Krea 2 Moodboard + Identity Fusion

The recommended way to combine style references with an identity edit, and the reason the pack has three different moodboard nodes instead of one. This is the fusion node: style from a moodboard plus identity from a source image, done in one encode. Same sources, same VAE, same geometry - so the classic wiring mistakes can't happen.

Why fusion beats bolting two nodes together

You can technically get style-plus-identity by feeding a Moodboard Encode into an identity edit's fuse_with input. But that concatenates two separate encodes, which has a real weakness: indirect mode (delete the moodboard rows after encoding) can't work there, because a deleted span in a concatenated pass carries nothing. This node does the whole thing as a single encode, so the instruction and the edit grounding attend the moodboard span inside the encoder. That's what makes indirect - default ON here - genuinely functional: moodboard rows get deleted after encoding, so people in your style references structurally cannot appear in the output, while the style still transfers. If your style refs contain humans, that's the difference between "nice vibe" and "your aunt shows up in every image."

The three modes

Both image inputs are optional, and which ones you connect decides what the node is:

• edit_source only → a plain identity edit (like Krea2IdentityEdit).
• moodboard_images only → plain vibe transfer (like Krea2MoodboardEncode).
• both → the fusion. This is why you're here.

Add edit_source2 for two-ref edit LoRAs (scene first, subject second), and sources for chained extra references (3+ blends identities, same caveat as everywhere in this pack).

Wiring it up

Connect vae - it's optional in the schema but the tooltip is blunt: required for actual editing, since the identity latents come from VAE-encoding the source. Wire the output CONDITIONING as your KSampler positive. The negative is a Krea2IdentityEdit with an empty prompt and the same source image - the training unconditional. Fuse the positive only; style in the negative cancels under CFG.

The identity-side dials from the edit node are here too:

• ref_boost / ref_boost_a - the reference-fidelity attention boosters, and note they hit the identity refs only. The moodboard span is unaffected, which is the behavior you want. 1.0 off; 2–6 recommended with the v1.2 edit LoRA.
• target_latent + fit_mode - connect your empty sampling latent to enable the v1.2 pixel-space fit geometry for the edit sources. This is the blur fix: refs are fitted in pixel space to the output resolution before VAE-encoding, removing both the blur and the old "match the source aspect ratio" requirement. With CFG > 1, connect the same latent to the negative edit node too.
• strength, extract, reference_processing, budget_px, style_directive, grounding_px - the moodboard and grounding dials, same semantics as the sibling nodes.

The bundled krea2_identity_edit_fusion.json example ships the v1.2 wiring (target_latent connected, ref_boost exposed, v1.2 LoRA) so you've got a working starting point.

Installing

Whole pack, one install - GPL-3.0, zero third-party node dependencies:

git clone https://github.com/RedNodeAI/ComfyUI-Krea2Moodboard ComfyUI/custom_nodes/ComfyUI-Krea2Moodboard

or ComfyUI Manager, search "Krea2Moodboard", restart. Needs ComfyUI 0.25.0+, qwen3vl_4b (CLIPLoader type krea2), qwen_image_vae, and a krea2_edit LoRA at strength 1.0 for the identity side. If you're doing vibe-plus-face at all, start here - it's the least likely to surprise you.