Krea 2 Identity Edit LoRA: Instruction-Based Identity-Preserving Image Editing for Krea 2 Raw
Conradlocke releases Krea 2 Identity Edit, a community LoRA fine-tune enabling instruction-based identity-preserving image editing on Krea 2 Raw — re-stage people, edit objects, and restyle with plain-language prompts.
Conradlocke released Krea 2 Identity Edit, a community LoRA fine-tune of Krea 2 Raw (12.9B single-stream MMDiT) that enables instruction-based, identity-preserving image editing.
Give it an image and a plain-language instruction — "create a photo of this person at a night market" — and it edits while preserving what you didn't ask to change, including the person's identity down to individual moles and marks, fully relit to the new scene.
Key capabilities:
- Person re-staging with likeness — relight and re-pose a person into a new scene while keeping their face, outfit, skin texture, and distinguishing marks
- Local edits — recolor, add/remove/replace objects with near-pixel preservation of untouched areas
- Full-image restyles — global style changes while preserving composition
- Replace-with-reference — "replace the woman with a big orangutan" with locality held
- Two-input compositing (experimental) — scene and person as separate references
- Composes with other LoRAs — character/body/style LoRAs stack on top
It requires the ComfyUI-Krea2Edit custom node pack for the specialized dual conditioning (in-context VAE tokens + image-grounded Qwen3-VL encoding). Two ready-made workflows are included.
Recommended settings (v1.1 — latest & recommended): use Turbo (8–12 steps, CFG 1.0) for most edits (add, recolor, restyle, re-stage), and Raw (20 steps, CFG 3.0) for removals. Match the output aspect ratio to the source image, and keep resolutions at or below 2MP (for two-person edits, prefer ~1–1.5MP and upscale). The grounding_px dial controls the edit vs. identity tradeoff: lower values = stronger edit adherence, higher values = stronger identity preservation. v1.1's trained range is 384–768 (default 768; 1024 often still works). For the older v1 release, grounding_px range was 512–1536.