Anima LoRA Block Weight

If you've been fighting an Anima LoRA that's half perfect and half wrong - right character, wrong style - you already know the problem: a single strength slider can't fix a LoRA that's strong where it shouldn't be. Anima LoRA Block Weight is the control panel that lets you decide where a LoRA's influence lands. It's the configuration half of the ComfyUI-LoRA-Block-Filter pack (by Danrisi, who also trains Anima's "UltraReal" realism line), and it pairs with Anima LoRA Filter (Blocks) - the node that actually loads and applies the LoRA. This one just builds the recipe.

What it configures

Anima's transformer has 28 blocks, each carrying six layer types the LoRA can touch: self_attn, cross_attn, mlp, and their adaln-modulated variants (adaln_self_attn, adaln_cross_attn, adaln_mlp). The whole model runs on Cosmos-Predict2, not SDXL, so don't carry over SDXL block-weight habits blindly - but the core intuition does transfer: cross-attention carries style, self-attention carries identity.

Inputs

• preset - eight options: all_enabled, character_detail, no_style_leak, face_focused, style_transfer, attn_only, mlp_only, custom. Read them as a cheat sheet: no_style_leak zeroes cross-attention and damps the adaln channels; character_detail zeroes cross-attn but keeps mlp at full; style_transfer inverts it, boosting cross-attn while cutting self-attn and mlp to a quarter.
• self_attn, cross_attn, mlp, adaln_self_attn, adaln_cross_attn, adaln_mlp - per-layer-type weights, 0 to 2 in 0.05 steps. Set custom and these become your active controls.
• block_vector - 28 comma-separated multipliers, one per block. This is the fine control: dampen block 12 alone, zero out the last four, whatever the situation calls for. Default is all 1.0.

Output

A single block_weights output of type ANIMA_BLOCK_WEIGHTS, wired into the block_weights input on the Anima LoRA Filter (Blocks) node. No model, no clip - the node is a pure config pack, so it costs you nothing to run.

Install

ComfyUI Manager finds it by pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/DanrisiUA/ComfyUI-LoRA-Block-Filter

Restart ComfyUI and it appears under Smart LoRA. There's no requirements.txt and no model to download - the pack only uses ComfyUI's own LoRA code, so install is as painless as custom nodes get.

Practical advice

The presets are genuinely usable defaults, not marketing labels - the no_style_leak preset is the one that solves the "style LoRA leaks into character" complaint that's all over Anima threads. But treat them as a starting point: the numbers that work depend on your LoRA and your prompt. And remember the version-churn rule that governs all of Anima: a LoRA trained on Preview 2 or Preview 3 won't behave on base v1.0, and block weighting can't rescue a LoRA whose keys don't line up with the model it was trained for. Retrain for the version you're actually running, then weight.