Anima Artist Cross-Attn (Manual/Advanced)

AnimaArtistCrossAttn is the original mixer node, the manual/advanced route this pack was forked around. Before presets existed, this was the whole show: you connected your model and your artist pack, picked how to combine and fuse the artists, dialed in a strength, and hoped. It still works, and it's still the right tool when you want direct control over the combine/fusion machinery without the preset layer translating it for you.

The knobs that matter

• combine_mode - how multiple artists become one attention signal:
  • concat - concatenate artist tokens as K/V.
  • output_avg - weighted average of per-artist attention outputs. The default, and the balanced preset's core.
  • lowrank_avg - deterministic low-rank constraint on the artist deltas. K set in options. More cross-seed stability, more "regression to the mean" with divergent styles.
• fusion_mode - how the artist signal meets the base:
  • interpolate - out = lerp(base, artist, strength).
  • concat_with_base - a single forward with K/V = [base; artist].
  • base_preserve - strips the artist component parallel to base, so the base content direction stays untouched: out = base + strength * proj_perp(artist − base). Compatible with lowrank/EMA.
• strength - 0–4. 0–1 interpolates between base and artist (1 = pure artist). 1–4 extrapolates (base + strength * (artist − base)), amplifying style. The sweet spot is 1.5–2.5; past 3 tends to oversaturate.
• enabled and apply_to_uncond (leave False - patching uncond usually breaks CFG).

Optional advanced_options (ANIMA_OPTS) and a preset input kept for compatibility with older workflows.

Outputs

model → KSampler; base_prompt (CONDITIONING) → KSampler positive. Same two outputs as PresetApply, same wiring.

The honest caveat

The README is upfront that this node can't hit SDXL-grade lossless artist mixing - Anima's LLM encoder is non-linear, so any mixing strategy introduces some distortion, and style-divergent artists may blend into a compromise. What you're buying here is controllability: weights, layer routes, timing routes, and a strength dial that behaves predictably. When a preset is connected, these manual widgets are ignored - the pack nudges you toward AnimaArtistPresetApply for preset-based work, and keeps this node around for the people who want the raw controls.

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 deps; Anima UNET + qwen_3_06b_base text encoder + qwen_image_vae, and the Anima-compatible CLIP loader. Don't install beside the original Anima-Artist-Mixer.

My take

Reach for this node when you're tuning one specific blend and you've already got a feel for what output_avg vs concat does to your artists. For everyone else, the preset route gets you 90% of the results with a fraction of the knobs - this one is for the curious and the opinionated.