Krea 2 Reference Encode

You know the ChatGPT trick - "the character from Picture 1, in the exact art style of Picture 2" - where the model actually does it? That's the whole job of this node. Krea's hosted product has a style-reference and moodboard module that Krea explicitly refused to open-source, so KonokoAz built the local replacement: Krea 2 Reference Encode feeds up to four images into Krea 2 as native vision tokens. No API call, no key, no img2img - it's all happening inside the text encoder.

How it actually works

Krea 2's text encoder is Qwen3-VL-4B, a vision-language model, and the DiT conditions on 12 tapped hidden states of it. Because the encoder can already see, this node doesn't need a reference-latent pathway like Flux Kontext does - it just interleaves your images into the prompt as vision tokens (Picture 1: <|vision_start|><|image_pad|><|vision_end|>), and the model reads character, style and pose from the same hidden states it was trained on. That's also the catch: it only works if your text encoder is the bf16 qwen3vl_4b build. The fp8 conversion breaks the vision tower, so references silently do nothing.

The inputs that matter

• clip - your CLIPLoader set to type krea2, loaded with the bf16 Qwen3-VL encoder. Non-negotiable.
• prompt - this is where the magic lives. Roles come from the prompt itself: character1..character4 mark identity references, style1..style4 mark style-only ones, and both expand into the exact clauses the reference LoRA trains on. So "The character1 driving a car in the style2" is a complete prompt, no extra wiring. Plain image1 / Picture 1 addressing also works, and any connected image you never mention gets a role clause appended automatically so it's not ignored.
• image1..image4 - the references, straight out of Load Image.
• imageN_detail - longest side in pixels before the vision encoder: low/normal/high/max = 256/512/1024/1280. More pixels = more vision tokens = a more literal, stronger reference. Use high for the character/subject, low/normal for pose or style hints.
• imageN_mask and bg_removal - both crop the reference to the subject and white out the background before encoding, so messy references (scenes, watermarks, stray characters) only contribute the actual subject. Wire a LoadBackgroundRemovalModel (birefnet) into bg_removal and every content-role image is cleaned automatically; style images are left alone, because their background often is the style.
• layout_split - the first fraction of denoise steps sees text-only conditioning. Set 0.25 and the layout locks in before the reference tokens can drag it into a side-by-side collage. It's a good zero-shot fix, but the LoRA makes it mostly unnecessary.

The single output is CONDITIONING, which goes into the positive input of your KSampler.

Install and gotchas

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

Restart ComfyUI (or grab it via Manager by searching "Krea 2 Reference"). It's a pure node pack - no Python dependencies to install. The models are the real cost: a Krea 2 checkpoint (UNETLoader), the bf16 qwen3vl_4b encoder (CLIPLoader, type krea2), and qwen_image_vae (VAELoader). If you're new to Krea 2, those three files are exactly what people routinely forget.

Where people get burned: the Encode node alone produces white-background character-sheet collages - you need the companion reference LoRA loaded too, which is the whole point of Krea2ReferenceLoraLoader in the same pack. Use 1024×1024 (wide canvases encourage figure duplication), and if you're on Turbo at CFG 1.0, don't be surprised the negative prompt is skipped - that's not a bug. Reference colors are character identity here, so to restyle an attribute, change it in the reference, not the prompt.