MiniMax H3 Enhancer Visual Reference (Legacy)

Before this pack's Plan v2 workflow existed, the enhancer pipeline worked differently: you built a prompt with the old Prompt Guide, fed reference media into Qwen's writing context through a chain of Visual Reference nodes, and a separate Prompt Enhancer node rewrote the prose. Enhancer Visual Reference (Legacy) is one of those chain nodes. It's still registered so old workflows load, but the pack's own docs are unambiguous: new projects should use Plan v2, where Image Reference and Video Reference do this job inside the typed plan.

Since it's a legacy node, let me be direct about when you should care: if you have an existing workflow using the Prompt Guide + Prompt Enhancer path, this node is what keeps it running. If you're starting fresh, use Plan v2 and don't build on this. The migration doc in the pack covers the move.

What it does

It adds one picture or reference-video frame batch to Qwen's prompt-writing context while passing native-ready H3 media through. Chain Visual Reference nodes to order media; the reference_context output feeds both the Prompt Guide's and Prompt Enhancer's reference_context inputs so their labels, roles, grouping, retention, and mode stay identical.

Inputs that matter

• media - exactly one image for "Picture," or the full video frame batch for "Video frames."
• media_type - Picture or Video frames. Picture can be an exact first/last endpoint or a Ref2VA picture; Video frames is always Ref2VA.
• reference_role - the single-role compatibility setting, with 14 choices (exact first frame, exact last frame, identity/appearance, object/prop, scene, style, keyframe, storyboard, motion, camera, source-to-edit...). The default is "Unassigned - choose a reference role", and that default stops execution rather than inventing a role. Leave it unassigned and the node refuses to run - that's the guardrail doing its job.
• notes - role notes (single-role path) or media-wide notes when role_bindings is connected.
• source_fps - video only; the input batch is resampled to native H3's fixed 24 FPS.
• analysis_fps and max_analysis_frames - video only; control the generic-Qwen evidence rate (default 1 FPS, cap 16 frames). MiniMax's native tokenizer gets a separate fixed 2-FPS sample regardless.
• analysis_long_edge - max long edge for both private Qwen analysis copies (default 768). Small media is never enlarged.
• h3_length - recommended for videos: connect the H3 target length so the node mirrors native truncation/alignment before Qwen analysis. Don't feed h3_length backward from a Prompt Guide that also consumes this reference_context - that creates a graph cycle.

Outputs

• reference_context - the chain value; fan the final one to both the Guide and the Enhancer.
• h3_media - the native-ready 24 FPS, 17k+5-aligned media (or the original picture).
• routing_report - exact semantic labels, mode family, native input routes, and effective duration.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMax-H3-Guide

Restart, look under MiniMax H3/Prompting. No extra dependencies. This is the old pipeline, so it also wants the legacy Prompt Guide and Prompt Enhancer nodes present, which ship in the same pack.

The honest take

The legacy enhancer path had the same ambition as Plan v2 - let Qwen write richer prose without destroying structure - but the structure was held by system-prompt discipline instead of a typed, validated plan. It worked, and the "unassigned role stops execution" behavior shows the care that went into it. But Plan v2's Image Reference is the strict upgrade: roles are enums the compiler validates, and media routing is deterministic instead of described. Keep this node for your old graphs; don't start new ones with it.