ComfyUI version 0.3.44

* update nightly workflow with cu129

* Remove unused file to lower standalone size.

.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/windows_release_nightly_pytorch.yml

@@ -7,7 +7,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "128"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
@@ -19,7 +19,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "2"
+        default: "5"
 #  push:
 #    branches:
 #      - master
@@ -53,6 +53,8 @@ jobs:
             ls ../temp_wheel_dir
             ./python.exe -s -m pip install --pre ../temp_wheel_dir/*
             sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
+
+            rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd

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/)
@@ -82,6 +86,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - Smart memory management: can automatically run models on GPUs with as low as 1GB vram.
 - Works even if you don't have a GPU with: ```--cpu``` (slow)
 - Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs and CLIP models.
+- Safe loading of ckpt, pt, pth, etc.. files.
 - Embeddings/Textual inversion
 - [Loras (regular, locon and loha)](https://comfyanonymous.github.io/ComfyUI_examples/lora/)
 - [Hypernetworks](https://comfyanonymous.github.io/ComfyUI_examples/hypernetworks/)
@@ -97,7 +102,6 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - [Model Merging](https://comfyanonymous.github.io/ComfyUI_examples/model_merging/)
 - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/)
 - Latent previews with [TAESD](#how-to-show-high-quality-previews)
-- Starts up very fast.
 - Works fully offline: core will never download anything unless you want to.
 - Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview).
 - [Config file](extra_model_paths.yaml.example) to set the search paths for models.
@@ -239,7 +243,7 @@ Nvidia users should install stable pytorch using this command:
 
 This is the command to install pytorch nightly instead which might have performance improvements.
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu129```
 
 #### Troubleshooting
 
@@ -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/gligen.py

@@ -1,55 +1,10 @@
 import math
 import torch
 from torch import nn
-from .ldm.modules.attention import CrossAttention
-from inspect import isfunction
+from .ldm.modules.attention import CrossAttention, FeedForward
 import comfy.ops
 ops = comfy.ops.manual_cast
 
-def exists(val):
-    return val is not None
-
-
-def uniq(arr):
-    return{el: True for el in arr}.keys()
-
-
-def default(val, d):
-    if exists(val):
-        return val
-    return d() if isfunction(d) else d
-
-
-# feedforward
-class GEGLU(nn.Module):
-    def __init__(self, dim_in, dim_out):
-        super().__init__()
-        self.proj = ops.Linear(dim_in, dim_out * 2)
-
-    def forward(self, x):
-        x, gate = self.proj(x).chunk(2, dim=-1)
-        return x * torch.nn.functional.gelu(gate)
-
-
-class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
-        super().__init__()
-        inner_dim = int(dim * mult)
-        dim_out = default(dim_out, dim)
-        project_in = nn.Sequential(
-            ops.Linear(dim, inner_dim),
-            nn.GELU()
-        ) if not glu else GEGLU(dim, inner_dim)
-
-        self.net = nn.Sequential(
-            project_in,
-            nn.Dropout(dropout),
-            ops.Linear(inner_dim, dim_out)
-        )
-
-    def forward(self, x):
-        return self.net(x)
-
 
 class GatedCrossAttentionDense(nn.Module):
     def __init__(self, query_dim, context_dim, n_heads, d_head):

comfy/k_diffusion/sampling.py

@@ -1,4 +1,5 @@
 import math
+from functools import partial
 
 from scipy import integrate
 import torch
@@ -142,6 +143,33 @@ class BrownianTreeNoiseSampler:
         return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
 
 
+def sigma_to_half_log_snr(sigma, model_sampling):
+    """Convert sigma to half-logSNR log(alpha_t / sigma_t)."""
+    if isinstance(model_sampling, comfy.model_sampling.CONST):
+        # log((1 - t) / t) = log((1 - sigma) / sigma)
+        return sigma.logit().neg()
+    return sigma.log().neg()
+
+
+def half_log_snr_to_sigma(half_log_snr, model_sampling):
+    """Convert half-logSNR log(alpha_t / sigma_t) to sigma."""
+    if isinstance(model_sampling, comfy.model_sampling.CONST):
+        # 1 / (1 + exp(half_log_snr))
+        return half_log_snr.neg().sigmoid()
+    return half_log_snr.neg().exp()
+
+
+def offset_first_sigma_for_snr(sigmas, model_sampling, percent_offset=1e-4):
+    """Adjust the first sigma to avoid invalid logSNR."""
+    if len(sigmas) <= 1:
+        return sigmas
+    if isinstance(model_sampling, comfy.model_sampling.CONST):
+        if sigmas[0] >= 1:
+            sigmas = sigmas.clone()
+            sigmas[0] = model_sampling.percent_to_sigma(percent_offset)
+    return sigmas
+
+
 @torch.no_grad()
 def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
     """Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
@@ -384,9 +412,13 @@ def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, o
             ds.pop(0)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
-        cur_order = min(i + 1, order)
-        coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
-        x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            cur_order = min(i + 1, order)
+            coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
+            x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
     return x
 
 
@@ -682,6 +714,7 @@ 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)."""
@@ -693,38 +726,49 @@ 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
 
 
@@ -753,6 +797,7 @@ def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=No
         old_denoised = denoised
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     """DPM-Solver++(2M) SDE."""
@@ -768,9 +813,12 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     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]])
 
+    model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling')
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+
     old_denoised = None
-    h_last = None
-    h = None
+    h, h_last = None, None
 
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
@@ -781,26 +829,29 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
             x = denoised
         else:
             # DPM-Solver++(2M) SDE
-            t, s = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = s - t
-            eta_h = eta * h
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
+            h_eta = h * (eta + 1)
 
-            x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_t * (-h_eta).expm1().neg() * denoised
 
             if old_denoised is not None:
                 r = h_last / h
                 if solver_type == 'heun':
-                    x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised)
+                    x = x + alpha_t * ((-h_eta).expm1().neg() / (-h_eta) + 1) * (1 / r) * (denoised - old_denoised)
                 elif solver_type == 'midpoint':
-                    x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised)
+                    x = x + 0.5 * alpha_t * (-h_eta).expm1().neg() * (1 / r) * (denoised - old_denoised)
 
-            if eta:
-                x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise
+            if eta > 0 and s_noise > 0:
+                x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise
 
         old_denoised = denoised
         h_last = h
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     """DPM-Solver++(3M) SDE."""
@@ -814,6 +865,10 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     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]])
 
+    model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling')
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+
     denoised_1, denoised_2 = None, None
     h, h_1, h_2 = None, None, None
 
@@ -825,13 +880,16 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
             # Denoising step
             x = denoised
         else:
-            t, s = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = s - t
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
             h_eta = h * (eta + 1)
 
-            x = torch.exp(-h_eta) * x + (-h_eta).expm1().neg() * denoised
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_t * (-h_eta).expm1().neg() * denoised
 
             if h_2 is not None:
+                # DPM-Solver++(3M) SDE
                 r0 = h_1 / h
                 r1 = h_2 / h
                 d1_0 = (denoised - denoised_1) / r0
@@ -840,20 +898,22 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
                 d2 = (d1_0 - d1_1) / (r0 + r1)
                 phi_2 = h_eta.neg().expm1() / h_eta + 1
                 phi_3 = phi_2 / h_eta - 0.5
-                x = x + phi_2 * d1 - phi_3 * d2
+                x = x + (alpha_t * phi_2) * d1 - (alpha_t * phi_3) * d2
             elif h_1 is not None:
+                # DPM-Solver++(2M) SDE
                 r = h_1 / h
                 d = (denoised - denoised_1) / r
                 phi_2 = h_eta.neg().expm1() / h_eta + 1
-                x = x + phi_2 * d
+                x = x + (alpha_t * phi_2) * d
 
-            if eta:
+            if eta > 0 and s_noise > 0:
                 x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noise
 
         denoised_1, denoised_2 = denoised, denoised_1
         h_1, h_2 = h, h_1
     return x
 
+
 @torch.no_grad()
 def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     if len(sigmas) <= 1:
@@ -863,6 +923,7 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
 
+
 @torch.no_grad()
 def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     if len(sigmas) <= 1:
@@ -872,6 +933,7 @@ def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
 
