pytraveler/MiniMax-H3-Prompt-Rewriter-ComfyUI

What it solves

This project provides ComfyUI nodes that transform short, simple prompts into structured, production-ready audio-video descriptions specifically formatted for the MiniMax-H3 video generation model. It eliminates the need to manually write complex, shot-by-shot descriptions, soundscapes, and musical scores required for high-quality MiniMax-H3 output.

How it works

The system uses various Large Language Models (LLMs) and LoRA adapters to expand a user's input. It offers four primary paths:

• LoRA-based Rewriters: Uses Qwen models (27B, 8B, or Omni) with specialized adapters to generate the structured output. The 8B and Omni versions are multimodal, meaning they can "see" reference images or "hear" audio clips to inform the rewrite.
• Writer Nodes: Use any instruction-following GGUF model guided by a system prompt based on MiniMax's own writing guide, offering a lightweight alternative for lower-end hardware.
• Reference Captioning: Converts images, audio, or video into text descriptions that the writer nodes can then use.
• Prompt Presets: A library of a thousand pre-made MiniMax-H3 prompts for quick use.

Who it’s for

• ComfyUI users generating video with MiniMax-H3.
• Creators who want high-quality, structured video prompts without writing them manually.
• Users with varying hardware capabilities, from those with 8GB VRAM (using writer nodes) to those with 16GB+ VRAM (using the 27B LoRA).

Highlights

• Multimodal Input: Supports Text-to-Video (T2VA), Image-to-Video (I2VA), First-and-Last-Frame-to-Video (FL2VA), and Reference-to-Video (Ref2VA).
• Local Execution: All prompt rewriting happens entirely locally.
• Flexible Model Options: Supports multiple model sizes (27B, 8B, Omni) and quantization formats (nf4, int8, GGUF) to fit different VRAM budgets.
• Integrated Workflow: Includes seven example workflows to get users started quickly with different prompt generation strategies.