H3 Full Reference Timeline Producer

If you queue this node in the current release, you'll get a validation error named perception_profile_unavailable, and the README says up front that it can't be queued - "with or without a connected visual_result." So the useful article here isn't how to run it. It's what it's for, what it needs, and what to do instead.

What it's supposed to do

It builds a partial Full-Reference timeline from qualified, decoded-CFR video observations. In plain terms: watch the reference video, work out what happens when, and produce a timeline the reference-set plan can build on. Its eleven required inputs tell you where it sits - this is the node that stitches the whole downstream producer chain together and asks, at the end, "and what did the perception say was in the video?"

The capability logic is explicit. Connect visual_result and the node reports itself as qualified_timeline. Don't, and it reports perception_profile_unavailable. In this build, execution always routes to the unavailable path - the fallback is retained on purpose ("Build qualified partial timelines and retain the legacy unavailable fallback"), so you get a typed refusal instead of a half-built timeline.

That refusal is honest about the underlying gap. Video observation in this pack is an injected boundary: the shipped package contains no decoder and no VLM runtime for it, and the perception profiles that would supply qualified observations report unavailable unless a host operator registers a local service. The reason the node still ships is that it's the declared seam - the shape the pipeline will have once qualified perception exists.

The inputs, since they map the whole route

Eleven required inputs, all typed values: request, reference_registry, media, evidence_graph and evidence_report, cross_reference_graph and cross_reference_report, directive_authority and directive_report, intent_graph and intent_report. And optionally visual_result (H3_VISUAL_PRODUCER_RESULT) - the perception output that everything hinges on.

Outputs are timeline (H3_FULL_REFERENCE_TIMELINE) and producer_report. The timeline would feed H3 Full-Reference Plan, which turns it into a plan for the shared Compiler. Reports go with graphs, always - the pack validates that pairing at every consumer.

If you take nothing else from this node, take the input list as documentation: it is the clearest statement anywhere in the pack of what a fully-qualified reference-set route is made of.

What to do instead, today

Use the base route. Build your prompt with H3 Context Request → H3 Context Plan → H3 Context Compiler, add H3 Hard Constraint Producer for dialogue and on-screen text you need word-for-word, and wire the result to the native H3 nodes with H3 Native MiniMax H3 Adapter. The sidebar does the same thing via Start H3 App Mode, which writes the official workflow for your task mode onto the canvas.

For reference-set jobs specifically, the README's guidance is to write the full-reference prompt through Plan and Compiler rather than waiting on the perception-driven timeline. You keep the generation path that works and lose only the automatic identity bookkeeping.

Install

Standard for the pack - it isn't in the Comfy Registry yet:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-MiniMaxH3-Studio.git
# restart ComfyUI

No Python dependencies declared, nothing fetched at install, prebuilt browser extension. Python 3.10+, and for generation: the native ComfyUI H3 nodes plus official H3 weights. The repo's workflows/ folder includes API-format graphs for the full-reference group, which is the least painful way to see how the eleven inputs are meant to line up.

Why it's worth knowing about

Because this node is where the pack's ambition is most visible. H3 is an omni-modal model - text, image, video, audio as one context (MiniMax H3 panel) - and the genuinely hard problem in reference-to-video isn't making a clip, it's keeping identity straight when your references arrive in different modalities: a face in a still, a body in motion in a clip, a voice in a soundtrack. That multi-subject attribution problem is the same one that deflates VLM captioners (character-consistency.md), and a timeline that binds observed evidence to directives is the sane way to attack it.

It just isn't finished on this release. If a "Full-Reference" node in your graph refuses, that's not your config - that's the release.