Krea2 Regional LoRA

Here's the thing about Krea 2 regional generation: the interesting part - drawing boxes, per-region prompts, the attention masking - is all handled by the Builder and Apply Regional. Krea2 Regional LoRA is just the plumbing that gets a list of LoRA files from point A to point B. It's a chainable list-builder: pick a file, set a strength, and hand the result to a region.

Don't skip it for being boring, though. Without it there is no per-region LoRA at all, because normal LoraLoader applies to the whole image - and two character LoRAs loaded globally is exactly how you get the blended-face mush this whole pack exists to prevent.

How it works

Mechanically it's a list node: you choose a LoRA from your models/loras folder, set a strength, optionally pass in a previous KREA2_LORAS list via prev_loras to chain several, and it returns a combined KREA2_LORAS list. The Builder uses these per-region; Apply Regional then gates each one's effect to the image tokens inside that region's mask at forward time.

The pack matches three LoRA families against Krea 2's blocks.* and txtfusion.* layers: PEFT/diffusers (lora_A/lora_B), kohya (lora_down/lora_up + alpha), and LoKr (Kronecker-factored - lokr_w1/lokr_w2, what recent ai-toolkit builds output). Tucker/conv LoKr variants are skipped with a warning rather than silently misapplied.

The strength number matters more than you think

Global LoRA weights are tuned for a LoRA acting on the whole image. A region-masked LoRA only gets to act on the fraction of the image inside its box - so the README's guidance is to run strengths around 1.3–1.6, not the ~1.0 you'd use globally. If your character comes out weak or half-formed in its region, this is the first thing to check, before you blame the masks.

You can chain region LoRAs and they stack with ComfyUI's own weight patching - a regular LoraLoader or Power Lora Loader for a global style LoRA sits on top of the regional ones without conflict. That combination (global style LoRA + per-region character LoRAs) is the setup the pack's base_loras_exclude_regions option in Apply Regional is designed to pair with.

The one input that matters

lora_name is just a dropdown of your loras folder. strength ranges −10 to 10; negative strengths invert the LoRA. prev_loras is only for chaining - wire one Krea2 Regional LoRA's output into the next's prev_loras to stack multiple LoRAs onto one region. That's the whole node.

Install

Part of the Krea2 Regional pack:

cd ComfyUI/custom_nodes
git clone https://github.com/januspluto/ComfyUI-Krea2-Regional.git

or search Krea2 Regional in ComfyUI Manager, then restart. Requires ComfyUI with native Krea 2 support (v0.26+), no extra dependencies.

Gotchas

If a checkpoint logs unmatched keys when you apply a region LoRA, that's usually a format the pack doesn't recognize - the README asks for an issue with a few key names. And note the dedup rule if you're feeding the same file through multiple paths: the Builder's own selection wins over extra_base_loras, which wins over inline <lora:...> tags in prompts, with the first occurrence kept after path resolution.