MiniMax H3 Prompt Rewriter 8B / 多模态提示词改写 (Advanced)

MiniMax H3's own text encoder is a 32B Qwen3-VL, and you can't LoRA a good rewrite habit into it - its parameter shape won't take the community's prompt-rewriting LoRA. MiniMaxH3PromptRewriter8BT8Advanced is the pack's answer: it runs a separate, dedicated Qwen3-VL-8B model with the pinned LightX2V MiniMax-H3-Prompt-Rewriter-LoRA-8B adapter, produces the same three-field structured rewrite the pack uses everywhere, and then unloads - by default it frees the model and cache after every generation, error, or OOM. A 16GB card can actually run this, though the README's verdict on CPU offloading is honest: it works and produces structured results, but it's slow.

Why a separate model? Because "their parameter shapes are not LoRA-compatible" - the description says it in so many words. The 32B CLIP is doing encoding duty for generation; you don't want a rewriter LoRA grafted onto it. The 8B is a sidecar: load, rewrite, unload.

The inputs that matter

• prompt - the raw short prompt. Dialogue, lyrics, and on-screen text are required to stay verbatim (the tooltip says so), which is how a rewrite doesn't quietly change your character's lines.
• task (T2VA default), resolution, duration - the same contract shape as the provider router.
• base_model_path - Qwen3-VL-8B-Instruct by default, resolved relative to ComfyUI/models/text_encoders (absolute paths also allowed). The precise base is Qwen/Qwen3-VL-8B-Instruct.
• adapter_path - MiniMax-H3-Prompt-Rewriter-LoRA-8B, resolved from models/loras. Exact upstream: lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA-8B.
• load_policy - auto_cpu_offload (recommended for 16GB), gpu_only if you have headroom.
• dtype (bfloat16), decoding (greedy), seed (42), temperature (0.7), top_p (0.8).
• max_new_tokens - 1024 default. Too low and the three-field structure gets truncated; too high just raises the ceiling, it doesn't force the model to fill it.
• min_image_pixels / max_image_pixels - the VLM's image resolution window for the optional first_frame/last_frame inputs.
• unload_after_generate (true) and free_comfy_models_before_load (true) - the two VRAM-hygiene switches; the latter frees current ComfyUI models before the 8B loads so a 16GB card has room, without a global unload afterwards.
• allow_hub_download - false by default, and that default is the feature: the node refuses surprise downloads of multi-GB files at runtime. If you haven't pre-downloaded the base and adapter, it errors locally instead of silently pulling them.

Outputs: enhanced_prompt, integrated_multimodal_description, overall_soundscape, non_diegetic_music, report_json.

The install that's actually different

This is the one node in the pack with real extra dependencies. From the README, only install them if you're using the rewriter, and install into ComfyUI's Python:

cd ComfyUI/custom_nodes/minimax-h3-audio-T8
pip install -r requirements-prompt-rewriter.txt

The file pins a conservative stack - transformers>=4.57.1,<5, accelerate, peft>=0.18,<0.21, safetensors, pillow>=10,<11 - deliberately constrained so installing this doesn't yank versions other nodes depend on. You also need the Qwen3-VL-8B base and the LightX2V LoRA on disk (models/text_encoders and models/loras respectively), because allow_hub_download won't fetch them for you. The pack itself otherwise installs plain: Manager search "MiniMax H3 Audio T8", or clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8, restart. Template: examples/workflows/14-prompt-relay/2026-08-22_H3_Prompt_Rewriter_8B_Advanced_EXP.json. It's a genuinely useful sidecar, but it's the slow, model-hauling one - budget for it.