README.md

metadata

title: Krea 2 Identity Edit
emoji: 🧬
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false
hardware: zero-a10g
python_version: '3.10'
startup_duration_timeout: 45m
short_description: Identity-preserving instruction image editing on Krea 2
models:
  - krea/Krea-2-Turbo
  - conradlocke/krea2-identity-edit

🧬 Krea 2 Identity Edit

Instruction-based, identity-preserving image editing built on Krea 2 Turbo with the community LoRA conradlocke/krea2-identity-edit.

Give it an image and a plain-language instruction; it edits while preserving what you didn't ask to change — including the person's likeness. Person re-staging, local edits (recolor / add / remove / replace), replace-with-reference, and full-image restyles.

How it works

Krea 2 is a text-to-image MMDiT; this LoRA turns it into an instruction editor via a dual-conditioning recipe that stock Krea2Pipeline does not provide. This Space reproduces the two custom pieces from the reference ComfyUI-Krea2Edit node pack in plain diffusers / PyTorch:

• Grounded encode — the instruction is encoded together with the source image through the Qwen3-VL text encoder (image tokens inserted via the grounded chat template), tapping the 12 selected decoder layers.
• Source patch — the VAE-encoded source latent is prepended to the transformer sequence as clean tokens (RoPE frame index: source = 1, target = 0), so appearance is carried in-context and only the target tokens are denoised.

Default recipe: Turbo, 8 steps, guidance off (Krea convention). grounding_px trades edit adherence (lower) against identity/likeness (higher); 768 is balanced, 1024+ for people. For removals / large deletions, raise steps and guidance.

An unofficial community fine-tune of Krea 2 Raw — not affiliated with or endorsed by Krea.ai, Inc. Weights are distributed under the Krea 2 Community License.