Fully local MiniMax H3 prompt enhancement: a ComfyUI workflow guide

Patient_Ratio4177 · reddit · 2026-08-19

The author shares a fully local workflow that expands shorthand prompts plus reference images into the six-section format MiniMax's official H3 video prompt guide expects (structured for 14-second videos, adjustable).

Setup: install the ComfyUI-H3-Qwen3VL-TextGen node pack to enable text generation with Qwen 3 VL 32B inside ComfyUI, then download the model's missing generation-tail weights from Hugging Face into models/textencoders/H3/generationtails — a standard ComfyUI Qwen 32B setup lacks the tail actually needed for text generation.

Why no official node: MiniMax offers a paid cloud prompt-expansion service unlikely to be released locally; API key holders can use the built-in MinimaxHailuo03ContextIRNode.

Design philosophy: deliberately modular, minimal custom nodes (each pack is a liability for your Python env, security, and startup time), fully local with no external dependencies.

Trade-offs: slow — on an RTX 5090 a 14s 0.4MP video takes 7 minutes including expansion; the system prompt is rigid but rewritable.

Original post →

More from Multimodal

Multimodal channel →