Merge fb5c22292e6302678f140f67e25f05d83dc33902 into dbc63f622dd095391335612d0c7d7bbff8745cc8

install-portable.bat - installs the requirements.txt for ComfyUI portable users (embedded python)

setup-onediff.bat - installs the recommended extras for ComfyUI Portable users (embedded python)

It can be easy to install to system python by accident when running the the Portable version of ComfyUI, so i use these to help make life easier.

install-portable.bat

@@ -0,0 +1,16 @@
+@echo off
+
+set "requirements_txt=%~dp0\requirements.txt"
+set "python_exec=..\..\..\python_embeded\python.exe"
+
+echo Installing node...
+
+if exist "%python_exec%" (
+    echo Installing with ComfyUI Portable
+    "%python_exec%" -s -m pip install -r "%requirements_txt%"
+) else (
+    echo Installing with system Python
+    pip install -r "%requirements_txt%"
+)
+
+pause

pipeline_cogvideox.py

@@ -658,10 +658,9 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                     latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                     counter = torch.zeros_like(latent_model_input)
                     noise_pred = torch.zeros_like(latent_model_input)
-                    
+
                     if image_cond_latents is not None:
                         latent_image_input = torch.cat([image_cond_latents] * 2) if do_classifier_free_guidance else image_cond_latents
-                        latent_model_input = torch.cat([latent_model_input, latent_image_input], dim=2)
 
                     # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                     timestep = t.expand(latent_model_input.shape[0])
@@ -724,7 +723,14 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
                             noise_pred = noise_pred.float()
                     else:
                         for c in context_queue:
+                            print("c:", c)
+
                             partial_latent_model_input = latent_model_input[:, c, :, :, :]
+                            if image_cond_latents is not None:
+                                partial_latent_image_input = latent_image_input[:, :len(c), :, :, :]
+                                partial_latent_model_input = torch.cat([partial_latent_model_input,partial_latent_image_input], dim=2)
+                            
+                            print(partial_latent_model_input.shape)
                             if (tora is not None and tora["start_percent"] <= current_step_percentage <= tora["end_percent"]):
                                 if do_classifier_free_guidance:
                                     partial_video_flow_features = tora["video_flow_features"][:, c, :, :, :].repeat(1, 2, 1, 1, 1).contiguous()
@@ -768,7 +774,13 @@ class CogVideoXPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
 
                     if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                         progress_bar.update()
-                        comfy_pbar.update(1)
+                        if callback is not None: 
+                            alpha_prod_t = self.scheduler.alphas_cumprod[t]
+                            beta_prod_t = 1 - alpha_prod_t
+                            callback_tensor = (alpha_prod_t**0.5) * latent_model_input[0][:, :16, :, :] - (beta_prod_t**0.5) * noise_pred.detach()[0]
+                            callback(i, callback_tensor * 5, None, num_inference_steps)
+                        else:
+                            comfy_pbar.update(1)
 
                 # region sampling
                 else:

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "comfyui-cogvideoxwrapper"
-description = "Diffusers wrapper for CogVideoX -models: [a/https://github.com/THUDM/CogVideo](https://github.com/THUDM/CogVideo)"
+description = "Diffusers wrapper for CogVideoX -models: https://github.com/THUDM/CogVideo"
 version = "1.5.1"
 license = {file = "LICENSE"}
 dependencies = ["huggingface_hub", "diffusers>=0.31.0", "accelerate>=0.33.0"]

readme.md

@@ -5,7 +5,7 @@ Spreadsheet (WIP) of supported models and their supported features: https://docs
 ## Update 9
 Added preliminary support for [Go-with-the-Flow](https://github.com/VGenAI-Netflix-Eyeline-Research/Go-with-the-Flow)
 
-This uses LoRA weights available here: https://huggingface.co/VGenAI-Netflix-Eyeline-Research/Go-with-the-Flow/tree/main
+This uses LoRA weights available here: https://huggingface.co/Eyeline-Research/Go-with-the-Flow/tree/main
 
 To create the input videos for the NoiseWarp process, I've added a node to KJNodes that works alongside my SplineEditor, and either [comfyui-inpaint-nodes](https://github.com/Acly/comfyui-inpaint-nodes) or just cv2 inpainting to create the cut and drag input videos.
 

setup-onediff.bat

@@ -0,0 +1,15 @@
+@echo off
+
+set "python_exec=..\..\..\python_embeded\python.exe"
+
+echo Installing node...
+
+if exist "%python_exec%" (
+    echo Installing with ComfyUI Portable
+    "%python_exec%" -s -m pip install --pre onediff onediffx && pip install nexfort"
+) else (
+    echo Installing with system Python
+    pip install --pre onediff onediffx && pip install nexfort"
+)
+
+pause