+
 @torch.no_grad()
 def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     if len(sigmas) <= 1:
@@ -1009,7 +1071,9 @@ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None,
         d_cur = (x_cur - denoised) / t_cur
 
         order = min(max_order, i+1)
-        if order == 1:      # First Euler step.
+        if t_next == 0:     # Denoising step
+            x_next = denoised
+        elif order == 1:    # First Euler step.
             x_next = x_cur + (t_next - t_cur) * d_cur
         elif order == 2:    # Use one history point.
             x_next = x_cur + (t_next - t_cur) * (3 * d_cur - buffer_model[-1]) / 2
@@ -1027,6 +1091,7 @@ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None,
 
     return x_next
 
+
 #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
 #under Apache 2 license
 def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
@@ -1050,7 +1115,9 @@ def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=Non
         d_cur = (x_cur - denoised) / t_cur
 
         order = min(max_order, i+1)
-        if order == 1:      # First Euler step.
+        if t_next == 0:     # Denoising step
+            x_next = denoised
+        elif order == 1:    # First Euler step.
             x_next = x_cur + (t_next - t_cur) * d_cur
         elif order == 2:    # Use one history point.
             h_n = (t_next - t_cur)
@@ -1090,6 +1157,7 @@ def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=Non
 
     return x_next
 
+
 #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
 #under Apache 2 license
 @torch.no_grad()
@@ -1140,6 +1208,7 @@ def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None,
 
     return x_next
 
+
 @torch.no_grad()
 def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
     extra_args = {} if extra_args is None else extra_args
@@ -1346,6 +1415,7 @@ def sample_res_multistep_ancestral(model, x, sigmas, extra_args=None, callback=N
 def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True)
 
+
 @torch.no_grad()
 def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False):
     """Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK"""
@@ -1372,31 +1442,32 @@ def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None,
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
         dt = sigmas[i + 1] - sigmas[i]
-        if i == 0:
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
             # Euler method
             if cfg_pp:
                 x = denoised + d * sigmas[i + 1]
             else:
                 x = x + d * dt
-        else:
-            # Gradient estimation
-            if cfg_pp:
+
+            if i >= 1:
+                # Gradient estimation
                 d_bar = (ge_gamma - 1) * (d - old_d)
-                x = denoised + d * sigmas[i + 1] + d_bar * dt
-            else:
-                d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
                 x = x + d_bar * dt
         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
@@ -1404,12 +1475,18 @@ 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
 
@@ -1420,41 +1497,45 @@ 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
 
+
 @torch.no_grad()
 def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5):
-    '''
-    SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VE Data Prediction) stage 2
-    Arxiv: https://arxiv.org/abs/2305.14267
-    '''
+    """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2.
+    arXiv: https://arxiv.org/abs/2305.14267
+    """
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
@@ -1462,6 +1543,11 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
 
     inject_noise = eta > 0 and s_noise > 0
 
+    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:
@@ -1469,80 +1555,96 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
         if sigmas[i + 1] == 0:
             x = denoised
         else:
-            t, t_next = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = t_next - t
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
             h_eta = h * (eta + 1)
-            s = t + r * h
+            lambda_s_1 = lambda_s + r * h
             fac = 1 / (2 * r)
-            sigma_s = s.neg().exp()
+            sigma_s_1 = sigma_fn(lambda_s_1)
+
+            # alpha_t = sigma_t * exp(log(alpha_t / sigma_t)) = sigma_t * exp(lambda_t)
+            alpha_s_1 = sigma_s_1 * lambda_s_1.exp()
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
 
             coeff_1, coeff_2 = (-r * h_eta).expm1(), (-h_eta).expm1()
             if inject_noise:
+                # 0 < r < 1
                 noise_coeff_1 = (-2 * r * h * eta).expm1().neg().sqrt()
-                noise_coeff_2 = ((-2 * r * h * eta).expm1() - (-2 * h * eta).expm1()).sqrt()
-                noise_1, noise_2 = noise_sampler(sigmas[i], sigma_s), noise_sampler(sigma_s, sigmas[i + 1])
+                noise_coeff_2 = (-r * h * eta).exp() * (-2 * (1 - r) * h * eta).expm1().neg().sqrt()
+                noise_1, noise_2 = noise_sampler(sigmas[i], sigma_s_1), noise_sampler(sigma_s_1, sigmas[i + 1])
 
             # Step 1
-            x_2 = (coeff_1 + 1) * x - coeff_1 * denoised
-            if inject_noise:
-                x_2 = x_2 + sigma_s * (noise_coeff_1 * noise_1) * s_noise
-            denoised_2 = model(x_2, sigma_s * s_in, **extra_args)
-
-            # Step 2
-            denoised_d = (1 - fac) * denoised + fac * denoised_2
-            x = (coeff_2 + 1) * x - coeff_2 * denoised_d
-            if inject_noise:
-                x = x + sigmas[i + 1] * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
-    return x
-
-@torch.no_grad()
-def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
-    '''
-    SEEDS-3 - Stochastic Explicit Exponential Derivative-free Solvers (VE Data Prediction) stage 3
-    Arxiv: https://arxiv.org/abs/2305.14267
-    '''
-    extra_args = {} if extra_args is None else extra_args
-    seed = extra_args.get("seed", 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]])
-
-    inject_noise = eta > 0 and s_noise > 0
-
-    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:
-            x = denoised
-        else:
-            t, t_next = -sigmas[i].log(), -sigmas[i + 1].log()
-            h = t_next - t
-            h_eta = h * (eta + 1)
-            s_1 = t + r_1 * h
-            s_2 = t + r_2 * h
-            sigma_s_1, sigma_s_2 = s_1.neg().exp(), s_2.neg().exp()
-
-            coeff_1, coeff_2, coeff_3 = (-r_1 * h_eta).expm1(), (-r_2 * h_eta).expm1(), (-h_eta).expm1()
-            if inject_noise:
-                noise_coeff_1 = (-2 * r_1 * h * eta).expm1().neg().sqrt()
-                noise_coeff_2 = ((-2 * r_1 * h * eta).expm1() - (-2 * r_2 * h * eta).expm1()).sqrt()
-                noise_coeff_3 = ((-2 * r_2 * h * eta).expm1() - (-2 * h * eta).expm1()).sqrt()
-                noise_1, noise_2, noise_3 = noise_sampler(sigmas[i], sigma_s_1), noise_sampler(sigma_s_1, sigma_s_2), noise_sampler(sigma_s_2, sigmas[i + 1])
-
-            # Step 1
-            x_2 = (coeff_1 + 1) * x - coeff_1 * denoised
+            x_2 = sigma_s_1 / sigmas[i] * (-r * h * eta).exp() * x - alpha_s_1 * coeff_1 * denoised
             if inject_noise:
                 x_2 = x_2 + sigma_s_1 * (noise_coeff_1 * noise_1) * s_noise
             denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
             # Step 2
-            x_3 = (coeff_2 + 1) * x - coeff_2 * denoised + (r_2 / r_1) * (coeff_2 / (r_2 * h_eta) + 1) * (denoised_2 - denoised)
+            denoised_d = (1 - fac) * denoised + fac * denoised_2
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * coeff_2 * denoised_d
+            if inject_noise:
+                x = x + sigmas[i + 1] * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
+    return x
+
+
+@torch.no_grad()
+def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
+    """SEEDS-3 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 3.
+    arXiv: https://arxiv.org/abs/2305.14267
+    """
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", 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]])
+
+    inject_noise = eta > 0 and s_noise > 0
+
+    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:
+            x = denoised
+        else:
+            lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])
+            h = lambda_t - lambda_s
+            h_eta = h * (eta + 1)
+            lambda_s_1 = lambda_s + r_1 * h
+            lambda_s_2 = lambda_s + r_2 * h
+            sigma_s_1, sigma_s_2 = sigma_fn(lambda_s_1), sigma_fn(lambda_s_2)
+
+            # alpha_t = sigma_t * exp(log(alpha_t / sigma_t)) = sigma_t * exp(lambda_t)
+            alpha_s_1 = sigma_s_1 * lambda_s_1.exp()
+            alpha_s_2 = sigma_s_2 * lambda_s_2.exp()
+            alpha_t = sigmas[i + 1] * lambda_t.exp()
+
+            coeff_1, coeff_2, coeff_3 = (-r_1 * h_eta).expm1(), (-r_2 * h_eta).expm1(), (-h_eta).expm1()
+            if inject_noise:
+                # 0 < r_1 < r_2 < 1
+                noise_coeff_1 = (-2 * r_1 * h * eta).expm1().neg().sqrt()
+                noise_coeff_2 = (-r_1 * h * eta).exp() * (-2 * (r_2 - r_1) * h * eta).expm1().neg().sqrt()
+                noise_coeff_3 = (-r_2 * h * eta).exp() * (-2 * (1 - r_2) * h * eta).expm1().neg().sqrt()
+                noise_1, noise_2, noise_3 = noise_sampler(sigmas[i], sigma_s_1), noise_sampler(sigma_s_1, sigma_s_2), noise_sampler(sigma_s_2, sigmas[i + 1])
+
+            # Step 1
+            x_2 = sigma_s_1 / sigmas[i] * (-r_1 * h * eta).exp() * x - alpha_s_1 * coeff_1 * denoised
+            if inject_noise:
+                x_2 = x_2 + sigma_s_1 * (noise_coeff_1 * noise_1) * s_noise
+            denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
+
+            # Step 2
+            x_3 = sigma_s_2 / sigmas[i] * (-r_2 * h * eta).exp() * x - alpha_s_2 * coeff_2 * denoised + (r_2 / r_1) * alpha_s_2 * (coeff_2 / (r_2 * h_eta) + 1) * (denoised_2 - denoised)
             if inject_noise:
                 x_3 = x_3 + sigma_s_2 * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
             denoised_3 = model(x_3, sigma_s_2 * s_in, **extra_args)
 
