Krea 2 Identity Edit is an unofficial community fine-tune of Krea 2 Raw that enables instruction-based, identity-preserving image editing through plain-language commands. Simply provide an image and a natural-language instruction — the model edits accordingly while preserving everything you didn't ask to change, including the person's identity.

The latest recommended version, v1.2, introduces a 1024 high-resolution pass for significantly better face likeness on restaged subjects. New capabilities include character reference-sheet creation, head/face/eye/person swap, outpainting, inpainting, and virtual try-on. The v1.2 nodes add automatic FIT reference geometry — mismatched source and output aspect ratios are handled gracefully, eliminating stretched or blurry results — along with a ref_boost fidelity dial, where ~4 is a strong starting point for likeness preservation.

Key capabilities include person re-staging with full likeness preservation (same face, same outfit down to individual moles and marks, fully relit to new scenes with new camera angles and poses), local edits such as recolor and object replacement, replace-with-reference operations, full-image restyles, and two-input edits where scene and person are provided as separate references. Crucially, Krea 2 Identity Edit composes with your existing character, body, and style LoRAs — a flexibility that closed editors structurally cannot offer.

The training code is fully public via the krea2edit-trainer extension. Low-VRAM users can choose from rank-reduced variants — 0.91GB (r128) and 0.46GB (r64) — retaining over 99% of the weight energy with near-identical quality. The head/face/eye/person swap capability was trained on stablellama's MIT-licensed dataset.

Recommended settings vary by task: Turbo mode at 8–12 steps and CFG 1.0 suits most edits, while Raw mode at 20 steps and CFG 3.0 is advised for removals. The ref_boost dial controls how strongly the edit holds the reference likeness, and grounding_px manages the balance between edit adherence and identity preservation. Generation should be kept at or below 2MP to avoid source bleed.

Known limitations are acknowledged honestly: likeness is texture-faithful but proportion-conservative, meaning distinctive facial geometry may regress toward typical proportions. Two-person edits can cause identity blending, and removals remain imperfect. Local edits can occasionally shift surrounding content. Stacking a subject LoRA on top addresses drift with unusual visual content. Requires the ComfyUI-Krea2Edit node pack. Not affiliated with or endorsed by Krea.ai, Inc.Krea 2 Identity Edit is an unofficial community fine-tune of Krea 2 Raw (a 12.9B single-stream MMDiT) that brings instruction-based, identity-preserving image editing to ComfyUI workflows. Give it an image and a plain-language instruction, and it edits accordingly while preserving everything you didn't ask to change — including the person's face, outfit, and distinguishing marks.

The latest recommended release, v1.2, introduces a 1024 high-resolution pass for stronger face likeness on restaged subjects. New capabilities include character reference-sheet creation, head/face/eye/person swap, outpainting, inpainting, and virtual try-on. The v1.2 nodes add automatic FIT reference geometry so mismatched source and output aspect ratios are handled gracefully — no more stretched or blurry results — plus a ref_boost fidelity dial (around 4 for strong likeness) that controls how firmly the edit holds to the reference.

Core capabilities cover person re-staging with full likeness preservation (same face, same outfit down to moles, fully relit to new scenes with new camera angles and poses), local edits like recolor and object replacement, replace-with-reference operations, full-image restyles, and two-input edits where a scene and a person are supplied as separate references. Notably, the LoRA composes with your existing character, body, and style LoRAs — they stack on top and steer the prior, a flexibility that closed editors structurally cannot offer.

Low-VRAM users can choose from rank-reduced variants at 0.91GB or 0.46GB, retaining over 99% of the weight energy with near-identical quality. The training code is fully public via the krea2edit-trainer extension, and the head/face/eye/person swap skill was trained on stablellama's MIT-licensed dataset.

Recommended settings: Turbo at 8–12 steps and CFG 1.0 for most edits, Raw at 20 steps and CFG 3.0 for removals. Keep generations at or below 2MP to avoid source bleed. The ref_boost and grounding_px dials offer fine control over likeness strength and edit locality respectively.