Anima Artist Blend
Blend multiple artist tags in Qwen3 0.6B hidden-state space for Anima (private test)
ComfyUI-AnimaArtistBlend — a research log
Status: research log, not a production tool. This node was built to answer one question experimentally:
Can several artist styles be blended inside the hidden states of Anima's Qwen3 0.6B text encoder, instead of stacking artist tags in one prompt string?
Short answer, measured on a real pod and macstudio: not by appending fabricated tokens. Pooled, amplified, per-artist and purity-boosted virtual tokens all failed to produce a perceptible style blend, while two approaches that stay inside real token sequences did get read — each with its own catch. The measurements are the point of this repo.
Test setup
- ComfyUI 0.44 (SwarmUI backend) on an RTX PRO 6000 MIG 24GB pod, mac studio m3 ultra 512gb
- Checkpoints: Anima base v1.0 and a 2.9B hybrid merge
- Fixed-seed controlled pairs throughout: same prompt, sampler, CFG and seed; only the node mode changed
Findings
Artist vectors were measured against their shared component: in a
base + artist encoding, the artist-specific deviation is only
20.4–22.2% of the shared component — the pooled span is mostly a
summary of the whole prompt.
| Approach (version) | Result |
| --- | --- |
| pooled_token — blend spans into one appended virtual token (v0.1) | Under a fixed seed, outfit details changed; swapping one artist for another (nardack ↔ fuzichoco) produced nearly identical images |
| deviation gain ×3 on the pooled vectors (v0.2) | Deviation amplified, image still barely moved |
| per_artist_tokens — one virtual token per artist (v0.3) | Detail-level changes only |
| extract_mode=isolated — encode the tag alone (v0.4) | Deviation purity rose to 35.0%; image still detail-level. Purity and token count were not the bottleneck |
| full_average — weighted average of whole real sequences (v0.1) | First visible shift: the whole tone went pale and soft — but that is the highest-weighted artist (cierra) winning, not a blend |
| adapter_delta — per-row perpendicular artist delta on the base sequence (v0.5) | Read by the DiT, but as a re-voicing of the base prompt (contrast/rendering shifts), not as imported artist identity. Measured delta: 14.9% of base norm, 95% of it perpendicular |
Interpretation
A fabricated vector — pooled, averaged out of context, or extracted in
isolation — is off-format for the DiT: it resembles no token state the
model was trained to read. Interventions survive only where the
sequence stays close to real token states (full_average), or where
real base rows are bent in place (adapter_delta) — and neither
amounts to style blending. On top of that, a 0.6B encoder's knowledge
of low-resource artist tags is thin to begin with; some artists in the
test set have only a few hundred Danbooru images.
Projects that mix in cross-attention output space instead — see Related work — sidestep the format problem entirely, at the cost of patching the model and per-block work.
The node
Anima Artist Blend (Qwen hidden states) (conditioning/anima)
- Inputs: a
CLIP(Qwen3 0.6B from an Anima checkpoint/loader), a sharedbase_prompt, and up to 8 artist slots, each with a weight. With no artist filled in, it behaves like a plain encode. - Modes:
pooled_token,full_average,per_artist_tokens,adapter_delta;extract_mode(contextual/isolated) applies to the pooled modes;sharpenexponentiates the weights before normalising;deviation_gainamplifies artist deviation in the pooled modes;strengthscales the intervention. - The node prints its span sizes and deviation measurements to the
ComfyUI log (
[AnimaArtistBlend] ...) — the numbers in the table above come from those lines.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/seren-alegro/ComfyUI-AnimaArtistBlend
# restart ComfyUI
No extra dependencies: PyTorch only, as provided by ComfyUI.
Related work
- An1X3R/Anima-Artist-Mixer
— cross-attention output mixing plus an Adapter Mixer
(perpendicular delta after the LLM adapter). The
adapter_deltamode here follows the formula as published in that project's README; it is an independent implementation, no code copied. (That project is GPL-3.0 since its 26.8.1; this repo is MIT.) - Localsmile's Anima Artist Cross-Attn Mix — per-block cross-attention output blending.
- n0va39's ComfyUI-EasyUseAnima — conditioning-space mixing modes including delta RMS.
License
MIT — see LICENSE.