Krea 2 Moodboard + Identity Edit for Forge Neo

Two features for the Krea 2 (K2) model in SD WebUI Forge Neo, built on a full native Qwen3-VL vision-encoder integration.

On Civitai (release zips + showcase): civitai.com/models/2794961 · ComfyUI sibling: ComfyUI-Krea2Moodboard

⚡ Plug and play: one backend patch + two extensions, everything in the normal txt2img UI — no ComfyUI, no node graphs, no dependency stack.

🎨 Krea2 Moodboard

Drop reference images into a gallery — generations inherit their style / vibe (like krea.ai's Moodboard). Training-free. Controls:

• Vibe strength — 1.0 raw reference, lower = purer extract
• Vibe extract — style (palette/texture/mood; subjects fade) or subject (composition survives, style is whitened away so your prompt controls the look)
• Reference processing — full image / quadrant crops / fine 4×4 tiles (tiles = strongest composition-scrambling, style-only transfer)
• Style directive, Indirect vibe (refs hidden from the image model — structurally cannot copy pose/subject), image tokens before/after prompt
• Multiple references are packed into ONE vision span → they blend into a joint vibe (and can't trigger grid/collage outputs)

👤 Krea2 Identity Edit

Instruction-based, identity-preserving editing via community krea2_edit LoRAs (krea2_identity_edit, weights also on HF conradlocke/krea2-identity-edit). Port of ComfyUI-Krea2Edit: dual conditioning — VAE source tokens at RoPE frame 1 (appearance) + image-grounded Qwen3-VL instruction encoding (semantics), grounded negative for CFG > 1, grounding_px likeness↔obedience dial, two-ref support, aspect-ratio handling (match source / crop source to your AR / fit).

v1.2 dials (matching the LoRA's v1.2 release):

• ref_boost — reference-fidelity dial: multiplies target→reference attention (1.0 = off; the LoRA author suggests 2–6). Separate ref_boost (scene) slider for the first image in two-ref setups.
• fit source to output (v1.2) AR mode — the source is fitted in pixel space to your output resolution before VAE-encoding: blur-proof (latents are never resized), keeps your chosen AR (no more matching the source's), and matches the v1.2 LoRA's training geometry. The older match/crop modes remain for v1/v1.1 weights.
• Auto face-ref prep (2-pass) — one toggle that rebuilds the subject reference before editing: pass 1 extracts the clean, unobstructed person (hats/glasses/hands/props removed, matte natural skin), pass 2 makes a front-facing identity headshot from it. Cached per reference image (content hash) so it never re-runs for the same picture.

The two compose: enable both to take identity from the edit source and style from the moodboard.

Requirements

1. Forge Neo (neo branch). This bundle was built and tested against a July-2026 neo build — see INSTALL.md for how the backend part is applied.
2. Qwen3-VL-4B text encoder with vision weights in your Text Encoder dropdown: Comfy-Org/Krea-2 → text_encoders/qwen3vl_4b_bf16.safetensors (the fp8_scaled variant also works — its vision tower is bf16 inside). Console logs Detected Qwen3-VL-4B (vision) text encoder when correct.
3. For Identity Edit: a krea2_edit LoRA at strength 1.0 (download from its civitai page — not bundled).

What's in this bundle

Item	What	Install
extensions/	the two UI extensions	copy into <forge>/extensions/
krea2-features-backend.patch	one ~680-line git apply patch for 4 backend files (activates Neo's dormant native Qwen3-VL path + the feature hooks; fixes a latent emphasis crash on image-spliced prompts)	git apply from <forge>

Targets current Forge Neo (neo branch, July 2026+) — Neo's own Krea 2 support is required (it ships the Qwen3-VL encoder this builds on). See INSTALL.md for step-by-step instructions.

Quick settings reference

• Moodboard "Krea vibe": extract style, strength 0.5, fine tiles 4×4, directive on, position after
• Identity Edit: 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); ≤2MP; grounding_px 768 (1024+ for people); with the v1.2 LoRA use AR mode fit and try ref_boost 2–6
• Prompting edits: describe only what changes; anchor with "this person"; one edit per pass

Credits & License

• Haoming02 — SD WebUI Forge Neo (this bundle's host application)
• ComfyUI — Qwen3-VL model code ported from qwen35.py / qwen3vl.py / llama.py (source URLs in file headers, following Forge Neo's existing porting convention)
• lbouaraba/ComfyUI-Krea2Edit (Apache-2.0) — the identity-edit dual-conditioning recipe and DiT in-context forward this port implements; and the krea2_identity_edit LoRA
• ethanfel (ComfyUI-Krea2TextEncoder) & ostris — validated K2 vision-conditioning recipes
• Krea.ai — Krea 2 (weights under the Krea 2 Community License; the LoRA is a Derivative Model — see its page)

Code in this bundle is distributed under the same license as SD WebUI Forge Neo (AGPL-3.0). Not affiliated with Krea.ai, Haoming02, or the LoRA author.