Krea 2 Moodboard Encode (packed)

The multi-reference specialist of the pack. Where the plain Krea2Moodboard node can end up producing a grid/collage when you feed it several images, Krea2MoodboardEncode packs all your references - or all their crops - into one vision span. Structurally grid-safe, and the references blend into a joint vibe instead of fighting each other. If your moodboard is a folder of eight images, this is the node.

How it works

The trick is in the packing: instead of one vision span per image (which is what pushes the model toward rendering a grid), the references are stitched into a single span that the Qwen3-VL encoder sees as one thing. The model reads the collection as a joint style signal rather than a set of subjects. That's also why it's the right feeder for the identity-edit path - more on that below.

The knobs

The controls are the same family as the plain moodboard node, so if you've used that one you know your way around:

• strength - 1.0 = raw reference detail; lower = purer extract of the selected aspect. Same information-knob semantics as the other node.
• extract - style / vibe vs subject / concept.
• reference_processing - full image, quadrant crops (2x2), or fine tiles (4x4). Crops scramble composition so only style survives.
• indirect - delete the reference tokens after encoding so the DiT never sees them; style arrives via prompt re-contextualization. Can't copy pose or subject. Structurally grid-proof on top of the single-span packing.
• position - before prompt or after prompt.
• style_directive - the declarative "style from the refs, subjects from the text" sentence.
• budget_px - this one's specific to the packed node: the area budget per reference fed to the vision encoder (384 default). Lower it to keep total tokens manageable when you batch a lot of images.

images is optional - batch several for a joint moodboard, or leave it unconnected for a plain Krea 2 text encode.

The two ways to use it

Standalone: the obvious one. Wire it as your KSampler positive, give it a prompt and a batch of references. Output is a single CONDITIONING.

As the fuse_with feeder for Krea2IdentityEdit: this is the trick worth knowing. Feed a Moodboard Encode conditioning into the identity edit's fuse_with input and you get style-from-references fused with identity-from-source in one pass. Two rules from the README: use an empty prompt on the feeder, and keep indirect OFF - fuse_with concatenates a separate encode, so if the vision rows were deleted they'd carry no image influence at all. (The dedicated Krea2MoodboardIdentityFusion node exists precisely because it can do indirect properly inside a single encode; use that when you want the safest fusion.)

Installing

Same as the whole pack - no third-party node deps, GPL-3.0:

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

then restart, or grab it through ComfyUI Manager by searching "Krea2Moodboard". Prerequisites are the usual Krea 2 stack: ComfyUI 0.25.0+, qwen3vl_4b text encoder via CLIPLoader type krea2, and qwen_image_vae.

If you catch yourself fighting grid artifacts with the plain moodboard node, this is your upgrade path.