Krea2 Multi-LoRA Composer

Two character LoRAs on Krea 2 in one image, and the model hands you a single person with a blended face. That smear is the single most complained-about thing in Krea 2 character work - the model routes either identity anywhere it likes, so both LoRAs fire on both figures and the characters collapse into an average. Krea2CharacterRouter (shown in ComfyUI as Krea2 Multi-LoRA Composer) is a purpose-built fix: one self-contained editor that assigns each character LoRA to a draggable region of the image, then makes the model respect the box.

It's part of the lokitsar/ComfyUI-Krea2-MultiLoRA-Composer pack and works only with Krea 2 (Raw or Turbo), its matching CLIP/VAE, and LoRAs trained for Krea 2. No API, no key, no extra Python packages - it's all local.

How it actually works

The node doesn't just stack LoRAs and hope. It routes each character adapter three ways at once, per the README:

1. Adapter-delta routing - each LoRA's image-token contribution is masked to its assigned region, so a character LoRA contributes (nearly) zero outside its box.
2. Token isolation - character text-fusion contributions apply to the matching subject phrase and are suppressed on competing phrases.
3. Attention bias - each subject phrase is encouraged to attend to its own image region and away from the others.

The base model still does the global coherence - lighting, scene, interaction - and an optional Canvas LoRA covers the unboxed scenery or the whole image with a separate style. The project credits FreeFuse for the token-separation direction. The result is one Krea 2 generation, not a paste-up: the composed prompt is built for you and exposed as the prompt output.

The inputs that matter

The node ships a custom web UI (drag regions, search LoRAs, Set + reset), and everything you set there serializes into a couple of widgets. The ones you'll actually touch:

• scene_prompt - shared scene language: camera, lighting, interaction. Keep character-specific stuff out of it; each character row's trigger and prompt describe only that character.
• regions_json - the serialized rows (LoRA path, trigger, strength, box, schedule). Usually you edit this through the canvas UI, not by hand.
• width / height - target canvas, 256–4096 in steps of 8.
• supersample_scale - how big the internal render gets (see the SuperSampler article; 1.25 is the sweet spot for likeness).
• feather (0.08 default) and overlap_policy (nearest) - how box edges blend and how overlapping regions divide. Leave them alone until you're chasing artifacts.
• canvas_lora_json - optional environment/style LoRA. Disabled by default.
• strict - on by default. It makes the node refuse to run when a routing check fails instead of silently degrading.

Outputs: model, conditioning, and latent feed the sampler; supersample_plan tells the sampler the target vs. working resolution; prompt shows the composed positive; mask_preview shows your regions; share_prompt_json is a portable scene export (import/paste from the UI).

Install

ComfyUI Manager (search "Krea2 Multi-LoRA Composer"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/lokitsar/ComfyUI-Krea2-MultiLoRA-Composer

Restart ComfyUI and hard-refresh the browser. You still need Krea 2 itself: the Raw/Turbo checkpoint, the Krea 2 CLIP, and VAE (it uses the Qwen-Image VAE under the hood). With Turbo, the README's starting point is steps 8, CFG 1.0, euler/simple - and no negative-prompt instructions in your positive fields.

Where people get burned

• Identity still bleeds if your triggers are vague or the LoRAs are weak. Enter the exact training trigger per character, keep boxes tight and separated, and start at strength 1.0.
• Regions are 2D. A face on a poster behind your character occupies the same box and can inherit the identity - the README flags this outright.
• strict mode errors when a trigger can't be aligned to Krea 2 tokens - usually means the trigger isn't in the prompt. Flip it off to get warnings instead, but you probably want to fix the prompt.
• Canvas LoRA leakage - a strongly character-trained environment LoRA can still sneak its subject into the scene.

It's one of several tools attacking this exact problem (another is CliffNodes' Krea 2 bounding-box node), and the smear it targets is real community pain. Give it a shot with two tight boxes and a shared scene prompt - this is the closest thing to a set-and-forget multi-character Krea 2 workflow right now.