             # Step 3
-            x = (coeff_3 + 1) * x - coeff_3 * denoised + (1. / r_2) * (coeff_3 / h_eta + 1) * (denoised_3 - denoised)
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * coeff_3 * denoised + (1. / r_2) * alpha_t * (coeff_3 / h_eta + 1) * (denoised_3 - denoised)
             if inject_noise:
                 x = x + sigmas[i + 1] * (noise_coeff_3 * noise_1 + noise_coeff_2 * noise_2 + noise_coeff_1 * noise_3) * s_noise
     return x

comfy/ldm/cosmos/predict2.py

@@ -70,11 +70,7 @@ def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H
     q_B_H_S_D = rearrange(q_B_S_H_D, "b ... h k -> b h ... k").view(in_q_shape[0], in_q_shape[-2], -1, in_q_shape[-1])
     k_B_H_S_D = rearrange(k_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
     v_B_H_S_D = rearrange(v_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
-    result_B_S_HD = rearrange(
-        optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True, skip_output_reshape=True), "b h ... l -> b ... (h l)"
-    )
-
-    return result_B_S_HD
+    return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True)
 
 
 class Attention(nn.Module):

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/lightricks/model.py

@@ -261,8 +261,8 @@ class CrossAttention(nn.Module):
         self.heads = heads
         self.dim_head = dim_head
 
-        self.q_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
-        self.k_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.q_norm = operations.RMSNorm(inner_dim, eps=1e-5, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(inner_dim, eps=1e-5, dtype=dtype, device=device)
 
         self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)

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)
@@ -1014,9 +1031,32 @@ class CosmosPredict2(BaseModel):
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
 
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+
         out['fps'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", None))
         return out
 
+    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])
+        return out
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        sigma = sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1))
+        sigma_noise_augmentation = 0 #TODO
+        if sigma_noise_augmentation != 0:
+            latent_image = latent_image + noise
+        latent_image = self.model_sampling.calculate_input(torch.tensor([sigma_noise_augmentation], device=latent_image.device, dtype=latent_image.dtype), latent_image)
+        sigma = (sigma / (sigma + 1))
+        return latent_image / (1.0 - sigma)
+
 class Lumina2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT)
@@ -1207,3 +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

@@ -441,11 +441,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["rope_h_extrapolation_ratio"] = 4.0
             dit_config["rope_w_extrapolation_ratio"] = 4.0
             dit_config["rope_t_extrapolation_ratio"] = 1.0
-        elif dit_config["in_channels"] == 17:
-            dit_config["extra_per_block_abs_pos_emb"] = False
-            dit_config["rope_h_extrapolation_ratio"] = 3.0
-            dit_config["rope_w_extrapolation_ratio"] = 3.0
-            dit_config["rope_t_extrapolation_ratio"] = 1.0
+        elif dit_config["in_channels"] == 17: # img to video
+            if dit_config["model_channels"] == 2048:
+                dit_config["extra_per_block_abs_pos_emb"] = False
+                dit_config["rope_h_extrapolation_ratio"] = 3.0
+                dit_config["rope_w_extrapolation_ratio"] = 3.0
+                dit_config["rope_t_extrapolation_ratio"] = 1.0
+            elif dit_config["model_channels"] == 5120:
+                dit_config["rope_h_extrapolation_ratio"] = 2.0
+                dit_config["rope_w_extrapolation_ratio"] = 2.0
+                dit_config["rope_t_extrapolation_ratio"] = 0.8333333333333334
 
         dit_config["extra_h_extrapolation_ratio"] = 1.0
         dit_config["extra_w_extrapolation_ratio"] = 1.0
@@ -454,6 +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/model_patcher.py

@@ -379,6 +379,9 @@ class ModelPatcher:
     def set_model_sampler_pre_cfg_function(self, pre_cfg_function, disable_cfg1_optimization=False):
         self.model_options = set_model_options_pre_cfg_function(self.model_options, pre_cfg_function, disable_cfg1_optimization)
 
+    def set_model_sampler_calc_cond_batch_function(self, sampler_calc_cond_batch_function):
+        self.model_options["sampler_calc_cond_batch_function"] = sampler_calc_cond_batch_function
+
     def set_model_unet_function_wrapper(self, unet_wrapper_function: UnetWrapperFunction):
         self.model_options["model_function_wrapper"] = unet_wrapper_function
 

comfy/ops.py

@@ -336,9 +336,12 @@ class fp8_ops(manual_cast):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            out = fp8_linear(self, input)
-            if out is not None:
-                return out
+            try:
+                out = fp8_linear(self, input)
+                if out is not None:
+                    return out
+            except Exception as e:
+                logging.info("Exception during fp8 op: {}".format(e))
 
             weight, bias = cast_bias_weight(self, input)
             return torch.nn.functional.linear(input, weight, bias)

comfy/samplers.py

@@ -373,7 +373,11 @@ def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_option
         uncond_ = uncond
 
     conds = [cond, uncond_]
-    out = calc_cond_batch(model, conds, x, timestep, model_options)
+    if "sampler_calc_cond_batch_function" in model_options:
+        args = {"conds": conds, "input": x, "sigma": timestep, "model": model, "model_options": model_options}
+        out = model_options["sampler_calc_cond_batch_function"](args)
+    else:
+        out = calc_cond_batch(model, conds, x, timestep, model_options)
 
     for fn in model_options.get("sampler_pre_cfg_function", []):
         args = {"conds":conds, "conds_out": out, "cond_scale": cond_scale, "timestep": timestep,
@@ -1039,13 +1043,13 @@ 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

@@ -462,7 +462,7 @@ class SDTokenizer:
             tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
         self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args)
         self.max_length = tokenizer_data.get("{}_max_length".format(embedding_key), max_length)
-        self.min_length = min_length
+        self.min_length = tokenizer_data.get("{}_min_length".format(embedding_key), min_length)
         self.end_token = None
         self.min_padding = min_padding
 
@@ -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

@@ -77,6 +77,7 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
         if safe_load or ALWAYS_SAFE_LOAD:
             pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args)
         else:
+            logging.warning("WARNING: loading {} unsafely, upgrade your pytorch to 2.4 or newer to load this file safely.".format(ckpt))
             pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)
         if "state_dict" in pl_sd:
             sd = pl_sd["state_dict"]

comfy_api/input/video_types.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 from abc import ABC, abstractmethod
-from typing import Optional
+from typing import Optional, Union
+import io
 from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
 
 class VideoInput(ABC):
@@ -31,6 +32,22 @@ class VideoInput(ABC):
         """
         pass
 
+    def get_stream_source(self) -> Union[str, io.BytesIO]:
+        """
+        Get a streamable source for the video. This allows processing without
+        loading the entire video into memory.
+
+        Returns:
+            Either a file path (str) or a BytesIO object that can be opened with av.
+
+        Default implementation creates a BytesIO buffer, but subclasses should
+        override this for better performance when possible.
+        """
+        buffer = io.BytesIO()
+        self.save_to(buffer)
+        buffer.seek(0)
+        return buffer
+
     # Provide a default implementation, but subclasses can provide optimized versions
     # if possible.
     def get_dimensions(self) -> tuple[int, int]:

comfy_api/input_impl/video_types.py

@@ -64,6 +64,15 @@ class VideoFromFile(VideoInput):
         """
         self.__file = file
 
+    def get_stream_source(self) -> str | io.BytesIO:
+        """
+        Return the underlying file source for efficient streaming.
+        This avoids unnecessary memory copies when the source is already a file path.
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+        return self.__file
+
     def get_dimensions(self) -> tuple[int, int]:
         """
         Returns the dimensions of the video input.

comfy_api_nodes/apis/__init__.py

@@ -1,6 +1,6 @@
 # generated by datamodel-codegen:
 #   filename:  filtered-openapi.yaml
-#   timestamp: 2025-05-19T21:38:55+00:00
+#   timestamp: 2025-07-06T09:47:31+00:00
 
 from __future__ import annotations
 
@@ -1355,6 +1355,158 @@ class ModelResponseProperties(BaseModel):
     )
 
