---
title: MiniMax-H3 Prompt Rewriter LoRA 8B
emoji: "🎬"
colorFrom: gray
colorTo: green
sdk: gradio
sdk_version: 6.24.0
app_file: app.py
short_description: Rewrite prompts into MiniMax-H3 audio-video prompts
python_version: "3.12"
startup_duration_timeout: 1h
---

# MiniMax-H3 Prompt Rewriter LoRA 8B

This Space uses a [PEFT LoRA adapter](https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA-8B) fine-tuned on [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct) to rewrite short user prompts into production-ready, structured prompts for the **MiniMax-H3** joint audio-video generation model.

## Supported Tasks

| Task | Description | Inputs |
|------|-------------|--------|
| **T2VA** | Text to audio-video | Text prompt only |
| **I2VA** | First-frame image + text to audio-video | First frame image + text prompt |
| **L2VA** | Last-frame image + text to audio-video | Last frame image + text prompt |
| **FL2VA** | First & last-frame images + text to audio-video | Both reference images + text prompt |

## How it works

1. Enter a short prompt (e.g. "A corgi runs through a rainy neon-lit alley.")
2. Select a task type — T2VA for text-only, or I2VA/L2VA/FL2VA with reference images
3. Set duration and resolution
4. Click **Rewrite** to get a structured MiniMax-H3 prompt with `integrated_multimodal_description`, `overall_soundscape`, and `non_diegetic_music` fields

The rewritten prompt is ready to feed into a downstream MiniMax-H3 video generation pipeline (e.g. via [LightX2V](https://github.com/lightx2v/lightx2v)).