Local AI Video Generator

First/last-frame animation and synchronized audio with ComfyUI

English | 简体中文

Windows 11 ComfyUI 0.30+ RTX 5090 verified MIT License

Run advanced image-to-video workflows locally on Windows. This repository installs and manages a dedicated ComfyUI environment for two model families:

• Wan2.2 Fun InP A14B — first-frame and last-frame video generation with quality and LightX2V fast modes.
• MiniMax H3 FL2VA — first-frame and last-frame video generation with synchronized stereo audio in the same inference pass.

• Animate the transition between a chosen start image and end image.
• Generate synchronized stereo audio with MiniMax H3 FL2VA.
• Switch between Wan2.2 quality and LightX2V fast modes.
• Install ComfyUI, Python dependencies, and approximately 75 GiB of models with setup.bat.
• Resume interrupted Hugging Face downloads and skip already verified files.
• Validate CUDA, GPU support, ComfyUI, workflows, and every model file.
• Keep model weights, prompts, input media, outputs, caches, and logs off GitHub.

The MiniMax workflow extends the official ComfyUI I2V template by connecting the last-frame input. It keeps the official sampling graph and requires no custom nodes.

The following configuration is the only one verified end to end:

Before installation, make sure you also have:

• Conda installed and available through conda --version in PowerShell.
• Access to GitHub, the PyTorch package index, and Hugging Face.
• A stable connection for approximately 75 GiB of model downloads.
• At least 100 GiB of free disk space; allow more if download caches are retained.
• Enough system memory for ComfyUI to offload model weights.

The scripts do not detect the GPU model or automatically reconfigure PyTorch for RTX 4090, RTX 3090, workstation, or lower-VRAM cards.

Open PowerShell and run:

git clone https://github.com/carrotProgrammer/local-ai-video-generator.git
cd local-ai-video-generator
.\setup.bat

You can also download the repository ZIP and double-click setup.bat.

The setup script will:

1. Create a project-local Conda environment in .venv with Python 3.12 and Git.
2. Clone the official ComfyUI repository into ComfyUI.
3. Install PyTorch 2.12.1, TorchVision 0.27.1, the CUDA 13.0 build, and ComfyUI dependencies.
4. Download all Wan2.2 and MiniMax H3 models, approximately 75 GiB in total.
5. Verify CUDA, GPU support, ComfyUI, all three workflows, and every model file size.

Successful model downloads are reused on later runs, while incomplete downloads are resumed when possible.

Start ComfyUI after setup and verification finish:

.\start_comfyui.bat

Open http://127.0.0.1:8188 if the browser does not open automatically.

1. Drag workflows/MiniMax_H3_FL2VA_官方.json onto the ComfyUI canvas.

2. Select your start and end images in the two Load Image nodes. Disconnect the end image for regular single-image I2V.

3. Use images with a similar aspect ratio. Begin with 0.4 MP, then increase to about 0.98 MP after validating the motion; 16:9 is approximately ×ばつ768.

4. Describe the visuals, motion, camera, and sound in the prompt. For example:

   The camera slowly pushes in as the subject moves naturally from the opening pose to the final pose, preserving identity and clothing. Audio: quiet indoor ambience, soft footsteps, no music.

5. Keep the default duration at 5 seconds for the first run. H3 generates at 24 fps and aligns the frame count to the required 17k+5 grid.

6. Press Ctrl+Enter. The MP4 with audio is saved under output/video.

• Start with 0.4 MP / 5 seconds on the RTX 5090, then increase the resolution.
• Do not add --highvram; ComfyUI needs to offload some H3 weights to system memory.
• Close other GPU-heavy applications and do not run Wan2.2 and H3 jobs simultaneously.
• Restart ComfyUI after an out-of-memory error to release fragmented VRAM.
• Keep the native H3 canvas short edge at or below 768 and dimensions divisible by 32.

Choose one of the included workflows:

• Fast check: workflows/Wan2.2_Fun_InP_快速测试_512_33帧.json
• Quality and fast modes: workflows/Wan2.2_Fun_InP_官方_质量与快速.json

Drag the workflow onto the canvas, choose the start and end images, enter a positive prompt, and press Ctrl+Enter. Wan frame counts follow 4n+1, such as 33, 49, or 81 frames. The workflows save video at 16 fps.

If Wan2.2 is already installed, download and verify only the H3 models, approximately 39.56 GiB:

.\download_h3_models.bat

Non-interactive equivalents:

.\.venv\python.exe .\scripts\download_models.py h3
.\.venv\python.exe .\scripts\verify_install.py h3

The workflow uses the official quantized components from Comfy-Org/MiniMax-H3:

# Start the installed ComfyUI service
.\start_comfyui.bat
# Stop the service
.\stop_comfyui.bat
# Validate the complete installation
.\.venv\python.exe .\scripts\verify_install.py all

Inputs belong in input; generated media is written to output. The launcher reserves 2 GiB of VRAM and uses comfyui.pid to prevent duplicate instances.

The public repository contains only setup and service scripts, the model manifest, and reusable workflow templates. .gitignore excludes:

• The cloned ComfyUI source and project environment.
• Model weights and interrupted downloads.
• Input images, videos, generated output, and logs.
• Hugging Face caches, PID files, environment files, and local experiment data.

Workflow templates use generic start_image.png and end_image.png placeholders. Review newly saved workflows before committing them if they contain private prompts or local paths.

• MISSING/INCOMPLETE — rerun the relevant download command; the script will attempt to resume.
• MiniMaxH3ImageToVideo is missing — update ComfyUI to version 0.30.0 or later.
• H3 model selectors are empty — confirm the four model files are in diffusion_models, text_encoders, and vae, then restart ComfyUI.
• A PID file exists but the service is stopped — the launcher removes stale PID files automatically; stop_comfyui.bat can also clean it up.
• Out of memory — lower resolution or duration, close other GPU applications, and restart ComfyUI.

MIT

This is an independent community project and is not affiliated with ComfyUI, Alibaba, MiniMax, NVIDIA, or Hugging Face. Follow the licenses and acceptable-use terms of the models and services you use.