Kevin-thu /StoryMem

About StoryMem

StoryMem makes a minute-long story video, one shot at a time. It builds on Wan2.2 and keeps characters looking the same by saving keyframes as memory. You need a big GPU and large model downloads to run it.

Best for: Researchers and advanced creators who want coherent multi-shot story videos on their own machine.

• Generates long story videos shot by shot with a memory model
• Uses a Memory-to-Video LoRA (a small add-on model) on Wan2.2
• Keeps characters consistent by saving keyframes after each shot
• Ships ST-Bench: 30 story scripts with 300 shot prompts
• Lets you set resolution, seed and memory size from the command line

1. 1
   Clone the repo and make a Python 3.11 environment: the command below, the next command, conda activate storymem

   git clone --single-branch --branch main git@github.com:Kevin-thu/StoryMem.git

   Copy

   conda create -n storymem python=3.11

   Copy
2. 2
   Install the dependencies: the command below then pip install flash_attn

   pip install -r requirements.txt

   Copy
3. 3
   Download the models: the command below, the next command, the next command

   huggingface-cli download Wan-AI/Wan2.2-T2V-A14B --local-dir ./models/Wan2.2-T2V-A14B

   Copy

   huggingface-cli download Wan-AI/Wan2.2-I2V-A14B --local-dir ./models/Wan2.2-I2V-A14B

   Copy

   huggingface-cli download Kevin-thu/StoryMem --local-dir ./models/StoryMem

   Copy
4. 4
   Run the example: bash run_example.sh
5. 5
   Change story_script_path, output_dir, max_memory_size, size and seed to fit your story