 
+class Keyframes(BaseModel):
+    image_url: Optional[str] = None
+
+
+class MoonvalleyPromptResponse(BaseModel):
+    error: Optional[Dict[str, Any]] = None
+    frame_conditioning: Optional[Dict[str, Any]] = None
+    id: Optional[str] = None
+    inference_params: Optional[Dict[str, Any]] = None
+    meta: Optional[Dict[str, Any]] = None
+    model_params: Optional[Dict[str, Any]] = None
+    output_url: Optional[str] = None
+    prompt_text: Optional[str] = None
+    status: Optional[str] = None
+
+
+class MoonvalleyTextToVideoInferenceParams(BaseModel):
+    add_quality_guidance: Optional[bool] = Field(
+        True, description='Whether to add quality guidance'
+    )
+    caching_coefficient: Optional[float] = Field(
+        0.3, description='Caching coefficient for optimization'
+    )
+    caching_cooldown: Optional[int] = Field(
+        3, description='Number of caching cooldown steps'
+    )
+    caching_warmup: Optional[int] = Field(
+        3, description='Number of caching warmup steps'
+    )
+    clip_value: Optional[float] = Field(
+        3, description='CLIP value for generation control'
+    )
+    conditioning_frame_index: Optional[int] = Field(
+        0, description='Index of the conditioning frame'
+    )
+    cooldown_steps: Optional[int] = Field(
+        None, description='Number of cooldown steps (calculated based on num_frames)'
+    )
+    fps: Optional[int] = Field(
+        24, description='Frames per second of the generated video'
+    )
+    guidance_scale: Optional[float] = Field(
+        12.5, description='Guidance scale for generation control'
+    )
+    height: Optional[int] = Field(
+        1080, description='Height of the generated video in pixels'
+    )
+    negative_prompt: Optional[str] = Field(None, description='Negative prompt text')
+    num_frames: Optional[int] = Field(64, description='Number of frames to generate')
+    seed: Optional[int] = Field(
+        None, description='Random seed for generation (default: random)'
+    )
+    shift_value: Optional[float] = Field(
+        3, description='Shift value for generation control'
+    )
+    steps: Optional[int] = Field(80, description='Number of denoising steps')
+    use_guidance_schedule: Optional[bool] = Field(
+        True, description='Whether to use guidance scheduling'
+    )
+    use_negative_prompts: Optional[bool] = Field(
+        False, description='Whether to use negative prompts'
+    )
+    use_timestep_transform: Optional[bool] = Field(
+        True, description='Whether to use timestep transformation'
+    )
+    warmup_steps: Optional[int] = Field(
+        None, description='Number of warmup steps (calculated based on num_frames)'
+    )
+    width: Optional[int] = Field(
+        1920, description='Width of the generated video in pixels'
+    )
+
+
+class MoonvalleyTextToVideoRequest(BaseModel):
+    image_url: Optional[str] = None
+    inference_params: Optional[MoonvalleyTextToVideoInferenceParams] = None
+    prompt_text: Optional[str] = None
+    webhook_url: Optional[str] = None
+
+
+class MoonvalleyUploadFileRequest(BaseModel):
+    file: Optional[StrictBytes] = None
+
+
+class MoonvalleyUploadFileResponse(BaseModel):
+    access_url: Optional[str] = None
+
+
+class MoonvalleyVideoToVideoInferenceParams(BaseModel):
+    add_quality_guidance: Optional[bool] = Field(
+        True, description='Whether to add quality guidance'
+    )
+    caching_coefficient: Optional[float] = Field(
+        0.3, description='Caching coefficient for optimization'
+    )
+    caching_cooldown: Optional[int] = Field(
+        3, description='Number of caching cooldown steps'
+    )
+    caching_warmup: Optional[int] = Field(
+        3, description='Number of caching warmup steps'
+    )
+    clip_value: Optional[float] = Field(
+        3, description='CLIP value for generation control'
+    )
+    conditioning_frame_index: Optional[int] = Field(
+        0, description='Index of the conditioning frame'
+    )
+    cooldown_steps: Optional[int] = Field(
+        None, description='Number of cooldown steps (calculated based on num_frames)'
+    )
+    guidance_scale: Optional[float] = Field(
+        12.5, description='Guidance scale for generation control'
+    )
+    negative_prompt: Optional[str] = Field(None, description='Negative prompt text')
+    seed: Optional[int] = Field(
+        None, description='Random seed for generation (default: random)'
+    )
+    shift_value: Optional[float] = Field(
+        3, description='Shift value for generation control'
+    )
+    steps: Optional[int] = Field(80, description='Number of denoising steps')
+    use_guidance_schedule: Optional[bool] = Field(
+        True, description='Whether to use guidance scheduling'
+    )
+    use_negative_prompts: Optional[bool] = Field(
+        False, description='Whether to use negative prompts'
+    )
+    use_timestep_transform: Optional[bool] = Field(
+        True, description='Whether to use timestep transformation'
+    )
+    warmup_steps: Optional[int] = Field(
+        None, description='Number of warmup steps (calculated based on num_frames)'
+    )
+
+
+class ControlType(str, Enum):
+    motion_control = 'motion_control'
+    pose_control = 'pose_control'
+
+
+class MoonvalleyVideoToVideoRequest(BaseModel):
+    control_type: ControlType = Field(
+        ..., description='Supported types for video control'
+    )
+    inference_params: Optional[MoonvalleyVideoToVideoInferenceParams] = None
+    prompt_text: str = Field(..., description='Describes the video to generate')
+    video_url: str = Field(..., description='Url to control video')
+    webhook_url: Optional[str] = Field(
+        None, description='Optional webhook URL for notifications'
+    )
+
+
 class Moderation(str, Enum):
     low = 'low'
     auto = 'auto'
@@ -3107,6 +3259,23 @@ class LumaUpscaleVideoGenerationRequest(BaseModel):
     resolution: Optional[LumaVideoModelOutputResolution] = None
 
 
+class MoonvalleyImageToVideoRequest(MoonvalleyTextToVideoRequest):
+    keyframes: Optional[Dict[str, Keyframes]] = None
+
+
+class MoonvalleyResizeVideoRequest(MoonvalleyVideoToVideoRequest):
+    frame_position: Optional[List[int]] = Field(None, max_length=2, min_length=2)
+    frame_resolution: Optional[List[int]] = Field(None, max_length=2, min_length=2)
+    scale: Optional[List[int]] = Field(None, max_length=2, min_length=2)
+
+
+class MoonvalleyTextToImageRequest(BaseModel):
+    image_url: Optional[str] = None
+    inference_params: Optional[MoonvalleyTextToVideoInferenceParams] = None
+    prompt_text: Optional[str] = None
+    webhook_url: Optional[str] = None
+
+
 class OutputContent(RootModel[Union[OutputTextContent, OutputAudioContent]]):
     root: Union[OutputTextContent, OutputAudioContent]
 

