LC Anima LoRA Editor

What it is

A LoRA loader that lets you edit the LoRA on the way in. It sits exactly where your normal LoRA loader sits - model in, model out, into the sampler - but instead of one strength slider you get 28 block multipliers, separate attention and MLP scaling, and two flavours of spike-softening. Then it hands you the edited LoRA on a second output so you can bake it to disk.

The pitch is: stop retraining, start editing. Fix a LoRA that locks the face, or one that repaints the entire image when you only wanted the character.

How it works

Loading goes through ComfyUI's own comfy.lora.load_lora, so anything core can read, this can read. Plain low-rank layers - the usual lora_A / lora_B files - are edited directly. Other formats (LoKr, LoHa, full diffs) get scaled per layer and converted to low rank only when you save.

Anima's blocks have four parts: self-attention (sa), cross-attention to the text (ca), an MLP, and the timestep modulation (adaln). The heatmap on the node shows 8 row groups across 28 blocks, before and after on the same scale, so you can see what your edits actually did rather than assuming. The attention slider scales sa and ca together, mlp scales the MLP, and outside_blocks covers the LLM adapter and the final layer - the two things that live outside the 28 blocks entirely.

spike_soften is the clever bit and the least obvious. layers finds a layer that's far stronger than the same layer in other blocks and pulls it down to spike_limit times their median. directions works inside a single layer and flattens whichever direction dominates it, thresholding against the second-strongest - well-behaved LoRAs sit around 2 there. both does both. Note it only applies to plain lora_A / lora_B files.

The inputs you'll actually touch

model (your Anima model), lora_name, strength (baked in when you save). Then either preset or the bars:

• Full is the LoRA as-is.
• Early blocks (0-6), Mid blocks: look (7-13), Mid blocks: look and character (14-20), Late blocks: finish (21-27) each turn those blocks down to 0.3. On the Animosity Anima LoRA the author tested, turning off 7-13 or 14-20 changed the picture most; 0-6 and 21-27 far less.
• blocks is the raw string, 28 multipliers from 0 to 2 where 1 is unchanged. Drag the bars; dragging switches preset to Custom automatically.
• blocks_in (optional, wire-only) takes a 28-value curve - the Block Scan's blocks output drops straight in and replaces the preset and the bars. Drag a bar after that and it unplugs the input, keeping its curve as your starting point.

Outputs are model (the patched model, to the sampler) and edited_lora, an LC_LORA for LC LoRA Save in the same pack. Save it and the edits and strength are baked in, so it loads at 1.0 in any loader.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC_ModelBuilder_nodes

Restart, then hard-refresh the browser - the bars and heatmap are JavaScript. No extra Python dependencies; this node runs on ComfyUI core alone.

Where people get burned

The author flags these as beta, and means it. Presets and defaults can shift between versions, so a saved workflow may need its bars set again after an update. Test every edit on a fixed seed before you save anything, and keep your original LoRA files - the editor never modifies them, but your saves can pile up and you'll want the untouched original to compare against.

And the honest ceiling: if a LoRA bakes one face into every layer, no single group of blocks frees it. Turn several down a little, or take the face out with LoRA Merge at a negative strength instead of hunting for a magic block.