comfyui_anima_artist_cross_attn_mix

Anima Artist Cross-Attn Mix

A ComfyUI node for the Anima model that blends multiple artist tags so their styles mix evenly instead of one artist dominating the rest.

Install

Clone into ComfyUI/custom_nodes/:

cd ComfyUI/custom_nodes
git clone https://github.com/Localsmile/comfyui_anima_artist_cross_attn_mix

Restart ComfyUI. The node appears under conditioning/anima as Anima Artist Cross-Attn Mix.

Usage

Wire model and clip in, then connect the model and conditioning outputs to your sampler. Write artist tags with an @ prefix in the text box:

masterpiece, 1girl, @(artist_a:0.8), @(artist_b:2.0), @artist_c

Artist tag syntax

• @artist_name - weight 1.0
• @(artist name:2.0) - explicit weight (spaces allowed)
• (@artist:1.5) - alternative explicit-weight form
• @(artist:-2) - negative weight subtracts that style

Mix settings

• mix_mode
  • legacy (default): the original mix. Each stacked artist adds its full weight (artists usually want 2.0+), so the push grows with every tag: six artists at 2.0 already break the image.
  • normalized: identical to legacy for a single artist, but the stacked total is divided by √(number of artists). Three artists look like legacy without its color fringing, and six or ten artists still produce a clean image. Weights mean the same as in legacy. Recommended when you stack artists.
• strength: style multiplier in normalized mode. 1.0 matches legacy for a single artist; lower it if a very large stack looks overcooked.
• start_percent / end_percent: the part of sampling where the artist mix applies. Steps outside it use the prompt without artist tags; for example, start_percent = 0.2 lets the composition settle before the style comes in.

Workflows saved with older versions load with the defaults above and keep the original behavior.

NegPiP

Non-artist tags with a negative weight are subtracted (NegPiP), with no extra node or dependency required:

1girl, @(artist_a:2.0), (speech bubble:-1.1), (signature:-1.3)

negpip_mode chooses how negative tags are applied:

• legacy (default): the negative tag's own tokens are flipped inside the prompt. This is the original behavior.
• branch: each negative tag is encoded on its own and its effect is subtracted from the result, scaled by the weight. The rest of the prompt never sees the negative concept, which avoids the tug-of-war that can cause odd results. Each negative tag costs one extra text encode and one extra attention pass per block. It also works when the prompt has no artist tags.

Compatibility

Works with Anima ControlNet-LLLite, both ComfyUI's native Apply Anima LLLite node and the sd-scripts custom node, in either order. Model-wrapper nodes applied before this one keep working.

License

MIT