anveshane-Comfyui_StoryMem/workflows

ComfyUI StoryMem Test Workflows

This directory contains test workflows for verifying the StoryMem custom nodes work correctly after fixing issue #3 (import errors).

Available Workflows

1. test_workflow_simple.json - Single Shot Test

Purpose: Basic functionality test Complexity: Simple (1 shot) Runtime: ~2-5 minutes

What it tests:

• Model loading
• Memory buffer creation
• First shot generation (T2V)
• Memory visualization

Nodes used:

• StoryMemModelLoader
• StoryMemMemoryBuffer
• StoryMemFirstShot
• StoryMemMemoryVisualizer
• PreviewImage (x2)

2. test_workflow_3shot_story.json - Multi-Shot Story

Purpose: Complete workflow test with memory continuity Complexity: Advanced (3 shots) Runtime: ~10-15 minutes

What it tests:

• Model loading
• Memory buffer creation
• First shot generation (T2V)
• Continuation shot with MI2V (memory + first frame)
• Continuation shot with M2V and scene cut
• Video combination
• Memory persistence across shots

Story outline:

1. Shot 1: "A young woman walks through a bustling city street at sunset"
2. Shot 2: "She stops at a café window, looking at her reflection" (MI2V mode)
3. Shot 3: "Inside the café, warm lighting illuminates the cozy interior" (M2V mode, scene cut)

Nodes used:

• StoryMemModelLoader
• StoryMemMemoryBuffer
• StoryMemFirstShot
• StoryMemContinuationShot (x2)
• StoryMemVideoCombine
• PreviewImage (x3)

How to Use

Prerequisites

1. Install ComfyUI (if not already installed)

   git clone https://github.com/comfyanonymous/ComfyUI
   cd ComfyUI
   pip install -r requirements.txt
   

2. Install StoryMem Custom Nodes

   cd custom_nodes
   git clone https://github.com/anveshane/Comfyui_StoryMem
   cd Comfyui_StoryMem
   git submodule update --init --recursive
   pip install -r requirements.txt
   

3. Download Models

   You need to download the StoryMem models:

   • Wan2.2-T2V-A14B (Text-to-Video) - ~20GB
   • Wan2.2-I2V-A14B (Image-to-Video) - ~20GB
   • StoryMem LoRA weights (MI2V/MM2V) - ~2GB

   Place models in: ComfyUI/models/checkpoints/storymem/

Loading a Workflow

1. Start ComfyUI:

   python main.py
   

2. Open your browser to http://127.0.0.1:8188

3. Load a workflow:

   • Click "Load" button
   • Navigate to custom_nodes/Comfyui_StoryMem/workflows/
   • Select test_workflow_simple.json or test_workflow_3shot_story.json

4. Click "Queue Prompt" to run

Expected Results

Simple Workflow

• Output: Single 25-frame video (832x480)
• Memory: 3 keyframes extracted
• Preview: Video frames + memory visualization grid

3-Shot Story Workflow

• Output: Combined video with 75 frames total (3 shots × 25 frames)
• Memory: Up to 10 keyframes maintained across shots
• Preview: Individual shot previews + final combined video

Troubleshooting

Import Errors

If you see errors like:

NameError: name 'Dict' is not defined

or

cannot import name 'WanMI2V' from 'wan.textimage2video'

Solution: Make sure you're using the fixed version from PR that addresses issue #3.

Model Not Found

FileNotFoundError: Model checkpoint not found

Solution:

• Verify models are downloaded and placed in correct directory
• Check model paths in StoryMemModelLoader node
• Use StoryMemInfo node to see expected model locations

CUDA Out of Memory

RuntimeError: CUDA out of memory

Solution:

• Enable "offload_model" in StoryMemModelLoader
• Reduce frame count from 25 to 17 or 13
• Reduce resolution from 832x480 to 704x384
• Close other GPU applications

Slow Generation

Solutions:

• Reduce sampling steps (from 50 to 30)
• Enable model offloading (trades VRAM for speed)
• Use smaller frame counts
• Consider using a machine with more powerful GPU

Testing the Fix for Issue #3

These workflows specifically test the fixes made in issue #3:

Test 1: Import Verification

The workflows will fail to load if the imports are broken. Successfully loading the workflow confirms:

• ✅ Dict and Any are properly imported in keyframe_extractor.py
• ✅ WanTI2V (not WanMI2V) is correctly imported in storymem_bridge.py

Test 2: Type Annotations

The analyze_frame_diversity method in keyframe_extractor.py uses Dict[str, Any]. If this executes without NameError, the fix is working.

Test 3: MI2V Pipeline

The 3-shot workflow uses MI2V mode which requires the WanTI2V class. Successfully running shot 2 confirms the class name fix.

Contributing

Found issues with these workflows? Please report them:

• GitHub: https://github.com/anveshane/Comfyui_StoryMem/issues
• Include: ComfyUI version, GPU model, error messages, workflow used

Version History

• v1.0.0 (2024-01-XX): Initial test workflows for issue #3 fix
  • Added simple single-shot test
  • Added 3-shot story test with memory continuity
  • Verified import fixes work correctly