MiniMax H3 Prompt Rewriter LoRA 8B: Image-Conditioned Rewriting

Lightx2v released an 8B variant of the MiniMax H3 prompt-rewriter LoRA (Hugging Face), and community member Nynxz converted it for ComfyUI as ComfyUI-NynxzH3 v0.0.2 (GitHub, Comfy Registry). Where the earlier 27B release was text-only, the 8B adapter accepts visual references and covers all four prompt-rewriting tasks used to train it.

What changed vs the 27B LoRA

The original MiniMax-H3-Prompt-Rewriter-LoRA builds on Qwen3.6-27B and rewrites short prompts into structured H3 audio-video descriptions. The new 8B variant swaps the base to Qwen3-VL-8B-Instruct and adds keyframe-conditioned rewriting:

• T2VA / T2AV: text to audio-video
• I2VA / I2AV: first-frame image + text to audio-video
• L2VA / L2AV: last-frame image + text to audio-video
• FL2VA / FL2AV / FLF2V: first- and last-frame images + text to audio-video

The rewritten prompt keeps the same structured output — shot timeline, synchronized physical and ambient sound, non-diegetic music guidance — that the 27B version produces. Ref2VA conditioning is not supported by this adapter release.

ComfyUI-NynxzH3 v0.0.2

Nynxz converted the adapter to ComfyUI and shipped the result as ComfyUI-NynxzH3 (GitHub). The v0.0.2 registry release bundles the prompt templates the LoRA was trained on, plus the converted weights:

• LoRA: minimax-h3-prompt-rewriter-8b_lora_rank256_bf16.safetensors (~1.3 GB, rank 256)
• Text encoder: qwen3vl_8b_int8_convrot.safetensors (~8.8 GB) — the Qwen3-VL-8B encoder for ComfyUI
• Optional: an abliterated Qwen3-VL-8B extract converted to a tiny (<4 MB) LoRA, also in the loras folder

Nynxz's own demo shows the pack wired into a ComfyUI graph with the prompt-rewriter nodes:

The ComfyUI-NynxzH3 graph with the 8B prompt-rewriter nodes wired into H3 conditioning.

The 8B route is a practical drop-in for H3 workflows that suffer from literal prompt reading, without the large VRAM footprint of the 27B base model.

Availability

Install the node pack from ComfyUI-Manager (search NynxzH3) or clone ComfyUI-NynxzH3 into ComfyUI/custom_nodes/, then place the LoRA and text encoder from nynxz/Qwen3-VL-8B-ComfyUI into the corresponding models/loras/ and models/text_encoders/ folders. The rewriter is a text-side adapter: it rewrites prompts only and still requires the MiniMax-H3 generator weights plus a ComfyUI MiniMax-H3 workflow to produce video.