MiniMax-H3 Prompt Writer (Ref2VA)

The ordinary writer turns text into a video description. Ref2VA is the full-reference mode from MiniMax's own guide: the target video doesn't just depict a scene, it reuses assets - a character's look, a specific style, a voice timbre, a clip being edited. MiniMax's full-reference guide knows how to express that, and this node turns your prompt plus a list of labeled references into the six-section description H3 wants.

This is the sibling of the MiniMax-H3 Prompt Writer (T2VA/I2VA/FL2VA/L2VA) - same no-LoRA approach, any instruction-following GGUF, the guide fetched from MiniMax on first use. The difference is which guide goes into the system prompt (the longer, ~24 KB full-reference one) and what comes out: six named outputs instead of three.

The one input that's the whole point

reference_assets is required, and the node refuses to run when it's empty - full-reference mode describes how a target video reuses assets, so with no assets it's just T2VA under another name. One asset per line, labeled, and the model does the rest:

Picture 1: young woman, long dark hair, blue cardigan, thin silver necklace
Picture 2: corner cafe interior, brick wall, brass lamps, rain on the window
Audio 1: voice-timbre reference for the woman - low, unhurried, slight rasp

The model binds <Subject 1> to the woman, <Picture 1> to the drawing's style, picks the [task type] prefix itself, and writes the retention analysis - fully_preserved, partially_preserved, attribute_transfer, weak_reference, fully_copy, that vocabulary. Or skip the typing: run a MiniMax-H3 Reference Caption (or the Multi Reference Caption) upstream and paste its reference_assets output straight into this field.

The outputs

• subject_definitions - what each <Subject N> / <Picture N> / <Video N> / <Audio N> label denotes.
• summary - the [task type] prefix and the reference relationships in one paragraph.
• retention_analysis - per label: how faithfully each asset is carried over.
• detailed_description - the body, shot by shot, labels cited where their roles apply.
• overall_soundscape - ambience and physical sound.
• non_diegetic_music - the audience-only score.

Plus rewritten_prompt, the whole thing in one string. Every output is a plain STRING - wire any of them into a text viewer.

Inputs to know

• prompt - what the target video should show and how it uses the references. The descriptions in reference_assets say what the assets are; this says what they're for.
• model - the guide is the longer of the two, so a 4B will hold the format, but a 9B keeps the labels consistent across all six sections. That consistency is where small models wobble.
• resolution / duration - keep equal to what you pass to H3. Duration caps at 15 s here.
• greedy - on. Small models drift out of format when they sample.

Install

Same pack, same two commands as every node in it:

cd ComfyUI/custom_nodes
git clone https://github.com/pytraveler/MiniMax-H3-Prompt-Rewriter-ComfyUI

(Or ComfyUI Manager → search "MiniMax-H3".) The writer runs the official llama.cpp binaries, so no Python deps are strictly needed for this node - first run fetches the guide and your model.

Where people get burned

• Empty reference_assets gets refused, not silently downgraded to T2VA. That's deliberate and it's the right call - but it means wiring a captioner upstream is the normal setup, not an optional extra.
• Labels stay within their own category. Picture 1, Picture 2, Video 1, Audio 1 - not 1 through 4. The caption nodes already number this way, so pasting their output just works.
• The rewrite may invent retention levels. It's following instructions, not reproducing a trained distribution. Skim the retention_analysis before generating - if a character's look must survive unchanged, that's the line to eyeball.