comfy_api_nodes/nodes_moonvalley.py

@@ -0,0 +1,639 @@
+import logging
+from typing import Any, Callable, Optional, TypeVar
+import random
+import torch
+from comfy_api_nodes.util.validation_utils import get_image_dimensions, validate_image_dimensions, validate_video_dimensions
+
+
+from comfy_api_nodes.apis import (
+    MoonvalleyTextToVideoRequest,
+    MoonvalleyTextToVideoInferenceParams,
+    MoonvalleyVideoToVideoInferenceParams,
+    MoonvalleyVideoToVideoRequest,
+    MoonvalleyPromptResponse
+)
+from comfy_api_nodes.apis.client import (
+    ApiEndpoint,
+    HttpMethod,
+    SynchronousOperation,
+    PollingOperation,
+    EmptyRequest,
+)
+from comfy_api_nodes.apinode_utils import (
+    download_url_to_video_output,
+    upload_images_to_comfyapi,
+    upload_video_to_comfyapi,
+)
+from comfy_api_nodes.mapper_utils import model_field_to_node_input
+
+from comfy_api.input.video_types import VideoInput
+from comfy.comfy_types.node_typing import IO
+from comfy_api.input_impl import VideoFromFile
+import av
+import io
+
+API_UPLOADS_ENDPOINT = "/proxy/moonvalley/uploads"
+API_PROMPTS_ENDPOINT = "/proxy/moonvalley/prompts"
+API_VIDEO2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/video-to-video"
+API_TXT2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/text-to-video"
+API_IMG2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/image-to-video"
+
+MIN_WIDTH = 300
+MIN_HEIGHT = 300
+
+MAX_WIDTH = 10000
+MAX_HEIGHT = 10000
+
+MIN_VID_WIDTH = 300
+MIN_VID_HEIGHT = 300
+
+MAX_VID_WIDTH = 10000
+MAX_VID_HEIGHT = 10000
+
+MAX_VIDEO_SIZE = 1024 * 1024 * 1024  # 1 GB max for in-memory video processing
+
+MOONVALLEY_MAREY_MAX_PROMPT_LENGTH = 5000
+R = TypeVar("R")
+class MoonvalleyApiError(Exception):
+    """Base exception for Moonvalley API errors."""
+    pass
+
+def is_valid_task_creation_response(response: MoonvalleyPromptResponse) -> bool:
+    """Verifies that the initial response contains a task ID."""
+    return bool(response.id)
+
+def validate_task_creation_response(response) -> None:
+    if not is_valid_task_creation_response(response):
+        error_msg = f"Moonvalley Marey API: Initial request failed. Code: {response.code}, Message: {response.message}, Data: {response}"
+        logging.error(error_msg)
+        raise MoonvalleyApiError(error_msg)
+
+def get_video_from_response(response):
+    video = response.output_url
+    logging.info(
+        "Moonvalley Marey API: Task %s succeeded. Video URL: %s", response.id, video
+    )
+    return video
+
+
+def get_video_url_from_response(response) -> Optional[str]:
+    """Returns the first video url from the Moonvalley video generation task result.
+    Will not raise an error if the response is not valid.
+    """
+    if response:
+        return str(get_video_from_response(response))
+    else:
+        return None
+
+
+def poll_until_finished(
+    auth_kwargs: dict[str, str],
+    api_endpoint: ApiEndpoint[Any, R],
+    result_url_extractor: Optional[Callable[[R], str]] = None,
+    node_id: Optional[str] = None,
+) -> R:
+    """Polls the Moonvalley API endpoint until the task reaches a terminal state, then returns the response."""
+    return PollingOperation(
+        poll_endpoint=api_endpoint,
+        completed_statuses=[
+            "completed",
+        ],
+        max_poll_attempts=240,  # 64 minutes with 16s interval
+        poll_interval=16.0,
+        failed_statuses=["error"],
+        status_extractor=lambda response: (
+            response.status
+            if response and response.status
+            else None
+        ),
+        auth_kwargs=auth_kwargs,
+        result_url_extractor=result_url_extractor,
+        node_id=node_id,
+    ).execute()
+
+def validate_prompts(prompt:str, negative_prompt: str, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH):
+    """Verifies that the prompt isn't empty and that neither prompt is too long."""
+    if not prompt:
+        raise ValueError("Positive prompt is empty")
+    if len(prompt) > max_length:
+        raise ValueError(f"Positive prompt is too long: {len(prompt)} characters")
+    if negative_prompt and len(negative_prompt) > max_length:
+        raise ValueError(
+            f"Negative prompt is too long: {len(negative_prompt)} characters"
+        )
+    return True
+
+def validate_input_media(width, height, with_frame_conditioning, num_frames_in=None):
+        # inference validation
+        # T = num_frames
+        # in all cases, the following must be true: T divisible by 16 and H,W by 8. in addition...
+        # with image conditioning: H*W must be divisible by 8192
+        # without image conditioning: T divisible by 32
+    if num_frames_in and not num_frames_in % 16 == 0 :
+        return False, (
+            "The input video total frame count must be divisible by 16!"
+        )
+
+    if height % 8 != 0 or width % 8 != 0:
+        return False, (
+            f"Height ({height}) and width ({width}) must be " "divisible by 8"
+        )
+
+    if with_frame_conditioning:
+        if (height * width) % 8192 != 0:
+            return False, (
+                f"Height * width ({height * width}) must be "
+                "divisible by 8192 for frame conditioning"
+            )
+    else:
+        if num_frames_in and not num_frames_in % 32 == 0 :
+            return False, (
+                "The input video total frame count must be divisible by 32!"
+            )
+
+
+def validate_input_image(image: torch.Tensor, with_frame_conditioning: bool=False) -> None:
+    """
+    Validates the input image adheres to the expectations of the API:
+    - The image resolution should not be less than 300*300px
+    - The aspect ratio of the image should be between 1:2.5 ~ 2.5:1
+
+    """
+    height, width = get_image_dimensions(image)
+    validate_input_media(width, height, with_frame_conditioning )
+    validate_image_dimensions(image, min_width=300, min_height=300, max_height=MAX_HEIGHT, max_width=MAX_WIDTH)
+
+def validate_input_video(video: VideoInput, num_frames_out: int, with_frame_conditioning: bool=False):
+    try:
+        width, height = video.get_dimensions()
+    except Exception as e:
+        logging.error("Error getting dimensions of video: %s", e)
+        raise ValueError(f"Cannot get video dimensions: {e}") from e
+
+    validate_input_media(width, height, with_frame_conditioning)
+    validate_video_dimensions(video, min_width=MIN_VID_WIDTH, min_height=MIN_VID_HEIGHT, max_width=MAX_VID_WIDTH, max_height=MAX_VID_HEIGHT)
+
+    trimmed_video = validate_input_video_length(video, num_frames_out)
+    return trimmed_video
+
+
+def validate_input_video_length(video: VideoInput, num_frames: int):
+
+    if video.get_duration() > 60:
+        raise MoonvalleyApiError("Input Video lenth should be less than 1min. Please trim.")
+
+    if num_frames == 128:
+       if video.get_duration() < 5:
+           raise MoonvalleyApiError("Input Video length is less than 5s. Please use a video longer than or equal to 5s.")
+       if video.get_duration() > 5:
+        #    trim video to 5s
+        video = trim_video(video, 5)
+    if num_frames == 256:
+        if video.get_duration() < 10:
+            raise MoonvalleyApiError("Input Video length is less than 10s. Please use a video longer than or equal to 10s.")
+        if video.get_duration() > 10:
+            # trim video to 10s
+            video = trim_video(video, 10)
+    return video
+
+def trim_video(video: VideoInput, duration_sec: float) -> VideoInput:
+    """
+    Returns a new VideoInput object trimmed from the beginning to the specified duration,
+    using av to avoid loading entire video into memory.
+
+    Args:
+        video: Input video to trim
+        duration_sec: Duration in seconds to keep from the beginning
+
+    Returns:
+        VideoFromFile object that owns the output buffer
+    """
+    output_buffer = io.BytesIO()
+
+    input_container = None
+    output_container = None
+
+    try:
+        # Get the stream source - this avoids loading entire video into memory
+        # when the source is already a file path
+        input_source = video.get_stream_source()
+
+        # Open containers
+        input_container = av.open(input_source, mode='r')
+        output_container = av.open(output_buffer, mode='w', format='mp4')
+
+        # Set up output streams for re-encoding
+        video_stream = None
+        audio_stream = None
+
+        for stream in input_container.streams:
+            logging.info(f"Found stream: type={stream.type}, class={type(stream)}")
+            if isinstance(stream, av.VideoStream):
+                # Create output video stream with same parameters
+                video_stream = output_container.add_stream('h264', rate=stream.average_rate)
+                video_stream.width = stream.width
+                video_stream.height = stream.height
+                video_stream.pix_fmt = 'yuv420p'
+                logging.info(f"Added video stream: {stream.width}x{stream.height} @ {stream.average_rate}fps")
+            elif isinstance(stream, av.AudioStream):
+                # Create output audio stream with same parameters
+                audio_stream = output_container.add_stream('aac', rate=stream.sample_rate)
+                audio_stream.sample_rate = stream.sample_rate
+                audio_stream.layout = stream.layout
+                logging.info(f"Added audio stream: {stream.sample_rate}Hz, {stream.channels} channels")
+
+        # Calculate target frame count that's divisible by 32
+        fps = input_container.streams.video[0].average_rate
+        estimated_frames = int(duration_sec * fps)
+        target_frames = (estimated_frames // 32) * 32  # Round down to nearest multiple of 32
+
+        if target_frames == 0:
+            raise ValueError("Video too short: need at least 32 frames for Moonvalley")
+
+        frame_count = 0
+        audio_frame_count = 0
+
+        # Decode and re-encode video frames
+        if video_stream:
+            for frame in input_container.decode(video=0):
+                if frame_count >= target_frames:
+                    break
+
+                # Re-encode frame
+                for packet in video_stream.encode(frame):
+                    output_container.mux(packet)
+                frame_count += 1
+
+            # Flush encoder
+            for packet in video_stream.encode():
+                output_container.mux(packet)
+
+            logging.info(f"Encoded {frame_count} video frames (target: {target_frames})")
+
+        # Decode and re-encode audio frames
+        if audio_stream:
+            input_container.seek(0)  # Reset to beginning for audio
+            for frame in input_container.decode(audio=0):
+                if frame.time >= duration_sec:
+                    break
+
+                # Re-encode frame
+                for packet in audio_stream.encode(frame):
+                    output_container.mux(packet)
+                audio_frame_count += 1
+
+            # Flush encoder
+            for packet in audio_stream.encode():
+                output_container.mux(packet)
+
+            logging.info(f"Encoded {audio_frame_count} audio frames")
+
+        # Close containers
+        output_container.close()
+        input_container.close()
+
+
+        # Return as VideoFromFile using the buffer
+        output_buffer.seek(0)
+        return VideoFromFile(output_buffer)
+
+    except Exception as e:
+        # Clean up on error
+        if input_container is not None:
+            input_container.close()
+        if output_container is not None:
+            output_container.close()
+        raise RuntimeError(f"Failed to trim video: {str(e)}") from e
+
+# --- BaseMoonvalleyVideoNode ---
+class BaseMoonvalleyVideoNode:
+    def parseWidthHeightFromRes(self, resolution: str):
+        # Accepts a string like "16:9 (1920 x 1080)" and returns width, height as a dict
+        res_map = {
+            "16:9 (1920 x 1080)": {"width": 1920, "height": 1080},
+            "9:16 (1080 x 1920)": {"width": 1080, "height": 1920},
+            "1:1 (1152 x 1152)": {"width": 1152, "height": 1152},
+            "4:3 (1440 x 1080)": {"width": 1440, "height": 1080},
+            "3:4 (1080 x 1440)": {"width": 1080, "height": 1440},
+            "21:9 (2560 x 1080)": {"width": 2560, "height": 1080},
+        }
+        if resolution in res_map:
+            return res_map[resolution]
+        else:
+            # Default to 1920x1080 if unknown
+            return {"width": 1920, "height": 1080}
+
+    def parseControlParameter(self, value):
+        control_map = {
+            "Motion Transfer": "motion_control",
