MiniMax-H3 Prompt Rewriter (FL2VA)

This model rewrites prompts for MiniMax-H3 video and audio generation. It supports text-only prompts and optional first/last-frame images.

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. The writing guide is included in this directory.

Run

Check MODEL_PATH and GPU settings in vllm_server.sh, then start the server:
```
bash vllm_server.sh
```
In another terminal, edit post_vllm.sh: set PROMPT, TASK_TYPE, DURATION, and any image paths, then run:
```
bash post_vllm.sh
```

Task	Required images
`T2VA`	None
`I2VA`	`FIRST_IMAGE`
`L2VA`	`LAST_IMAGE`
`FL2VA`	`FIRST_IMAGE` and `LAST_IMAGE`

Image paths must be absolute. Keep client MODEL and SERVER_URL aligned with the server's MODEL_NAME, HOST, and PORT. Commands above assume this directory is the working directory; the scripts also accept absolute 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 the input and output; the output limit is 16384.

The client saves the enhanced prompt, reasoning, input, generation parameters, usage, and timing to infer_outputs/result.json, overwriting an existing file. Run python post_vllm.py --help for all options.