Krea 2 Reference LoRA (layered)

Here's the dirty secret of reference-image conditioning on Krea 2: feed images in with the Encode node alone and the model tends to ignore the scene text and just collage your character sheet onto a white background. krea2_reference_v1 is the fix for that, and this node is the only loader that applies it the way it was trained. Two strength knobs instead of one, each aimed at a different part of the network.

Why two knobs?

The LoRA was trained on ~2.7k reference-following triplets, and its weights don't land in one place. The reference addressing, anti-bleed and style transfer all live in the conditioning path - the text_fusion and txt_in tensors. The rest of the DiT-block tensors mostly carry identity drift. So this loader splits the file into those two groups and gives each its own strength, instead of forcing one number across the whole model like a normal LoraLoaderModelOnly would.

The recommended settings are blunt about it: strength_text_fusion 1.0, strength_dit_body 0.0. That keeps the base model's rendering completely untouched while the references are followed - best identity, best for comics. Raise strength_dit_body toward 0.3 if a hard prompt needs extra scene adherence, but know that identity drifts above that, and the tooltip warns it pulls toward the LoRA's (colorful) training distribution. If you want to restyle an attribute, change the reference image, don't crank this knob.

The inputs

• model - your Krea 2 model, straight from UNETLoader.
• lora_name - an enum populated from your models/loras/ folder. You need krea2_reference_v1.safetensors downloaded from Civitai or Hugging Face; it is not bundled with the pack.
• strength_text_fusion - default 1.0, range -2 to 2. The reference addressing / anti-bleed / style-transfer layer. This is the knob that makes references follow.
• strength_dit_body - default 0.0, range -2 to 2. Extra scene adherence, at the cost of identity drift above ~0.3.

The single output is MODEL, which feeds your KSampler. Chain it right after the UNETLoader and before the sampler.

Install

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/KonokoAz/ComfyUI-Krea2-Reference

Restart ComfyUI. No Python dependencies. Then grab krea2_reference_v1.safetensors from Civitai/Hugging Face and drop it in models/loras/ - the node shows nothing in lora_name until you do. The rest of the stack is the usual Krea 2 trio: the checkpoint, the bf16 qwen3vl_4b text encoder (fp8 breaks the vision tower), and qwen_image_vae.

Troubleshooting

• The LoRA dropdown is empty - the file isn't in models/loras/. Check the path, refresh, restart.
• References have no effect - either the Encode node isn't in the graph (this LoRA does nothing without vision tokens to address), or strength_text_fusion is at 0. The two nodes are halves of one system.
• Identity drifts or the scene looks off - strength_dit_body too high. Keep it at 0.0 and let strength_text_fusion do the work.
• Character sheets produce multi-view clones - that's a known LoRA limit, not your settings; use single-pose references instead.
• You hate the second knob - a regular LoraLoaderModelOnly at strength 0.7–1.4 also works (slight identity drift at the top end), but you lose the ability to isolate the conditioning path.

Known limits worth pricing in before you build a whole pipeline: style references strongly tied to a franchise will drag its cast along, and two-reference prompts (character + style together) work only partially - v2 is in progress. One last tip: the reference's colors are the character's identity here. Feed a grayscale copy of the character plus "the entire image is black and white" and you get a monochrome scene; keep the color ref and the character keeps her palette no matter what the prompt demands.