MAI Concept Capture Flush (alpha)

The closing bracket of a Concept Lab capture. MAI Concept Capture Arm taps a model clone and writes block-state tensors to disk as a render runs; MAI Concept Capture Flush is what finalizes the thing - it writes the manifest into the concept space and reports what landed. Two nodes, one measurement.

The after trick is the whole trick

The input that matters most is the one that doesn't do anything except order execution. after accepts anything (* type), and it's there because ComfyUI runs a graph by data dependency. A flush that isn't downstream of the render can fire before the sampler has taken a single step - the node would happily write a manifest for a capture that never happened. Wire the decoded images (or whatever your last node emits) into after, and the flush is guaranteed to run after the render finishes.

If that sounds like a hack, it's the honest ComfyUI idiom: you can't express "run this last" except as a data dependency, so this node makes one out of a wire you'd otherwise ignore. The after tensor is also hashed and recorded in the capture notes as free evidence of what the render came out as - never part of the capture's identity, just a fingerprint.

The rest

• capture - the CONCEPT_CAPTURE from the Arm node
• output: a report STRING listing everything that landed in the concept space

The heavy lifting already happened during the render. The tap writes as it captures, so by the time you flush, the tensors are on disk - this node just writes the index that makes them findable, then tells you what it wrote. Nothing is held in VRAM for a whole run's worth of state, which is exactly what you want in an experiment harness that might sit next to a 33B video model.

It's alpha (2026-08-15), an output node, and part of Concept Lab - the pack's research subsystem whose stated status is "alpha, nothing measured yet." The three nodes are built and registered, but the pipeline that turns a capture into a reusable factor is explicitly not built yet. So: if you're capturing, you're doing it because you're building that research, not because this improves your renders today. It ships in ComfyUI-MAINodes (GPL-3.0, matlowai), no Python deps, installed with the pack-wide clone:

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

Restart ComfyUI and it appears under MAI/concept (alpha), guarded like every alpha node so a broken one can't take the pack down.