Krea 2 Moodboard + Identity Edit - ComfyUI Nodes (+ Forge Neo)

krea.ai-style vibe transfer and identity-preserving editing for the open Krea 2 (K2) model. Ships for both platforms: a ComfyUI node pack (main download) and a matching SD WebUI Forge Neo extension set — same algorithms, same knobs.

⚡ Plug and play. These are custom nodes purpose-built for K2 — not a wrapper around a stack of helper packs. Install this one pack and the bundled example workflows run on ComfyUI core nodes alone: one node replaces your CLIPTextEncode, and that's the whole integration. No 400-node mega-workflow to untangle, no dependency shopping list. Prefer a plain UI? The same features run in Forge Neo's txt2img tab with two extensions and one patch.

🆕 New in v1.3 : Krea 2 RedNode

One node, the whole toolkit. Subject photo + optional scene photo + style references + your instruction in — positive AND matched negative out. Presets: balanced / max identity / style only / outfit transfer. Two-ref ordering, grounded negatives, and fit geometry are handled internally, so the classic wiring mistakes are impossible. Plug the RedNode Settings node in for full manual control with plain-language dials.

👗 Outfit transfer — no prompting: drop clothing/product photos into moodboard_style, pick the preset — your character wears the outfit, patterns and accessories included, same face every time. For strict garment construction, also feed the garment into scene_image at style_strength 0.8. Multi-character scenes work too (two locked identities, one generation).

Also new: RedNode Prompt Combine (growable textboxes with an order field), friendly error messages for every classic wiring mistake, and the full production pipeline Krea_Workflow_Public rebuilt on RedNode with a SAM 3 face inpainter and Detail Daemon sampler stack.

Coming next

Region-masked reference boost (boost just the face), and timestep-windowed reference influence, (identity lock without composition lock). In testing now.

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 pass your references through as semantic description only (the model knows what's in the picture, but you get no style-vs-subject control, no strength dial, and multi-reference inputs can grid). And core's ReferenceLatent node attaches latents that K2's stock model ignores — no in-context pixel path means no true identity preservation. This pack adds both halves on top of the stock implementation: the moodboard controls and the full identity-edit recipe (in-context source latents + K2-template grounded encode + grounded negative + v1.2 fit/ref_boost), plus single-encode fusion of the two. All additive — leave the nodes unused and everything is bit-identical to stock ComfyUI.

🎨 Moodboard

Feed reference images — generations inherit their style/vibe. Training-free, built on K2's native Qwen3-VL vision conditioning:

• Vibe strength + extract mode (style ↔ subject: keep the palette/texture/mood, or keep the subject and let your prompt restyle it)

• Style crops (scramble composition so only style transfers), indirect mode (references are hidden from the image model — structurally cannot copy poses)

• Multiple references blend into one joint vibe (packed vision span — no grid/collage artifacts)

👤 Identity Edit

Instruction-based editing with the community krea2_identity_edit LoRA (port of ComfyUI-Krea2Edit): "create a photo of this person at a night market" — same face, same outfit, relit to the new scene.

• Dual conditioning: in-context VAE tokens (RoPE frame 1) + image-grounded instruction encoding

• Grounded negative (needed for CFG > 1 recipes), grounding_px likeness↔obedience dial, two-ref support (scene + subject), chainable extra sources

• v1.2 support: ref_boost reference-fidelity dial (try 2–6 on the subject ref) and the fit geometry — sources are fitted in pixel space to your output resolution (blur-proof, no more AR-matching requirement, matches the v1.2 LoRA's training)

• Fusion: identity from the edit source + style from the moodboard in one single-encode pass

• Auto face-ref prep (Forge Neo, one toggle): rebuilds your reference before editing — extracts the clean, unobstructed person (hats/glasses/hands/props removed), then builds a front-facing identity headshot. Cached per image, so it only ever runs once per reference.

ComfyUI Nodes (main download)

Unzip into ComfyUI/custom_nodes/ (or git clone the repo), restart. Nodes: Krea 2 Moodboard, Moodboard Encode (packed), Identity Edit, Moodboard + Identity Fusion, Edit Source Chain, Conditioning Rebalance (per-layer 12-tap reweighting, RMS-renormalized, presets).

Example workflows included: basic moodboard t2i, identity edit + moodboard fusion (v1.2 wiring), and Krea_Workflow_Public — the full production pipeline (JoyCaption auto-captioning, wildcards, fusion, AR handling, upscale pass; needs ~12 common node packs, dependency table in the GitHub README, all installable via ComfyUI-Manager).

Example 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. (I know Joy Caption is quite slow and also a bit outdated, you remove it or not use it you like)

Requires these custom node packs (all installable via ComfyUI-Manager):

rgthree-comfyswitches
group bypassers
Power Lora Loader
ComfyUI-mxToolkitsliders
ComfyUI-Impact-Packwildcard processor
ComfyUI-KJNodesSet/Get nodes
ComfyUI_Comfyroll_CustomNodesprompt combine, aspect ratio, text replace
ComfyUI-JoyCaptionimage to prompt captioning (downloads its captioner model on first run)
ComfyUI-Custom-Scripts (pythongosssss)text display
comfyui-ollama-describerText Transformer (regex caption cleanup). Needs the ollama pip package to load (Manager installs it); NO Ollama server required
ComfyUI-Image-Saversampler selector
comfyui_layerstyleseed node
ComfyUI-WhiteRabbitbatch Lanczos resize
Derfuu_ComfyUI_ModdedNodestext box
sam3_smartinpainter + ComfyUI-Easy-Sam3text-prompted face inpainter. Needs sam3.pt in models/sams (e.g. HF jetjodh/sam3)
comfyui-detail-daemondetail - boost in the sampler stack (keep detail_amount at 0.05 or less on short/Turbo schedules)
ComfyUI_UltimateSDUpscaletiled upscale pass. Uses 4x_NMKD-Siax_200k.pth in models/upscale_models

The two lean example workflows 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

Forge Neo (additional download)

1. git apply krea2-features-backend.patch from your Forge Neo root (current neo branch)

2. Copy the two folders from extensions/ into extensions/

Everything runs in the txt2img UI — no ComfyUI needed. Full steps in the bundled INSTALL.md.

Shared requirements

• Text encoder: Comfy-Org/Krea-2 qwen3vl_4b_bf16.safetensors (or fp8_scaled) — the VISION variant

• For editing: the identity-edit LoRA at strength 1.0 · Euler/Simple · Turbo 8 steps CFG 1 (most edits; v1.2 LoRA: 8–12 steps — 8 favors composition, 12 face detail) or Raw 20–40 steps CFG 3 (removals) · generate ≤2MP

GitHub (full docs):

ComfyUI nodes: github.com/RedNodeAI/ComfyUI-Krea2Moodboard
Forge Neo toolkit: github.com/RedNodeAI/forge-neo-krea2-toolkit

Credits

ComfyUI (Qwen3-VL ports) · Haoming02 (Forge Neo) · lbouaraba (identity-edit recipe + LoRA, Apache-2.0) · nova452 & huwhitememes (conditioning-rebalance mechanic, Apache-2.0) · ethanfel & ostris (K2 vision-conditioning recipes) · Krea.ai (Krea 2, Community License)

Code: GPL/AGPL-3.0 per repo. Not affiliated with Krea.ai, Haoming02, or the LoRA author. No model weights are redistributed here — download them from their linked sources.