lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.5-9B-Ref2VA

MiniMax-H3 Prompt Rewriter (Ref2VA)

This model rewrites prompts for MiniMax-H3 video and audio generation. It supports Ref2VA prompts with image, video, and audio references.

This model was trained through supervised fine-tuning (SFT) of Qwen3.5-9B.

Setup

Use a Python environment with vLLM and GPU support on the server:

pip install vllm requests pillow qwen-vl-utils

All scripts use python from the active environment. For video references, install ffmpeg and ffprobe and add them to PATH. The writing guide is included here.

Run

1. Check MODEL_PATH and GPU settings in vllm_server.sh, then start the server:

   bash vllm_server.sh
   

2. In another terminal, edit post_vllm.sh: set PROMPT, DURATION, and at least one absolute path in IMAGES, VIDEOS, or AUDIOS, then run:

   bash post_vllm.sh
   

For a dry run, add --dry_run to the Python command in post_vllm.sh.

Each modality is numbered from 1 in array order, matching references such as image 1, video 1, and audio 1 in the prompt. Videos are sampled into VIDEO_FRAMES images each (default: 8); images plus sampled frames must fit both the client and server image limits (default: 50). Audio files are checked for existence and represented by labels; their contents are not read or uploaded.

Keep client MODEL and SERVER_URL aligned with the server's MODEL_NAME, HOST, and PORT. Run the commands from this directory, or use absolute script paths from elsewhere.

Both example servers use GPU 0 and port 8000. To run both models at once, assign separate GPUs and ports and update each client's SERVER_URL. The 32768-token context includes text, image/frame tokens, and output; the output limit is 16384. Reference counts alone do not guarantee that a request fits.

Outputs in infer_outputs/ are overwritten on each run: result.request.json contains the request, result.prompt.txt previews its text, and result.json contains the response, reasoning, parameters, usage, and timing. Dry runs save only the request and preview. Run python post_vllm.py --help for all options.