt):
+        g = GraphBuilder()
+
+        # Create a basic image generation workflow using the input model, clip and vae
+        # 1. Setup text prompts using the provided CLIP model
+        positive_prompt = g.node("CLIPTextEncode",
+                               text=prompt,
+                               clip=clip)
+        negative_prompt = g.node("CLIPTextEncode",
+                                text=negative_prompt,
+                                clip=clip)
+
+        # 2. Create empty latent with specified size
+        empty_latent = g.node("EmptyLatentImage", width=512, height=512, batch_size=1)
+
+        # 3. Setup sampler and generate image latent
+        sampler = g.node("KSampler",
+                        model=model,
+                        positive=positive_prompt.out(0),
+                        negative=negative_prompt.out(0),
+                        latent_image=empty_latent.out(0),
+                        seed=seed,
+                        steps=steps,
+                        cfg=cfg,
+                        sampler_name="euler_ancestral",
+                        scheduler="normal")
+
+        # 4. Decode latent to image using VAE
+        output = g.node("VAEDecode", samples=sampler.out(0), vae=vae)
+
+        return {
+            "result": (output.out(0),),
+            "expand": g.finalize(),
+        }
+
+class TestSleep(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "seconds": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 9999.0, "step": 0.01, "tooltip": "The amount of seconds to sleep."}),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "sleep"
+
+    CATEGORY = "_for_testing"
+
+    async def sleep(self, value, seconds, unique_id):
+        pbar = ProgressBar(seconds, node_id=unique_id)
+        start = time.time()
+        expiration = start + seconds
+        now = start
+        while now < expiration:
+            now = time.time()
+            pbar.update_absolute(now - start)
+            await asyncio.sleep(0.01)
+        return (value,)
+
+class TestParallelSleep(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image1": ("IMAGE", ),
+                "image2": ("IMAGE", ),
+                "image3": ("IMAGE", ),
+                "sleep1": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step": 0.01}),
+                "sleep2": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step": 0.01}),
+                "sleep3": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step": 0.01}),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "parallel_sleep"
+    CATEGORY = "_for_testing"
+    OUTPUT_NODE = True
+
+    def parallel_sleep(self, image1, image2, image3, sleep1, sleep2, sleep3, unique_id):
+        # Create a graph dynamically with three TestSleep nodes
+        g = GraphBuilder()
+
+        # Create sleep nodes for each duration and image
+        sleep_node1 = g.node("TestSleep", value=image1, seconds=sleep1)
+        sleep_node2 = g.node("TestSleep", value=image2, seconds=sleep2)
+        sleep_node3 = g.node("TestSleep", value=image3, seconds=sleep3)
+
+        # Blend the results using TestVariadicAverage
+        blend = g.node("TestVariadicAverage",
+                       input1=sleep_node1.out(0),
+                       input2=sleep_node2.out(0),
+                       input3=sleep_node3.out(0))
+
+        return {
+            "result": (blend.out(0),),
+            "expand": g.finalize(),
+        }
+
 TEST_NODE_CLASS_MAPPINGS = {
     "TestLazyMixImages": TestLazyMixImages,
     "TestVariadicAverage": TestVariadicAverage,
@@ -345,6 +475,9 @@ TEST_NODE_CLASS_MAPPINGS = {
     "TestCustomValidation5": TestCustomValidation5,
     "TestDynamicDependencyCycle": TestDynamicDependencyCycle,
     "TestMixedExpansionReturns": TestMixedExpansionReturns,
+    "TestSamplingInExpansion": TestSamplingInExpansion,
+    "TestSleep": TestSleep,
+    "TestParallelSleep": TestParallelSleep,
 }
 
 TEST_NODE_DISPLAY_NAME_MAPPINGS = {
@@ -359,4 +492,7 @@ TEST_NODE_DISPLAY_NAME_MAPPINGS = {
     "TestCustomValidation5": "Custom Validation 5",
     "TestDynamicDependencyCycle": "Dynamic Dependency Cycle",
     "TestMixedExpansionReturns": "Mixed Expansion Returns",
+    "TestSamplingInExpansion": "Sampling In Expansion",
+    "TestSleep": "Test Sleep",
+    "TestParallelSleep": "Test Parallel Sleep",
 }