ACEdit: training scheme keeps 15-turn image edits stable via asymmetric conditioning
linoy_tsaban · x · 2026-10-18
Tel Aviv University researchers introduce ACEdit, a training scheme that stabilizes multi-turn image editing. Models are trained on clean images but must edit their own outputs at inference, so small errors accumulate into drift.
Key ideas:
- Brings Self-Forcing to image editing: training forces the editor to operate on its own previous outputs, closing the train–inference distribution gap
- Self-generated trajectories deviate from ground truth, so direct supervision is unreliable; a distillation objective uses a teacher editor anchored to the original clean image with compound instructions (asymmetric conditioning)
- Experiments show substantially reduced degradation over 15 consecutive edits while preserving quality and instruction adherence
More from Multimodal
- Dev training a pixel-art image model tired of being asked why bother — Angaisb_ · 2026-10-18
- How to keep FLUX.2 Klein 9B edits consistent in lighting across photos — CemreKArtworks_ · 2026-10-18
- Generated flow now indistinguishable from real footage, but 8-frame chunks break continuity — pixlpa · 2026-10-18
- MiniMax H3 ref2va renders 15s music video with vocals from 3 reference images — michael_goldenhour · 2026-10-18
- Creator compiles one-month progress of experimental video model project — pixlpa · 2026-10-18
- DeepSeek 4.1 Flash wows with procedural Boeing 747 Blender scene renders — victormustar · 2026-10-18