KiraLora_EQ
Kira LoRA EQ is a 20-band “equalizer” for a single LoRA in ComfyUI. Instead of just one strength slider, it slices the LoRA into bands across the network depth, separates cross-attention, self-attention and MLP parts, and lets you rebalance them like you would on a graphic EQ. It runs entirely in memory and does not touch your LoRA file on disk. The idea came from a very simple annoyance: character LoRAs tend to leak into the rest of the image. No LoRA is perfect, and sometimes you want to surgically keep one part of its effect while dialing down the rest. I'm an audio engineer, so my brain naturally wanted something that feels like a familiar tool: a graphic equalizer. That is why the node has a custom skin that looks more like studio hardware than a default ComfyUI box. On load, the node takes one LoRA and: - inspects its keys and detects UNet layers - classifies LoRA patches into three rough groups: - cross-attention (prompt match, text conditioning) - self-attention (shape, layout, overall structure) - feed-forward / MLP (texture, detail, color) - maps layers to 20 “bands” along the depth of the network For Z-Image, bands 1–15 are currently used. Qwen might use 16–20 as well. From my own use: faces for character LoRAs tend to live roughly in bands 10 to 15–16, and they react the most to the Texture & Color control. Your mileage will absolutely vary from LoRA to LoRA. With Z-Image With Qwen Edit 2509 Put the folder kiraloraeq into your ComfyUI/customnodes directory so that the structure looks like this: ComfyUI/customnodes/kiraloraeq/init.py ComfyUI/customnodes/kiralora…
Kira/LoRA EQ · 2 Ausgaben
