H3 Prompt Expander

ComfyUI-H3-PromptExpander

One node, MiniMax H3 Prompt Expander (local GGUF), that turns a short request plus optional images or reference videos into a MiniMax-H3 structured prompt. It runs a local GGUF Qwen vision model through llama-cpp-python and feeds it the official h3-prompt-writing guide from MiniMax-AI/MiniMax-H3 (references/base-en.txt, references/ref-en.txt). The Output Rules and Tips for Better Results sections of references/SKILL.md are spliced into the system prompt too, so extra house rules go there. Nothing leaves the machine.

Models

Put the language GGUF and its mmproj-*.gguf vision projector in models/LLM/. Tested with Qwen_Qwen3.5-9B-Q8_0.gguf + mmproj-Qwen_Qwen3.5-9B-bf16.gguf and with Qwen3.8-27B-Heretic-ara-noMTP-Q4_K_M.gguf + mmproj-Qwen3.8-27B-Heretic-ara-F16.gguf; Qwen3.8-27B GGUFs use the same qwen3.5 / qwen3.8 chat family. Qwen3-VL GGUFs use qwen3-vl.

Qwen3.8 GGUFs carry an MTP (nextn) layer that llama-cpp-python 0.3.35 cannot load (missing tensor 'blk.64.ssm_conv1d.weight'). llama-cpp-python 0.4.3 (JamePeng, 2026-10-17) loads them; on an older build use a noMTP file instead. MTP only speeds up decoding and does not change the output.

Mode selection

| connected inputs | mode | guide | |---|---|---| | none | T2VA | base-en | | first_frame | I2VA | base-en | | first_frame + last_frame | FL2VA | base-en | | last_frame | L2VA | base-en | | any ref_image / ref_video | Ref2VA (six sections) | ref-en |

Reference videos are IMAGE frame batches, the same convention as the core MiniMax H3 Reference to Video node; frames_per_video frames are sampled evenly and shown with their timestamps.

Outputs

prompt is the complete text for the H3 text encoder. description, soundscape and music are the three fields split out for hand edits.

tools/strip_nextn.py in.gguf out.gguf writes that noMTP copy from a GGUF that has the layer.