Sara Dorfman, Nir Goren, Gal Barak, Daniel Cohen-Or, Or Patashnik
9 min
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.
Sam: The paper uses a handbag. The student adds it, makes it red, makes it leather, then adds a gold buckle. The teacher instead receives the original image and the instruction to add a red leather handbag with a gold buckle.
Alex: So the student practises the messy route, while the teacher judges against a cleaner route to the intended result.
Sam: That's asymmetric conditioning: different reference images and instructions for student and teacher. The teacher stays anchored to the source, without treating accumulated artifacts as details worth keeping. But it still receives the earlier edits through the combined instruction.
Alex: How does that become a training signal without making the teacher generate an exact answer for every intermediate image?
Sam: They use score distillation, learning from a teacher's predictions on noisy versions of the student's output. A separate, trainable auxiliary editor learns to describe the distribution the student currently produces. The difference between its predictions and the fixed teacher's predictions guides the student's update.
Alex: So there are two references in training: what the student currently produces, and what the clean-conditioned teacher favours.
Sam: Both teacher and auxiliary editor receive the original image and cumulative instructions. Only the student follows its own successive outputs. After training, the student edits alone, using the same sampling procedure as its unmodified backbone.
Alex: What's actually new compared with earlier attempts to stabilise editing? Training on generated histories already existed.
Sam: The paper distinguishes its method from MT-OPSD, which trains through identity edits, where intended content stays fixed. Here, generated histories contain successive semantic changes, so both content and preservation requirements evolve. The contribution is combining those histories with clean-source, cumulative teacher guidance, without paired intermediate targets.
Alex: Let's turn to evidence. A model could look stable simply because it stops making changes. How do they separate preservation from refusing edits?
Sam: Their benchmark, EdiVal-10, contains 252 chains on real photographs. Each chain contains ten cumulative local edits. Instructions target particular objects, while reserved anchor objects must remain untouched.
Alex: Local changes make the untouched areas useful checks. A global style change would make comparison with the original much harder.
Sam: That's their reason for constructing this benchmark. They measure image quality, unrequested changes, and instruction following separately. An automated vision-language judge supplies ratings, alongside learned image-distance measures on untouched objects and backgrounds.
Alex: What's the result that best captures whether the whole scene survives?
Sam: On the judge's five-point consistency scale, ACEdit scores about 3.85. Its unmodified backbone scores about 3.03. The authors report better whole-image consistency than every compared method, and the lowest background difference.
Alex: Automated judges have their own preferences. Do the other measures tell the same story?
Sam: Broadly, but not uniformly. On untouched anchor objects, ACEdit and Emu Edit are statistically indistinguishable. Final-image quality favours ACEdit over every comparison, while instruction following stays comparable to the strongest training-free methods. So the claim isn't that it wins every metric.
Alex: And the comparison with MT-OPSD isn't quite as controlled as comparing changes to the same backbone.
Sam: Correct. MT-OPSD uses a different, undistilled backbone setup, with different sampling and resolution. The paper includes that backbone separately. I'd put more weight on the matched-backbone comparisons and the experiments changing individual parts of ACEdit.
Alex: What happens when they remove the clean-reference separation? That's the most direct test of their explanation.
Sam: Giving the teacher the student's previous image substantially worsens preservation. Giving the student the original image as an extra reference improves source consistency, but weakens instruction following. Earlier colour changes get undone, as objects return to their original appearance.
Alex: So preserving the source too strongly can erase the very history the user wanted to build.
Sam: That trade-off is why the asymmetry matters. They also find shorter training histories give weaker preservation at evaluation. Training used just 96 synthetic source images, separate from the real-photo benchmark. For this setup, the reported training cost was about four-and-a-half GPU-hours on one accelerator.
Alex: That sounds accessible for this model, but where does the method still fail?
Sam: It inherits the base editor's limitations. In one example, asking a hiker to smile removes sunglasses added earlier. Both the original editor and ACEdit do this. Reducing accumulated drift doesn't guarantee that a difficult new instruction will respect every previous change.
Alex: And the evidence is strongest for local edits where some content should persist, not arbitrary transformations over indefinitely long conversations.
Sam: That's the scope. The authors show stability beyond their training lengths, but much longer sequences remain future work. They also note that repeatedly replacing subject and background can preserve visual quality without this method, because little old content must survive.
Alex: Who should read the full document, and where should they start?
Sam: Researchers building iterative image editors, or studying generators that repeatedly consume their own outputs. Start with the conditioning ablations: they show why clean guidance and faithful editing aren't the same objective. Then read the Training section for the teacher-and-auxiliary mechanism, and Quantitative Evaluation for the comparisons and judging setup.
Alex: And for someone who only needs the central lesson?
Sam: Let the editor practise on its own mistakes, but don't let those mistakes become the teacher's reference.
Alex: Keep that distinction in mind when the edits start piling up.