ACEdit Stabilizes 15-Turn Image Editing With Asymmetric Conditioning
tau · hf · 2026-10-20
tau released ACEdit on Hugging Face, tackling the drift that accumulates over successive image edits. The method combines self-forcing training, teacher distillation, and asymmetric conditioning to keep results stable across many turns.
The demo compares Flux.2 Klein with and without ACEdit over 15 consecutive edits (add a bird, green socks, purple cage, ribbon, flowers): the baseline visibly degrades subject identity (e.g. distorted cat eyes), while ACEdit preserves consistency and quality throughout.
More from Multimodal
- AI video models finally good enough, says creator of standout scripted malfoid video — enggirlfriend · 2026-10-20
- Anam launches Cara-4.1 Express: 1.2s latency digital human beats HeyGen for 85% of blind raters — Div_pradeep · 2026-10-20
- A few prompts rebuilt Wembley in the browser with 90,360 clickable seats — DynamicWebPaige · 2026-10-20
- Fibo 2.0 takes #1 open-weights spot in image editing arena with 1087 Elo — multimodalart · 2026-10-20
- Dreamlike AI-generated 'departure' series turns airports into neon-blurred memories — egeberkina · 2026-10-20
- kohya's sd-scripts speeds up block swap by up to 20% for Anima and MOSS-Image training — bdsqlsz · 2026-10-20