diff --git a/backend/modules/k_model.py b/backend/modules/k_model.py index f8d57130..e795b94d 100644 --- a/backend/modules/k_model.py +++ b/backend/modules/k_model.py @@ -12,11 +12,11 @@ class KModel(torch.nn.Module): self.computation_dtype = computation_dtype self.diffusion_model = huggingface_components['unet'] - self.prediction = k_prediction_from_diffusers_scheduler(huggingface_components['scheduler']) + self.predictor = k_prediction_from_diffusers_scheduler(huggingface_components['scheduler']) def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs): sigma = t - xc = self.prediction.calculate_input(sigma, x) + xc = self.predictor.calculate_input(sigma, x) if c_concat is not None: xc = torch.cat([xc] + [c_concat], dim=1) @@ -24,7 +24,7 @@ class KModel(torch.nn.Module): dtype = self.computation_dtype xc = xc.to(dtype) - t = self.prediction.timestep(t).float() + t = self.predictor.timestep(t).float() context = context.to(dtype) extra_conds = {} for o in kwargs: @@ -35,7 +35,7 @@ class KModel(torch.nn.Module): extra_conds[o] = extra model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float() - return self.prediction.calculate_denoised(sigma, model_output, x) + return self.predictor.calculate_denoised(sigma, model_output, x) def memory_required(self, input_shape): area = input_shape[0] * input_shape[2] * input_shape[3] diff --git a/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py b/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py index 3e07d1f3..0ff78fa1 100644 --- a/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py +++ b/extensions-builtin/forge_preprocessor_inpaint/scripts/preprocessor_inpaint.py @@ -154,7 +154,7 @@ class PreprocessorInpaintLama(PreprocessorInpaintOnly): def process_before_every_sampling(self, process, cond, mask, *args, **kwargs): cond, mask = super().process_before_every_sampling(process, cond, mask, *args, **kwargs) - sigma_max = process.sd_model.forge_objects.unet.model.prediction.sigma_max + sigma_max = process.sd_model.forge_objects.unet.model.predictor.sigma_max original_noise = kwargs['noise'] process.modified_noise = original_noise + self.latent.to(original_noise) / sigma_max.to(original_noise) return cond, mask diff --git a/extensions-builtin/forge_preprocessor_reference/scripts/forge_reference.py b/extensions-builtin/forge_preprocessor_reference/scripts/forge_reference.py index 51f48012..cc06ffba 100644 --- a/extensions-builtin/forge_preprocessor_reference/scripts/forge_reference.py +++ b/extensions-builtin/forge_preprocessor_reference/scripts/forge_reference.py @@ -67,8 +67,8 @@ class PreprocessorReference(Preprocessor): gen_cpu = torch.Generator().manual_seed(gen_seed) unet = process.sd_model.forge_objects.unet.clone() - sigma_max = unet.model.prediction.percent_to_sigma(start_percent) - sigma_min = unet.model.prediction.percent_to_sigma(end_percent) + sigma_max = unet.model.predictor.percent_to_sigma(start_percent) + sigma_min = unet.model.predictor.percent_to_sigma(end_percent) self.recorded_attn1 = {} self.recorded_h = {} diff --git a/extensions-builtin/forge_preprocessor_tile/scripts/preprocessor_tile.py b/extensions-builtin/forge_preprocessor_tile/scripts/preprocessor_tile.py index 40e4c39a..9ed7bb31 100644 --- a/extensions-builtin/forge_preprocessor_tile/scripts/preprocessor_tile.py +++ b/extensions-builtin/forge_preprocessor_tile/scripts/preprocessor_tile.py @@ -43,7 +43,7 @@ class PreprocessorTileColorFix(PreprocessorTile): latent = self.register_latent(process, cond) unet = process.sd_model.forge_objects.unet.clone() - sigma_data = process.sd_model.forge_objects.unet.model.prediction.sigma_data + sigma_data = process.sd_model.forge_objects.unet.model.predictor.sigma_data if getattr(process, 'is_hr_pass', False): k = int(self.variation * 2) diff --git a/extensions-builtin/sd_forge_dynamic_thresholding/lib_dynamic_thresholding/dynthres.py b/extensions-builtin/sd_forge_dynamic_thresholding/lib_dynamic_thresholding/dynthres.py index fa286cc0..e02aab73 100644 --- a/extensions-builtin/sd_forge_dynamic_thresholding/lib_dynamic_thresholding/dynthres.py +++ b/extensions-builtin/sd_forge_dynamic_thresholding/lib_dynamic_thresholding/dynthres.py @@ -38,7 +38,7 @@ class DynamicThresholdingNode: cond = input - args["cond"] uncond = input - args["uncond"] cond_scale = args["cond_scale"] - time_step = model.model.prediction.timestep(args["sigma"]) + time_step = model.model.predictor.timestep(args["sigma"]) time_step = time_step[0].item() dynamic_thresh.step = 999 - time_step diff --git a/extensions-builtin/sd_forge_fooocus_inpaint/scripts/forge_fooocus_inpaint.py b/extensions-builtin/sd_forge_fooocus_inpaint/scripts/forge_fooocus_inpaint.py index 88066585..c62b2240 100644 --- a/extensions-builtin/sd_forge_fooocus_inpaint/scripts/forge_fooocus_inpaint.py +++ b/extensions-builtin/sd_forge_fooocus_inpaint/scripts/forge_fooocus_inpaint.py @@ -102,8 +102,8 @@ class FooocusInpaintPatcher(ControlModelPatcher): if not_patched_count > 0: print(f"[Fooocus Patch Loader] Failed to load {not_patched_count} keys") - sigma_start = unet.model.prediction.percent_to_sigma(self.start_percent) - sigma_end = unet.model.prediction.percent_to_sigma(self.end_percent) + sigma_start = unet.model.predictor.percent_to_sigma(self.start_percent) + sigma_end = unet.model.predictor.percent_to_sigma(self.end_percent) def conditioning_modifier(model, x, timestep, uncond, cond, cond_scale, model_options, seed): if timestep > sigma_start or timestep < sigma_end: diff --git a/extensions-builtin/sd_forge_ipadapter/lib_ipadapter/IPAdapterPlus.py b/extensions-builtin/sd_forge_ipadapter/lib_ipadapter/IPAdapterPlus.py index 3fe54d3a..dcced0b2 100644 --- a/extensions-builtin/sd_forge_ipadapter/lib_ipadapter/IPAdapterPlus.py +++ b/extensions-builtin/sd_forge_ipadapter/lib_ipadapter/IPAdapterPlus.py @@ -760,8 +760,8 @@ class IPAdapterApply: if attn_mask is not None: attn_mask = attn_mask.to(self.device) - sigma_start = model.model.prediction.percent_to_sigma(start_at) - sigma_end = model.model.prediction.percent_to_sigma(end_at) + sigma_start = model.model.predictor.percent_to_sigma(start_at) + sigma_end = model.model.predictor.percent_to_sigma(end_at) patch_kwargs = { "number": 0, diff --git a/extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py b/extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py index a5cb522f..3403340c 100644 --- a/extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py +++ b/extensions-builtin/sd_forge_latent_modifier/lib_latent_modifier/sampler_mega_modifier.py @@ -919,10 +919,10 @@ class ModelSamplerLatentMegaModifier: cond = args["cond"] uncond = args["uncond"] cond_scale = args["cond_scale"] - timestep = model.model.prediction.timestep(args["timestep"]) + timestep = model.model.predictor.timestep(args["timestep"]) sigma = args["sigma"] sigma = sigma.view(sigma.shape[:1] + (1,) * (cond.ndim - 1)) - #print(model.model.prediction.timestep(timestep)) + #print(model.model.predictor.timestep(timestep)) x = x_input / (sigma * sigma + 1.0) cond = ((x - (x_input - cond)) * (sigma ** 2 + 1.0) ** 0.5) / (sigma) diff --git a/ldm_patched/modules/controlnet.py b/ldm_patched/modules/controlnet.py index 46848f8c..df88f49d 100644 --- a/ldm_patched/modules/controlnet.py +++ b/ldm_patched/modules/controlnet.py @@ -285,7 +285,7 @@ class ControlNet(ControlBase): def pre_run(self, model, percent_to_timestep_function): super().pre_run(model, percent_to_timestep_function) - self.model_sampling_current = model.prediction + self.model_sampling_current = model.predictor def cleanup(self): self.model_sampling_current = None diff --git a/modules_forge/forge_sampler.py b/modules_forge/forge_sampler.py index fe3bd46a..124db5b3 100644 --- a/modules_forge/forge_sampler.py +++ b/modules_forge/forge_sampler.py @@ -108,7 +108,7 @@ def sampling_prepare(unet, x): real_model = unet.model - percent_to_timestep_function = lambda p: real_model.prediction.percent_to_sigma(p) + percent_to_timestep_function = lambda p: real_model.predictor.percent_to_sigma(p) for cnet in unet.list_controlnets(): cnet.pre_run(real_model, percent_to_timestep_function)