krea.ai-style moodboard / vibe transfer and identity-preserving editing for the open Krea 2 model, as ComfyUI nodes. Companion to the Forge Neo version — same algorithms, same knobs.

On Civitai (release zips + showcase): civitai.com/models/2794961

⚡ Plug and play: purpose-built nodes, zero third-party node dependencies — the bundled example workflows run on ComfyUI core nodes + this pack alone. One node replaces your CLIPTextEncode; that's the whole integration.

Why not just the stock nodes?

Stock ComfyUI runs Krea 2 text-to-image perfectly — this pack exists for what the stock nodes can't do with reference images:

Stock image-reference encodes (the qwen-edit style nodes) pass references through the encoder as semantic description only, using QwenImage's template rather than K2's: the model learns what's in your reference, with no control over which aspect transfers (style vs subject), no strength dial, and multi-reference inputs can collapse into grid/collage outputs.
Core's ReferenceLatent node attaches latents that K2's stock model ignores — there is no in-context pixel path, so no true identity preservation, no edit-LoRA support.

This pack adds both halves on top of the stock implementation: the moodboard controls (strength, style↔subject extraction, crops, indirect mode, grid-safe packed spans) and the full identity-edit recipe (in-context source latents at RoPE frames 1..N, K2-template grounded instruction + grounded negative, v1.2 fit geometry, ref_boost) — plus single-encode fusion of the two. All additive: leave the nodes unused and every patched path is bit-identical to stock ComfyUI.

Nodes

Krea 2 RedNode (Moodboard + Identity) — start here

The simple front door: subject_image (the face to keep), optional scene_image (two-ref order handled for you), moodboard_style references, your instruction, and a preset (balanced / max identity / style only). Outputs positive AND matched grounded negative — same sources, same VAE, same fit geometry by construction, so the classic wiring mistakes can't happen. Connect output_latent to your sampling latent for the v1.2 blur-proof geometry. For full manual control, plug a Krea 2 RedNode Settings (Advanced) node into settings — it replaces the preset entirely and names every dial in plain language (technical terms in the tooltips). The nodes below remain as advanced/legacy building blocks.

Krea 2 Moodboard

One-node vibe transfer: prompt + reference image(s) in, conditioning out (replaces CLIPTextEncode on the positive). The image inputs are optional — with nothing connected it behaves exactly like a plain Krea 2 text encode, so you can leave it wired in and just unplug the references. Knobs:

strength — 1.0 = raw reference detail (layout/pose can leak); lower = purer extract. This is an information knob, not a multiplier (per-token RMSNorms erase plain scaling).
extract — style: palette/lighting/texture/mood survive, subjects fade (spans collapse toward a mean/±std statistics signature). subject: statistics are whitened away, subject/composition survives and your prompt controls the look.
reference_processing — full / 2×2 crops / 4×4 tiles (tiles: subjects are largely never encoded — strongest style-only setting).
indirect — reference tokens are deleted after the text encoder ran: the DiT never sees them, style arrives only through prompt re-contextualization. Cannot copy poses; also the safe mode for crops/multi-ref (deleted spans cannot grid).
style_directive — declarative "style from the refs, subjects from the text" sentence (auto-matches the extract mode).

Krea 2 Moodboard Encode (packed)

The multi-reference specialist: all references (or crops) are packed into ONE vision span — structurally grid-safe, references blend into a joint vibe. Same knobs; images is optional (unconnected = plain Krea 2 text encode). Use it standalone (with a prompt) or as the fuse_with feeder for the identity node (empty prompt, indirect OFF — fuse_with concatenates a separate encode, so deleted rows would carry no image influence).