+            "Canny": "canny_control",
+            "Pose Transfer": "pose_control",
+            "Depth": "depth_control"
+        }
+        if value in control_map:
+            return control_map[value]
+        else:
+            return control_map["Motion Transfer"]
+
+    def get_response(
+        self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
+    ) -> MoonvalleyPromptResponse:
+        return poll_until_finished(
+            auth_kwargs,
+            ApiEndpoint(
+                path=f"{API_PROMPTS_ENDPOINT}/{task_id}",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MoonvalleyPromptResponse,
+            ),
+            result_url_extractor=get_video_url_from_response,
+            node_id=node_id,
+        )
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "prompt": model_field_to_node_input(
+                    IO.STRING, MoonvalleyTextToVideoRequest, "prompt_text",
+                    multiline=True
+                ),
+                "negative_prompt": model_field_to_node_input(
+                    IO.STRING,
+                    MoonvalleyTextToVideoInferenceParams,
+                    "negative_prompt",
+                    multiline=True,
+                    default="gopro, bright, contrast, static, overexposed, bright, vignette, artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, flare, saturation, distorted, warped, wide angle, contrast, saturated, vibrant, glowing, cross dissolve, texture, videogame, saturation, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, blown out, horrible, blurry, worst quality, bad, transition, dissolve, cross-dissolve, melt, fade in, fade out, wobbly, weird, low quality, plastic, stock footage, video camera, boring, static",
+                ),
+
+                "resolution": (IO.COMBO, {
+                        "options": ["16:9 (1920 x 1080)",
+                                    "9:16 (1080 x 1920)",
+                                    "1:1 (1152 x 1152)",
+                                    "4:3 (1440 x 1080)",
+                                    "3:4 (1080 x 1440)",
+                                    "21:9 (2560 x 1080)"],
+                        "default": "16:9 (1920 x 1080)",
+                        "tooltip": "Resolution of the output video",
+                    }),
+                # "length": (IO.COMBO,{"options":['5s','10s'], "default": '5s'}),
+                "prompt_adherence": model_field_to_node_input(IO.FLOAT,MoonvalleyTextToVideoInferenceParams,"guidance_scale",default=7.0, step=1, min=1, max=20),
+                "seed": model_field_to_node_input(IO.INT,MoonvalleyTextToVideoInferenceParams, "seed", default=random.randint(0, 2**32 - 1), min=0, max=4294967295, step=1, display="number", tooltip="Random seed value", control_after_generate=True),
+                "steps": model_field_to_node_input(IO.INT, MoonvalleyTextToVideoInferenceParams, "steps", default=100, min=1, max=100),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+            "optional": {
+                "image": model_field_to_node_input(
+                    IO.IMAGE,
+                    MoonvalleyTextToVideoRequest,
+                    "image_url",
+                    tooltip="The reference image used to generate the video",
+                ),
+            }
+        }
+
+    RETURN_TYPES = ("STRING",)
+    FUNCTION = "generate"
+    CATEGORY = "api node/video/Moonvalley Marey"
+    API_NODE = True
+
+    def generate(self, **kwargs):
+        return None
+
+# --- MoonvalleyImg2VideoNode ---
+class MoonvalleyImg2VideoNode(BaseMoonvalleyVideoNode):
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return super().INPUT_TYPES()
+
+    RETURN_TYPES = ("VIDEO",)
+    RETURN_NAMES = ("video",)
+    DESCRIPTION = "Moonvalley Marey Image to Video Node"
+
+    def generate(self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs):
+        image = kwargs.get("image", None)
+        if (image is None):
+            raise MoonvalleyApiError("image is required")
+        total_frames = get_total_frames_from_length()
+
+        validate_input_image(image,True)
+        validate_prompts(prompt, negative_prompt, MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
+        width_height = self.parseWidthHeightFromRes(kwargs.get("resolution"))
+
+        inference_params=MoonvalleyTextToVideoInferenceParams(
+                    negative_prompt=negative_prompt,
+                    steps=kwargs.get("steps"),
+                    seed=kwargs.get("seed"),
+                    guidance_scale=kwargs.get("prompt_adherence"),
+                    num_frames=total_frames,
+                    width=width_height.get("width"),
+                    height=width_height.get("height"),
+                    use_negative_prompts=True
+                )
+        """Upload image to comfy backend to have a URL available for further processing"""
+        # Get MIME type from tensor - assuming PNG format for image tensors
+        mime_type = "image/png"
+
+        image_url = upload_images_to_comfyapi(image, max_images=1, auth_kwargs=kwargs, mime_type=mime_type)[0]
+
+        request = MoonvalleyTextToVideoRequest(
+                image_url=image_url,
+                prompt_text=prompt,
+                inference_params=inference_params
+            )
+        initial_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(path=API_IMG2VIDEO_ENDPOINT,
+                                 method=HttpMethod.POST,
+                                 request_model=MoonvalleyTextToVideoRequest,
+                                 response_model=MoonvalleyPromptResponse
+                                 ),
+            request=request,
+            auth_kwargs=kwargs,
+        )
+        task_creation_response = initial_operation.execute()
+        validate_task_creation_response(task_creation_response)
+        task_id = task_creation_response.id
+
+        final_response = self.get_response(
+            task_id, auth_kwargs=kwargs, node_id=unique_id
+        )
+        video = download_url_to_video_output(final_response.output_url)
+        return (video, )
+
+# --- MoonvalleyVid2VidNode ---
+class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
+    def __init__(self):
+        super().__init__()
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        input_types = super().INPUT_TYPES()
+        for param in ["resolution", "image"]:
+            if param in input_types["required"]:
+                del input_types["required"][param]
+            if param in input_types["optional"]:
+                del input_types["optional"][param]
+        input_types["optional"] = {
+                "video": (IO.VIDEO, {"default": "", "multiline": False, "tooltip": "The reference video used to generate the output video. Input a 5s video for 128 frames and a 10s video for 256 frames. Longer videos will be trimmed automatically."}),
+                "control_type": (
+                    ["Motion Transfer", "Pose Transfer"],
+                    {"default": "Motion Transfer"},
+                ),
+                "motion_intensity": (
+                    "INT",
+                    {
+                        "default": 100,
+                        "step": 1,
+                        "min": 0,
+                        "max": 100,
+                        "tooltip": "Only used if control_type is 'Motion Transfer'",
+                    },
+                )
+            }
+
+        return input_types
+
+    RETURN_TYPES = ("VIDEO",)
+    RETURN_NAMES = ("video",)
+
+    def generate(self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs):
+        video = kwargs.get("video")
+        num_frames = get_total_frames_from_length()
+
+        if not video :
+            raise MoonvalleyApiError("video is required")
+
+
+        """Validate video input"""
+        video_url=""
+        if video:
+            validated_video = validate_input_video(video, num_frames, False)
+            video_url = upload_video_to_comfyapi(validated_video, auth_kwargs=kwargs)
+
+        control_type = kwargs.get("control_type")
+        motion_intensity = kwargs.get("motion_intensity")
+
+        """Validate prompts and inference input"""
+        validate_prompts(prompt, negative_prompt)
+        inference_params=MoonvalleyVideoToVideoInferenceParams(
+            negative_prompt=negative_prompt,
+            steps=kwargs.get("steps"),
+            seed=kwargs.get("seed"),
+            guidance_scale=kwargs.get("prompt_adherence"),
+            control_params={'motion_intensity': motion_intensity}
+        )
+
+        control = self.parseControlParameter(control_type)
+
+        request = MoonvalleyVideoToVideoRequest(
+                control_type=control,
+                video_url=video_url,
+                prompt_text=prompt,
+                inference_params=inference_params
+            )
+
+        initial_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(path=API_VIDEO2VIDEO_ENDPOINT,
+                                 method=HttpMethod.POST,
+                                 request_model=MoonvalleyVideoToVideoRequest,
+                                 response_model=MoonvalleyPromptResponse
+                                 ),
+            request=request,
+            auth_kwargs=kwargs,
+        )
+        task_creation_response = initial_operation.execute()
+        validate_task_creation_response(task_creation_response)
+        task_id = task_creation_response.id
+
+        final_response = self.get_response(
+            task_id, auth_kwargs=kwargs, node_id=unique_id
+        )
+
+        video = download_url_to_video_output(final_response.output_url)
+
+        return (video, )
+
+# --- MoonvalleyTxt2VideoNode ---
+class MoonvalleyTxt2VideoNode(BaseMoonvalleyVideoNode):
+    def __init__(self):
+        super().__init__()
+
+    RETURN_TYPES = ("VIDEO",)
+    RETURN_NAMES = ("video",)
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        input_types = super().INPUT_TYPES()
+        # Remove image-specific parameters
+        for param in ["image"]:
+            if param in input_types["optional"]:
+                del input_types["optional"][param]
+        return input_types
+
+    def generate(self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs):
+        validate_prompts(prompt, negative_prompt, MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
+        width_height = self.parseWidthHeightFromRes(kwargs.get("resolution"))
+        num_frames = get_total_frames_from_length()
+
+        inference_params=MoonvalleyTextToVideoInferenceParams(
+                    negative_prompt=negative_prompt,
+                    steps=kwargs.get("steps"),
+                    seed=kwargs.get("seed"),
+                    guidance_scale=kwargs.get("prompt_adherence"),
+                    num_frames=num_frames,
+                    width=width_height.get("width"),
+                    height=width_height.get("height"),
+                )
+        request = MoonvalleyTextToVideoRequest(
+                prompt_text=prompt,
+                inference_params=inference_params
+            )
+
+        initial_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(path=API_TXT2VIDEO_ENDPOINT,
+                                 method=HttpMethod.POST,
+                                 request_model=MoonvalleyTextToVideoRequest,
+                                 response_model=MoonvalleyPromptResponse
+                                 ),
+            request=request,
+            auth_kwargs=kwargs,
+        )
+        task_creation_response = initial_operation.execute()
+        validate_task_creation_response(task_creation_response)
+        task_id = task_creation_response.id
+
+        final_response = self.get_response(
+            task_id, auth_kwargs=kwargs, node_id=unique_id
+        )
+
+        video = download_url_to_video_output(final_response.output_url)
+        return (video, )
+
+
+
+NODE_CLASS_MAPPINGS = {
+    "MoonvalleyImg2VideoNode": MoonvalleyImg2VideoNode,
+    "MoonvalleyTxt2VideoNode": MoonvalleyTxt2VideoNode,
+    # "MoonvalleyVideo2VideoNode": MoonvalleyVideo2VideoNode,
+}
+
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "MoonvalleyImg2VideoNode": "Moonvalley Marey Image to Video",
+    "MoonvalleyTxt2VideoNode": "Moonvalley Marey Text to Video",
+    # "MoonvalleyVideo2VideoNode": "Moonvalley Marey Video to Video",
+}
+
+def get_total_frames_from_length(length="5s"):
+    # if length == '5s':
+    #     return 128
+    # elif length == '10s':
+    #     return 256
+    return 128
+    # else:
+    #     raise MoonvalleyApiError("length is required")

