ACEditStable Multi‑Turn Image Editing
via Asymmetric Conditioning

Sara Dorfman* Nir Goren* Gal Barak* Daniel Cohen‑Or Or Patashnik

* Equal contribution

Tel Aviv University

TL;DR

Image editors are trained on clean images, but in multi-turn editing they must edit their own outputs, causing small errors to accumulate. We reduce this train–inference gap by training on the editor’s own rollouts, guided by a teacher anchored to the original image and conditioned on a compound editing instruction.

Fifteen consecutive edits, each applied to the previous result.

Abstract

Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns. Each editing turn introduces small distribution shifts that accumulate over successive edits, creating a gap between the input distributions encountered during the editor’s original training and multi-turn inference. We introduce a training scheme that makes multi-turn image editing stable under this accumulating distribution shift.

A key is to force the editor to operate on its own outputs from previous editing turns during training, bringing the principle of Self-Forcing to image editing. The main challenge is that self-generated editing trajectories deviate from ground-truth trajectories, making direct supervision unreliable. We address this challenge through a distillation objective that leverages a teacher image editor to provide a reliable training signal when the model operates on its own previous outputs.

We introduce an asymmetric conditioning scheme: while the student performs edits sequentially on its own outputs, the teacher remains anchored to the original clean image and receives a compound editing instruction. Experiments show substantially reduced degradation over long multi-turn sequences while maintaining editing quality and instruction adherence.

The problem

In multi-turn editing, every output becomes the next input. Small errors compound: colors drift and details wash out, even when the editor is asked to change nothing.

FLUX.2 Klein over 20 turns: four edits, each followed by four “keep the image as is” instructions.

Method

Method overview: self-generated rollout and score-distillation supervision with asymmetric conditioning
Left: the student performs edits along a self-generated rollout, where each output becomes the input to the next turn. Given the previous output Ie−1 and the current instruction, the few-step student produces a clean estimate x̂se. Right: score distillation with asymmetric conditioning. The estimate is noised to level τ and fed to a frozen teacher and a trainable auxiliary editor. Unlike the student, both are conditioned on the original clean image I0 and a compound instruction aggregating all edits up to turn e. The teacher represents the desired clean-conditioned distribution, while the auxiliary editor models the distribution the student actually produces; the difference between their predictions pushes the student’s multi-turn outputs toward the teacher’s clean, non-degraded distribution.

Self-generated rollouts

The student trains on chains of edits, each applied to its own previous output, just as at inference.

Asymmetric conditioning

The teacher sees the clean source image and a compound instruction, never the student’s drifted outputs.

Score distillation

No ground-truth targets needed: a teacher and an auxiliary editor supervise the student’s output distribution.

Asymmetric conditioning, by example

Student

I0“Add a handbag”I1“Make it red”I2“Make it leather”I3“Add a gold buckle”x4

Edits its own outputs, one instruction at a time. I1–I3 may already have drifted.

Teacher

I0“Add a red leather handbag with a gold buckle”x4

Scores the same output as one compound edit of the clean source, so its signal never depends on drifted images.

clean source student’s own output output being trained (x4)

FLUX.2 Klein 9B4-step distilled student
0.3%parameters trained (LoRA r=16)
4.4 htraining on one B200 GPU

Results

FLUX.2 Klein before and after our fine-tuning, with the same instructions and seed. Each displayed instruction is followed by no-op instructions (“keep the image as is”), and the images show the results after those no-ops.