Anima image editing: two methods for SDXL character edits with loras
AI image editing via inpainting and reference latent loras. But watch for third hands.
Get AI news that actually matters
One email a day. Zero fluff. Join 10,000+ professionals.
Two distinct methods have emerged for editing images using the Anima model in Stable Diffusion. The first method uses a split-screen approach combined with the 'Anima-lllite-inpainting-v2' ControlNet. This allows for significant changes like pose alteration, clothing swaps, and facial expression edits, but remains unstable and may fail entirely. The more extensive the changes, the less character detail remains.
The second method leverages Cosmos Reference Latent with dedicated Edit lorAs: AnimaEditV1 (by mattehe) and lora_edit_ZeroTwo (by GOOKLE). These lorAs are trained mainly on dressing/undressing, making them better at clothing and expression changes but poor at pose modification. Users report artifacts like extra hands, and the models struggle with already-clothed characters. Both approaches are experimental but promising for character consistency.
- Anima-lllite-inpainting-v2 ControlNet enables pose changes but is unstable.
- Two lorAs (AnimaEditV1, lora_edit_ZeroTwo) use Cosmos Reference Latent for clothing/expression edits.
- Both methods risk extra limbs (e.g., third hands) and cannot reliably change character poses.
Why It Matters
These workflows push AI image editing closer to controllable character consistency, but instability and artifacts remain key hurdles.