MAI Concept Capture Arm (alpha)

Before you read another word: this is a researcher's node, not a workflow node. It won't make your renders better. It exists to answer one question about MiniMax-H3 - what a piece of conditioning (a reference image, say) actually does to the model's internal block states - so that the answer can be measured, stored, and maybe compiled back into conditioning later.

That's the bet Concept Lab is making: a reusable concept doesn't have to live in trained weights. It can live in a measured functional delta. This node is the measuring instrument.

How it works

MAI Concept Capture Arm installs a tap on a clone of your model and hands the clone on. Your sampler downstream renders exactly as it would have - bit-identical, that's the whole discipline - while selected DiT block states get written to the concept space. Tensors go to disk as it captures rather than accumulating in VRAM.

One node is one arm of one comparison. The arm enum is the experiment design: R is the actual reference, N is a shape-matched control (same layout, same presentation tokens, different content), 0 is no reference, free is uncontrolled. That R-versus-N distinction is the clever part: adding a reference changes the packed layout, the presentation tokens, and the time origin of the target stream all at once, so "reference vs no reference" is a muddle. "Reference vs a shape-matched control" holds all three fixed and leaves only the content. That's the delta worth measuring.

The inputs that matter

• model - in, a patched clone out; the input is untouched
• seed - must equal the sampler's noise seed. The node can't read it (the seed lives in the sampler, not the model), so you copy it by hand, and a mismatch silently mislabels the capture - the tap now refuses on mismatch
• blocks - which DiT blocks to tap, default 0,5,10,20,30,40,49. An experiment config, not a format fact
• step_quantiles - schedule-relative positions, default 0.15,0.5,0.85
• store - what to keep: stats,frame_mean,sketch by default, or full for the whole state, which costs accordingly
• stack_label - the checkpoint name, typed by you, because a ModelPatcher doesn't carry the UNET filename and the node won't guess one
• enabled=False - the pass-through control: tap installed, records nothing. That's how you prove the instrumentation is free

Outputs are model (the patched clone) and capture, a CONCEPT_CAPTURE you feed to MAI Concept Capture Flush.

The catch that will bite you

All arms of a comparison have to run in one ComfyUI process. This pipeline is bit-identical within a launch and only close across launches, so restarting between arms makes a new replicate rather than the same condition (the pack cites DEC-CL-0015). If you're building a real comparison, queue all arms before touching anything else. It's alpha from 2026-08-15, loads behind the pack's guarded loader, and ships in ComfyUI-MAINodes with no Python deps of its own:

cd ComfyUI/custom_nodes
git clone https://github.com/matlowai/ComfyUI-MAINodes

Restart, and remember it targets H3 - whose weights are licence-locked out of the US, EU, UK and Korea, so the whole exercise assumes you're somewhere the model is legal to run.