Krea 2 Moodboard + Identity Fusion

The recommended way to combine style refs with an identity edit: one node, one encode — the instruction and the edit grounding attend the moodboard span inside the encoder, so indirect (default ON) genuinely works: moodboard rows are deleted after encoding, people in your style refs cannot appear in the output, and the style still transfers. Wire it as the KSampler positive (connect vae!), keep the negative a Krea 2 Identity Edit with an empty prompt + the same source. Both image inputs are optional: connect only edit_source for a plain identity edit, only moodboard_images for plain vibe transfer, both for fusion — one node covers all three modes. The identity-side dials from the edit node are here too: ref_boost/ref_boost_a (they boost the identity refs only — the moodboard span is unaffected) and target_latent + fit_mode (v1.2 fit geometry for the edit sources).

Krea 2 Identity Edit

Instruction-based identity-preserving editing with community krea2_edit LoRAs (krea2_identity_edit, weights also on HF conradlocke/krea2-identity-edit): "create a photo of this person at a night market" — same face, same outfit, relit. Dual conditioning: clean source latents ride in-context at RoPE frames 1..N (the LoRA's preserve-this signal) + the instruction is grounded on the source through Qwen3-VL. grounding_px = likeness↔obedience dial (768 balanced, 1024+ for people).

Use two of these: positive (instruction) + negative (empty prompt, same image) — the training unconditional, needed for CFG > 1 recipes.
ref_boost — reference-fidelity dial: multiplies target→reference attention (additive logit bias). 1.0 = off; >1 pulls harder toward the reference's appearance (the v1.2 LoRA author suggests 2–6); <1 loosens. Applies to the LAST ref (= the subject); ref_boost_a is the same dial for the scene ref in two-ref workflows. Set on the positive node only.
target_latent + fit_mode — connect your (empty) sampling latent to enable the v1.2 fit geometry: refs are fitted in pixel space to the output resolution before VAE-encoding. Fixes blurry results from resolution mismatches (latents are never resized) and removes the old "match the source aspect ratio" requirement (AR-preserving fit at a centered stride-1 offset, matching v1.2 training). crop (legacy) keeps the v1/v1.1 geometry for older weights. With CFG > 1, connect the same latent to the negative node too so both passes share one geometry.
fuse_with input: feed a Moodboard Encode conditioning to fuse style-from-moodboard with identity-from-source. Fuse the POSITIVE only (style in the negative cancels under CFG).
Two-ref (experimental upstream): scene in image, subject in image2.

v1.2 LoRA notes (krea2_identity_edit_v1_2.safetensors): adds head/face swap, inpaint/outpaint grounding, try-on, character sheets, and a 1024 high-res pass; on Turbo run 8–12 steps (8 favors composition, 12 favors face detail). The fit default matches how v1.2 was trained; use crop (legacy) with v1/v1.1 weights.

Krea2 Edit Source Chain

Chainable multi-reference input: each node appends one image; connect chains into the sources input on Identity Edit or the Fusion node (frames 3..N after image/image2). Unlimited by the architecture — but the edit LoRA trained on 1–2 references, so 3+ tends to blend identities (the LoRA author's multi-person recipe is chaining edit passes instead: place person A, then run a second edit adding person B from their reference).

Krea 2 Conditioning Rebalance

Per-layer reweighting of K2's conditioning: the model conditions on a 12-layer Qwen3-VL stack whose taps the DiT mixes linearly — shallow taps carry broad syntax/composition, deep taps carry fine detail (identity, texture, precise attributes). Insert between any conditioning node and the sampler. Presets: balanced (the classic community profile — deep taps 2.5/5.0/1.1/4.0), detail, subtle, uniform, or custom weights. renormalize (default ON) holds the overall magnitude so only the tap ratios change — the quality-preserving mode; global amplification (multiplier > 1) mostly degrades likeness/color. Compatible drop-in for workflows built around the community "Conditioning Krea2 Rebalance" node (its behavior = renormalize OFF, multiplier 4.0).

