Anima Artist Layer Probe

Layer routing (wlop@0-8, krenz@33%-67%) is the mixer's most powerful lever, and the one people guess at. AnimaArtistProbe replaces the guessing with data. Drop it between your model loader and the sampler, run one generation, and it measures each artist's per-layer style influence across the DiT blocks - ||artist_out − base_out|| / ||base_out|| - without altering the generated image. Then AnimaArtistProbeReport turns those numbers into a route suggestion you can paste straight back into your chain.

How to wire it

• model - your Anima model, between the loader and whatever applies the preset/sampler.
• artist_pack - from AnimaArtistPack.
• probe_steps - how many sampling steps to measure. Early steps carry the most style signal, so 4–8 is usually enough (default 6).

Outputs

• model - the patched model, passed through (this is why it sits in your normal graph).
• base_prompt - the conditioning, passed through.
• probe_id - a STRING id that ties this probe run to its report. You feed this into AnimaArtistProbeReport.

The probe pass itself doesn't change what gets generated - the README is explicit: "The probe pass does not alter the generated image." It's measuring, not mixing.

Install

Manager → search "Anima Artist Mixer Forge", or:

cd ComfyUI/custom_nodes
git clone https://github.com/peter119lee/Anima-Artist-Mixer-Forge

then restart. No extra dependencies; standard Anima files and the Anima-compatible CLIP loader. Uninstall the original Anima-Artist-Mixer if it's present (same node names).

My take

This is the node that turns "where do I route this artist?" from tribal knowledge into an experiment you can run in one queue. The natural workflow: wire the probe in, run once, read the report, take the suggested artist@lo-hi routes back into your chain, and remove the probe. And if you're wondering whether a subtle artist is being drowned out by a loud one, the probe's contribution split (in the report) answers it - it's the difference between "this artist is weak" and "this artist is weak at these layers, this late in sampling," which are different problems with different fixes.