ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning

ACEdit. Stable Multi-Turn Image Editing via Asymmetric Conditioning. By Sara Dorfman and four others, from Tel Aviv University. Published on the fifteenth of October. No code repository is linked to it.

Image editing models do well on a single edit, but their quality often deteriorates over many turns. Each turn introduces a small distribution shift, and those shifts accumulate across successive edits. The editor was trained on one distribution of inputs, and several turns in it is being fed another: its own slightly drifted outputs. That gap, between the inputs seen in the editor's original training and the inputs met during multi-turn inference, is where the degradation comes from.

The authors introduce a training scheme meant to keep multi-turn editing stable under that accumulating shift. The key is to make the editor operate, during training, on its own outputs from previous editing turns. This brings the principle of Self-Forcing to image editing: train the model on the kind of input it will actually encounter, which is its own work.

The obstacle is supervision. Once the editor works on its own outputs, its editing trajectory drifts away from the ground-truth trajectory, and direct supervision against ground truth stops being reliable. The answer is distillation. A teacher image editor supplies the training signal while the student operates on its own previous outputs.

What makes that signal reliable is an asymmetric conditioning scheme. The student performs its edits in sequence, each on its own last result. The teacher stays anchored to the original clean image, and receives a compound editing instruction in place of the separate steps. The two see different inputs on purpose: the student sees the drift, and the teacher does not.

Experiments show substantially reduced degradation over long multi-turn sequences, while editing quality and instruction adherence are maintained. The abstract reports this without figures, so the size of the improvement is left to the paper itself.

Each half of the scheme covers the other's weakness. Training on self-generated inputs fixes the mismatch between training and use, but breaks the supervision. Anchoring a teacher to the clean image, with all the edits asked for at once, restores the supervision without bringing the mismatch back. The asymmetry in the name is exactly that split: one model sees the drift, the other gives the answer it should have led to.