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_extras/nodes_audio.py

@@ -133,14 +133,6 @@ def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=Non
             if sample_rate != audio["sample_rate"]:
                 waveform = torchaudio.functional.resample(waveform, audio["sample_rate"], sample_rate)
 
-        # Create in-memory WAV buffer
-        wav_buffer = io.BytesIO()
-        torchaudio.save(wav_buffer, waveform, sample_rate, format="WAV")
-        wav_buffer.seek(0)  # Rewind for reading
-
-        # Use PyAV to convert and add metadata
-        input_container = av.open(wav_buffer)
-
         # Create output with specified format
         output_buffer = io.BytesIO()
         output_container = av.open(output_buffer, mode='w', format=format)
@@ -150,7 +142,6 @@ def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=Non
             output_container.metadata[key] = value
 
         # Set up the output stream with appropriate properties
-        input_container.streams.audio[0]
         if format == "opus":
             out_stream = output_container.add_stream("libopus", rate=sample_rate)
             if quality == "64k":
@@ -175,18 +166,16 @@ def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=Non
         else: #format == "flac":
             out_stream = output_container.add_stream("flac", rate=sample_rate)
 
-
-        # Copy frames from input to output
-        for frame in input_container.decode(audio=0):
-            frame.pts = None  # Let PyAV handle timestamps
-            output_container.mux(out_stream.encode(frame))
+        frame = av.AudioFrame.from_ndarray(waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[0] == 1 else 'stereo')
+        frame.sample_rate = sample_rate
+        frame.pts = 0
+        output_container.mux(out_stream.encode(frame))
 
         # Flush encoder
         output_container.mux(out_stream.encode(None))
 
         # Close containers
         output_container.close()
-        input_container.close()
 
         # Write the output to file
         output_buffer.seek(0)

comfy_extras/nodes_cosmos.py

@@ -2,6 +2,7 @@ import nodes
 import torch
 import comfy.model_management
 import comfy.utils
+import comfy.latent_formats
 
 
 class EmptyCosmosLatentVideo:
@@ -75,8 +76,53 @@ class CosmosImageToVideoLatent:
         out_latent["noise_mask"] = mask.repeat((batch_size, ) + (1,) * (mask.ndim - 1))
         return (out_latent,)
 
+class CosmosPredict2ImageToVideoLatent:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"vae": ("VAE", ),
+                             "width": ("INT", {"default": 848, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "length": ("INT", {"default": 93, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                },
+                "optional": {"start_image": ("IMAGE", ),
+                             "end_image": ("IMAGE", ),
+                }}
+
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/inpaint"
+
+    def encode(self, vae, width, height, length, batch_size, start_image=None, end_image=None):
+        latent = torch.zeros([1, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        if start_image is None and end_image is None:
+            out_latent = {}
+            out_latent["samples"] = latent
+            return (out_latent,)
+
+        mask = torch.ones([latent.shape[0], 1, ((length - 1) // 4) + 1, latent.shape[-2], latent.shape[-1]], device=comfy.model_management.intermediate_device())
+
+        if start_image is not None:
+            latent_temp = vae_encode_with_padding(vae, start_image, width, height, length, padding=1)
+            latent[:, :, :latent_temp.shape[-3]] = latent_temp
+            mask[:, :, :latent_temp.shape[-3]] *= 0.0
+
+        if end_image is not None:
+            latent_temp = vae_encode_with_padding(vae, end_image, width, height, length, padding=0)
+            latent[:, :, -latent_temp.shape[-3]:] = latent_temp
+            mask[:, :, -latent_temp.shape[-3]:] *= 0.0
+
+        out_latent = {}
+        latent_format = comfy.latent_formats.Wan21()
+        latent = latent_format.process_out(latent) * mask + latent * (1.0 - mask)
+        out_latent["samples"] = latent.repeat((batch_size, ) + (1,) * (latent.ndim - 1))
+        out_latent["noise_mask"] = mask.repeat((batch_size, ) + (1,) * (mask.ndim - 1))
+        return (out_latent,)
 
 NODE_CLASS_MAPPINGS = {
     "EmptyCosmosLatentVideo": EmptyCosmosLatentVideo,
     "CosmosImageToVideoLatent": CosmosImageToVideoLatent,
+    "CosmosPredict2ImageToVideoLatent": CosmosPredict2ImageToVideoLatent,
 }

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:
     """
@@ -523,6 +583,49 @@ class GetImageSize:
 
         return width, height, batch_size
 
+class ImageRotate:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": (IO.IMAGE,),
+                              "rotation": (["none", "90 degrees", "180 degrees", "270 degrees"],),
+                              }}
+    RETURN_TYPES = (IO.IMAGE,)
+    FUNCTION = "rotate"
+
+    CATEGORY = "image/transform"
+
+    def rotate(self, image, rotation):
+        rotate_by = 0
+        if rotation.startswith("90"):
+            rotate_by = 1
+        elif rotation.startswith("180"):
+            rotate_by = 2
+        elif rotation.startswith("270"):
+            rotate_by = 3
+
+        image = torch.rot90(image, k=rotate_by, dims=[2, 1])
+        return (image,)
+
+class ImageFlip:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": (IO.IMAGE,),
+                              "flip_method": (["x-axis: vertically", "y-axis: horizontally"],),
+                              }}
+    RETURN_TYPES = (IO.IMAGE,)
+    FUNCTION = "flip"
+
+    CATEGORY = "image/transform"
+
+    def flip(self, image, flip_method):
+        if flip_method.startswith("x"):
+            image = torch.flip(image, dims=[1])
+        elif flip_method.startswith("y"):
+            image = torch.flip(image, dims=[2])
+
+        return (image,)
+
+
 NODE_CLASS_MAPPINGS = {
     "ImageCrop": ImageCrop,
     "RepeatImageBatch": RepeatImageBatch,
@@ -532,5 +635,8 @@ NODE_CLASS_MAPPINGS = {
     "SaveAnimatedPNG": SaveAnimatedPNG,
     "SaveSVGNode": SaveSVGNode,
     "ImageStitch": ImageStitch,
+    "ResizeAndPadImage": ResizeAndPadImage,
     "GetImageSize": GetImageSize,
+    "ImageRotate": ImageRotate,
+    "ImageFlip": ImageFlip,
 }

comfy_extras/nodes_load_3d.py

@@ -5,6 +5,8 @@ import os
 from comfy.comfy_types import IO
 from comfy_api.input_impl import VideoFromFile
 