Example workflows (`workflows/`)

Both basic examples run the positive through Krea 2 Conditioning Rebalance (balanced, renormalized — set preset to uniform to bypass), and the fusion example ships the v1.2 wiring (target_latent connected, ref_boost dials exposed, v1_2 LoRA).

krea2_rednode_identity.json — start here: identity edit on the RedNode (max identity preset)
krea2_rednode_style_transfer.json — start here: vibe transfer on the RedNode (style only, no LoRA needed)
krea2_moodboard_t2i.json — basic vibe transfer text-to-image (legacy nodes)
krea2_identity_edit_fusion.json — identity edit + moodboard style fusion (legacy nodes)

Full pipeline: `Krea_Workflow_Public.json` (advanced)

The author's complete daily-driver workflow: JoyCaption auto-captioning of scene/subject references, wildcard prompting, identity edit + moodboard fusion, AR handling, optional upscale pass, group bypass switches. Requires these custom node packs (all installable via ComfyUI-Manager):

| Pack | Used for | |---|---| | rgthree-comfy | switches, group bypassers, Power Lora Loader | | ComfyUI-mxToolkit | sliders | | ComfyUI-Impact-Pack | wildcard processor | | ComfyUI-KJNodes | Set/Get nodes | | ComfyUI_Comfyroll_CustomNodes | prompt combine, aspect ratio, text replace | | ComfyUI-JoyCaption | image -> prompt captioning (downloads its captioner model on first run) | | ComfyUI-Custom-Scripts (pythongosssss) | text display | | comfyui-ollama-describer | Text Transformer (regex caption cleanup) — needs the ollama pip package to load (Manager installs it); NO Ollama server required | | ComfyUI-Image-Saver | sampler selector | | comfyui_layerstyle | seed node | | ComfyUI-WhiteRabbit | batch Lanczos resize | | Derfuu_ComfyUI_ModdedNodes | text box | | sam3_smartinpainter + ComfyUI-Easy-Sam3 | text-prompted face inpainter — needs sam3.pt in models/sams (e.g. HF jetjodh/sam3) | | comfyui-detail-daemon | detail boost in the sampler stack (keep detail_amount ≤ 0.05 on short/Turbo schedules) | | ComfyUI_UltimateSDUpscale | tiled upscale pass — uses 4x_NMKD-Siax_200k.pth in models/upscale_models |

The two lean example workflows above need NONE of these — core nodes + this pack only.

Settings baked in: ModelSamplingAuraFlow shift 1.15 (= ComfyUI's stock Krea 2 default — the node is there as a handle; raise it for Raw-checkpoint recipes), Euler/Simple, Turbo 8 steps CFG 1 (removals: Raw checkpoint, 20–40 steps, CFG 3). With the v1.2 LoRA, 8–12 steps (8 = composition, 12 = face detail). Generate ≤2MP. Matching the output AR to the source is no longer required when target_latent is connected (fit geometry) — but staying close still gives the best results.

Requirements

ComfyUI with native Krea 2 support; qwen3vl_4b text encoder (vision weights) via CLIPLoader type krea2; qwen_image_vae.
For editing: a krea2_edit LoRA at strength 1.0 (LoraLoaderModelOnly).

Install

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

How it works / credits

Small additive patches at import: packed list-spans in Qwen3-VL preprocessing, moodboard effects inside Krea 2's encode_token_weights, and the in-context ref-latents branch on the Krea 2 DiT (reference_latents conditioning contract, like QwenImage/Flux edit models). All paths are bit-identical to stock when the nodes aren't used.

Credits: ComfyUI · lbouaraba/ComfyUI-Krea2Edit (Apache-2.0 — the identity-edit dual-conditioning recipe this reimplements) · nova452/ComfyUI-ConditioningKrea2Rebalance & huwhitememes/comfyui-krea2-conditioning (Apache-2.0 — the per-layer rebalance mechanic and its RMS-renormalized variant) · ethanfel & ostris (K2 vision-conditioning recipes) · Krea.ai (Krea 2, Community License). License: GPL-3.0 (ComfyUI-compatible). Not affiliated with Krea.ai.

ComfyUI-Krea2Moodboard

Nodes (9)