Krea2 SVDQuant Capture Save

The short version

The save half of the capture pair. Krea2 SVDQuant Capture Start hooks your model and records activation statistics while sampling runs; this node, Krea2 SVDQuant Capture Save, sits downstream of the KSampler and writes those statistics to a .safetensors file under ComfyUI/output/. That file is what makes your quantizer build an activation-aware low-rank branch - the one lever in this whole pack that's measured to improve fidelity for free at inference.

The pair exists because the statistics have to come from real sampling - real conditioning, real scheduler, real graph - and reimplementing that outside ComfyUI to run a calibration pass would be a second, subtly different sampler. So the hooks ride along on the actual generation instead.

The inputs that matter

• latent - wire the KSampler's output here. This is what forces the node to run after sampling rather than before it. ComfyUI is free to execute nodes in any order unless there's a data dependency, and this wire is the dependency. If you ever get a "no activation statistics were collected" error, it's because this wire is missing or the graph executed out of order.
• filename - default krea2_act_stats.safetensors, written under ComfyUI/output/. Keep it relative - the node rejects absolute paths and anything that escapes the output directory, on purpose.
• keep_capturing - default false. Leave it on for every prompt but the last, so several prompts accumulate into one file (that's the point of calibrating - you want a spread of prompts, not one). On the final prompt, leave it off so the hooks detach and stop recording unrelated generations.

Outputs: latent (passed through unchanged), status (where the file went and what's in it - layer count and tokens per layer), and act_stats_path (the absolute path of the file just written). Wire act_stats_path into the Krea2 SVDQuant Quantize node's act_stats input - that wire does double duty: it feeds the statistics and forces the quantizer to run after the calibration pass, instead of ComfyUI helpfully quantizing before you've captured anything.

How to install

cd ComfyUI/custom_nodes && git clone https://github.com/alperktt/Krea-2-SVDQuant-ComfyUI krea-2-svdquant

No Python dependencies. The full calibrated workflow is workflows/krea2_quantize_calibrated.json, which wires the capture and the quantize together so ordering can't go wrong.

Gotchas

• Use calibration prompts that aren't the ones you'll judge the checkpoint on. Measuring and then judging on the same prompts is how you fool yourself.
• The pre-made files are for stock weights only. If your model is a finetune or merge, sharing the architecture isn't enough - Turbo's stock calibration file loads into a base build without complaint and describes the wrong model. That's the case this capture rig exists for.
• Watch the status line for uneven token counts. A warning there means the hooks stayed live through something other than your calibration run - most likely a graph that errored between Start and Save. The stats are still usable, just not only about what you meant to calibrate on; re-run with reset=true on the first prompt if they're meant to describe one model.
• The output filename gains an -actaware tag when you pass it to the quantizer, and the file's metadata records which statistics built the checkpoint (krea2_svdquant_act_stats) - nice for keeping builds straight. Missing stats for any branched layer is a hard error, checked before any GPU work, rather than a silent fallback.