+from pathlib import Path
+
 
 def normalize_path(path):
     return path.replace('\\', '/')
@@ -16,7 +18,14 @@ class Load3D():
 
         os.makedirs(input_dir, exist_ok=True)
 
-        files = [normalize_path(os.path.join("3d", f)) for f in os.listdir(input_dir) if f.endswith(('.gltf', '.glb', '.obj', '.fbx', '.stl'))]
+        input_path = Path(input_dir)
+        base_path = Path(folder_paths.get_input_directory())
+
+        files = [
+            normalize_path(str(file_path.relative_to(base_path)))
+            for file_path in input_path.rglob("*")
+            if file_path.suffix.lower() in {'.gltf', '.glb', '.obj', '.fbx', '.stl'}
+        ]
 
         return {"required": {
             "model_file": (sorted(files), {"file_upload": True}),
@@ -61,7 +70,14 @@ class Load3DAnimation():
 
         os.makedirs(input_dir, exist_ok=True)
 
-        files = [normalize_path(os.path.join("3d", f)) for f in os.listdir(input_dir) if f.endswith(('.gltf', '.glb', '.fbx'))]
+        input_path = Path(input_dir)
+        base_path = Path(folder_paths.get_input_directory())
+
+        files = [
+            normalize_path(str(file_path.relative_to(base_path)))
+            for file_path in input_path.rglob("*")
+            if file_path.suffix.lower() in {'.gltf', '.glb', '.fbx'}
+        ]
 
         return {"required": {
             "model_file": (sorted(files), {"file_upload": True}),

comfy_extras/nodes_lt.py

@@ -134,8 +134,8 @@ class LTXVAddGuide:
         _, num_keyframes = get_keyframe_idxs(cond)
         latent_count = latent_length - num_keyframes
         frame_idx = frame_idx if frame_idx >= 0 else max((latent_count - 1) * time_scale_factor + 1 + frame_idx, 0)
-        if guide_length > 1:
-            frame_idx = frame_idx // time_scale_factor * time_scale_factor # frame index must be divisible by 8
+        if guide_length > 1 and frame_idx != 0:
+            frame_idx = (frame_idx - 1) // time_scale_factor * time_scale_factor + 1 # frame index - 1 must be divisible by 8 or frame_idx == 0
 
         latent_idx = (frame_idx + time_scale_factor - 1) // time_scale_factor
 
@@ -144,7 +144,7 @@ class LTXVAddGuide:
     def add_keyframe_index(self, cond, frame_idx, guiding_latent, scale_factors):
         keyframe_idxs, _ = get_keyframe_idxs(cond)
         _, latent_coords = self._patchifier.patchify(guiding_latent)
-        pixel_coords = latent_to_pixel_coords(latent_coords, scale_factors, True)
+        pixel_coords = latent_to_pixel_coords(latent_coords, scale_factors, causal_fix=frame_idx == 0)  # we need the causal fix only if we're placing the new latents at index 0
         pixel_coords[:, 0] += frame_idx
         if keyframe_idxs is None:
             keyframe_idxs = pixel_coords

comfy_extras/nodes_mask.py

@@ -152,7 +152,7 @@ class ImageColorToMask:
     def image_to_mask(self, image, color):
         temp = (torch.clamp(image, 0, 1.0) * 255.0).round().to(torch.int)
         temp = torch.bitwise_left_shift(temp[:,:,:,0], 16) + torch.bitwise_left_shift(temp[:,:,:,1], 8) + temp[:,:,:,2]
-        mask = torch.where(temp == color, 255, 0).float()
+        mask = torch.where(temp == color, 1.0, 0).float()
         return (mask,)
 
 class SolidMask:

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_slg.py

@@ -78,7 +78,75 @@ class SkipLayerGuidanceDiT:
 
         return (m, )
 
+class SkipLayerGuidanceDiTSimple:
+    '''
+    Simple version of the SkipLayerGuidanceDiT node that only modifies the uncond pass.
+    '''
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"model": ("MODEL", ),
+                             "double_layers": ("STRING", {"default": "7, 8, 9", "multiline": False}),
+                             "single_layers": ("STRING", {"default": "7, 8, 9", "multiline": False}),
+                             "start_percent": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001}),
+                             "end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001}),
+                                }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "skip_guidance"
+    EXPERIMENTAL = True
+
+    DESCRIPTION = "Simple version of the SkipLayerGuidanceDiT node that only modifies the uncond pass."
+
+    CATEGORY = "advanced/guidance"
+
+    def skip_guidance(self, model, start_percent, end_percent, double_layers="", single_layers=""):
+        def skip(args, extra_args):
+            return args
+
+        model_sampling = model.get_model_object("model_sampling")
+        sigma_start = model_sampling.percent_to_sigma(start_percent)
+        sigma_end = model_sampling.percent_to_sigma(end_percent)
+
+        double_layers = re.findall(r'\d+', double_layers)
+        double_layers = [int(i) for i in double_layers]
+
+        single_layers = re.findall(r'\d+', single_layers)
+        single_layers = [int(i) for i in single_layers]
+
+        if len(double_layers) == 0 and len(single_layers) == 0:
+            return (model, )
+
+        def calc_cond_batch_function(args):
+            x = args["input"]
+            model = args["model"]
+            conds = args["conds"]
+            sigma = args["sigma"]
+
+            model_options = args["model_options"]
+            slg_model_options = model_options.copy()
+
+            for layer in double_layers:
+                slg_model_options = comfy.model_patcher.set_model_options_patch_replace(slg_model_options, skip, "dit", "double_block", layer)
+
+            for layer in single_layers:
+                slg_model_options = comfy.model_patcher.set_model_options_patch_replace(slg_model_options, skip, "dit", "single_block", layer)
+
+            cond, uncond = conds
+            sigma_ = sigma[0].item()
+            if sigma_ >= sigma_end and sigma_ <= sigma_start and uncond is not None:
+                cond_out, _ = comfy.samplers.calc_cond_batch(model, [cond, None], x, sigma, model_options)
+                _, uncond_out = comfy.samplers.calc_cond_batch(model, [None, uncond], x, sigma, slg_model_options)
+                out = [cond_out, uncond_out]
+            else:
+                out = comfy.samplers.calc_cond_batch(model, conds, x, sigma, model_options)
+
+            return out
+
+        m = model.clone()
+        m.set_model_sampler_calc_cond_batch_function(calc_cond_batch_function)
+
+        return (m, )
 
 NODE_CLASS_MAPPINGS = {
     "SkipLayerGuidanceDiT": SkipLayerGuidanceDiT,
+    "SkipLayerGuidanceDiTSimple": SkipLayerGuidanceDiTSimple,
 }

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",
+}

comfy_extras/nodes_train.py

@@ -552,6 +552,9 @@ class LoraModelLoader:
 
 
 class SaveLoRA:
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+
     @classmethod
     def INPUT_TYPES(s):
         return {
@@ -565,7 +568,7 @@ class SaveLoRA:
                 "prefix": (
                     "STRING",
                     {
-                        "default": "trained_lora",
+                        "default": "loras/ComfyUI_trained_lora",
                         "tooltip": "The prefix to use for the saved LoRA file.",
                     },
                 ),
@@ -588,12 +591,13 @@ class SaveLoRA:
     OUTPUT_NODE = True
 
     def save(self, lora, prefix, steps=None):
-        date = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(prefix, self.output_dir)
         if steps is None:
-            output_file = f"models/loras/{prefix}_{date}_lora.safetensors"
+            output_checkpoint = f"{filename}_{counter:05}_.safetensors"
         else:
-            output_file = f"models/loras/{prefix}_{steps}_steps_{date}_lora.safetensors"
-        safetensors.torch.save_file(lora, output_file)
+            output_checkpoint = f"{filename}_{steps}_steps_{counter:05}_.safetensors"
+        output_checkpoint = os.path.join(full_output_folder, output_checkpoint)
+        safetensors.torch.save_file(lora, output_checkpoint)
         return {}
 
 

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.40"
+__version__ = "0.3.44"

execution.py

@@ -429,17 +429,20 @@ 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)
 

main.py

@@ -55,6 +55,9 @@ 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 +69,6 @@ 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 +81,9 @@ 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))
@@ -185,7 +188,13 @@ 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)
@@ -270,7 +279,10 @@ 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 = []
@@ -2305,6 +2310,7 @@ def init_builtin_api_nodes():
         "nodes_pika.py",
         "nodes_runway.py",
         "nodes_tripo.py",
+        "nodes_moonvalley.py",
         "nodes_rodin.py",
         "nodes_gemini.py",
     ]

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.40"
+version = "0.3.44"
 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==0.1.35
+comfyui-embedded-docs==0.2.4
 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