conradlocke/krea2-identity-edit

Krea 2 Identity Edit

v1.2 (recommended): krea2_identity_edit_v1_2.safetensors — better face likeness on restaged subjects and higher fidelity across every skill, from a 1024 high-resolution pass (v1.1 had none). New capabilities: character reference-sheet use and creation, head / face / eye / person swap, outpainting, inpainting, and virtual try-on. With the v1.2 nodes it also gains the FIT reference geometry — mismatched source/output aspect ratios are handled automatically, so no more stretched or blurry results — and a ref_boost fidelity dial (how hard the edit holds the reference; ~4 is a strong-likeness starting point). Requires the v1.2 nodes.

Head / face / eye / person swap is trained on stablellama's MIT-licensed change_eye_face_head_person dataset — thank you for making it available.

Low-VRAM variants: ..._v1_2_r128.safetensors (0.91GB) and ..._v1_2_r64.safetensors (0.46GB) — SVD rank-reduced from v1.2 (>99% of the weight energy retained), near-identical quality.

Prior versions remain available: ..._v1_1.safetensors (+ its _r128/_r64 variants) and the original ..._v1.safetensors.

Instruction-based, identity-preserving image editing for Krea 2 (12.9B single-stream MMDiT). Give it an image and a plain-language instruction; it edits while preserving what you didn't ask to change — including the person.

An unofficial community fine-tune of Krea 2 Raw. Not an official Krea product; not affiliated with or endorsed by Krea.ai, Inc.

Requires the ComfyUI-Krea2Edit node pack — the LoRA is trained with dual conditioning (in-context VAE tokens + image-grounded Qwen3-VL encoding) that stock nodes don't provide. Two ready-made workflows ship with it.

What it does

• Person re-staging with likeness: "create a photo of this person at a night market" — same face, same outfit down to individual moles and marks, fully relit to the new scene. New camera angles and poses included.
• Local edits: recolor, add/remove/replace objects, attribute and outfit changes, with near-pixel preservation of the rest of the frame.
• Replace-with-reference: "replace the woman with a big orangutan" — the replace verb is trained, locality holds.
• Full-image restyles: global style with preserved composition.
• Two-input edits: scene + person as separate references — "create a photo of this man next to the tractor." Input order matters and is fixed: the scene is always image 1 (source_latent/image), the person is always image 2 (source_latent_b/image_b). Swapping them sharply degrades results (this matches the training layout).
• Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior — something closed editors structurally can't offer.

Recommended settings

Task type	Model	Steps	CFG
Most edits (add, recolor, restyle, re-stage)	Turbo	8–12	1.0
Removals / large deletions	Raw	20	3.0

• ref_boost (v1.2 nodes) is the fidelity dial — ~4 gives strong face/body likeness (the example workflow ships at 4); >10 starts breaking removals; <1 loosens toward creative freedom.
• Aspect ratio: with the v1.2 nodes' fit geometry (default) a mismatched source/output aspect ratio is handled automatically. On v1/v1.1 weights or crop (legacy), match the source AR — otherwise preservation degrades.
• Generate at ≤2MP (source bleed / duplication above). For v1.1 two-person edits, prefer ~1–1.5MP — at higher resolutions the two identities may blend together; generate lower and upscale instead.
• Step count is a mild dial too: fewer steps (8) favor composition adherence, more (12) favor face detail; ~10 is a good balance.
• grounding_px is a real dial. Lower values = stronger edit adherence and more uniform scene changes; higher = stronger identity/likeness. v1.1's trained range is 384–768 (768 default); 1024 often still works nicely. If you get duplicated/split compositions ("double pictures"), lower grounding_px — running far above the trained range is the most common cause. (v1's trained range was 512–1536.)
• At CFG > 1, ground the negative too (empty prompt + same image).
• LoRA strength 1.0.

Known limitations (honest list)

• Likeness is texture-faithful, proportion-conservative. Moles, skin character, hair, and lighting adapt beautifully; strongly distinctive facial geometry (unusual nose, eye spacing, face length) regresses toward typical proportions. People whose identity lives in texture and structure transfer best; geometry-defined faces read as a "close relative."
• Two-person inputs keep outfits distinct but faces drift toward each other. Workaround that works today: chain single-ref inserts (place person A, then a second edit pass adding person B from their reference).
• Removal works but is not yet reliable — always use the Raw/CFG 3 recipe; expect occasional re-renders instead of deletions.
• Outfit swaps are hit-or-miss — changing what a person wears sometimes works cleanly and sometimes doesn't apply; reroll or rephrase.
• Local edits aren't always perfectly local — add/remove/replace operations can sometimes alter other parts of the frame or shift the overall color grade (substantially improved in v1.1). If preservation matters, compare against the source and reroll.
• Highly unusual visual content (extravagant hairstyles, extreme body types) can drift toward the base prior — a subject LoRA stacked on top fixes this.

License

The LoRA weights are a Derivative Model of Krea 2 and are distributed under the Krea 2 Community License Agreement (see also NOTICE). Key points for users: commercial use is permitted under the license's revenue threshold (§2.3, currently <$1M/yr — above that, contact Krea for an enterprise license); deployments must implement reasonable content moderation (§4.2); AI disclosure obligations apply where required (§4.3). This repository modifies the Krea Model as permitted by §3; it is not endorsed by Krea.

Research/portfolio release by a self-funded hobbyist.

Showcase

All reference people below are themselves AI-generated — no real likenesses. Prompts are embedded in each image.

v1.1

v1 gallery

---

Support

This is a solo hobby project, no company or funding behind it, and it stays free and open either way. If it's useful to you and you'd like to chip in toward GPU compute for future versions, there's a tip jar: ko-fi.com/conradlocke. Completely optional. Using the model, sharing results, and reporting issues already helps a lot. Thank you.