MiniMax-H3 Prompt-Rewriter (Qwen3.8-27B)

Prompt-rewriter models built on Qwen3.8-27B:

Using in ComfyUI

Option 1 — Native "Generate Text" node (recommended)

Both FL2VA and Ref2VA rewriter loras can be used directly with ComfyUI's native Generate Text (TextGenerateV2) node — no extra custom node packs are needed.

The comfy/ directory of this repo contains everything for this setup:

Loras (converted for the native node's LoadLoraTextEncoder loader):
- lightx2v_minimax_h3_FL2VA_PR-Q3.8-27B_rank32_bf16_comfy.safetensors
- lightx2v_minimax_h3_Ref2VA_PR-Q3.8-27B_rank32_bf16_comfy.safetensors
Example workflow: comfy/H3_FL2VA_prompt_gen_27B_9B.json

The workflow loads the Qwen3.8-27B text encoder with CLIPLoader (e.g. qwen3.8_27b_w4a8.safetensors), attaches the lora through Load Lora (Text Encoder) (LoadLoraTextEncoder), and runs prompt generation in Generate Text (TextGenerateV2). A switch also lets you flip between the 27B and a 9B encoder variant.

The loras are converted on the fly during runtime, so just point the loader at the .safetensors file.

Option 2 — ComfyUI-QwenVL nodes

These models can also be used with the ComfyUI-QwenVL (separate branch with changes) nodes as prompt enhancers via the QwenVL (GGUF) node, or any other LLM lora loaders.

An example workflow is available here: Q3.8_27B_H3_promt_enhance.json.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for siraxe/lightx2v_MiniMax-H3-Prompt-Rewriter_loras

Base model

lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-FL2VA

Finetuned

(1)

this model