A training loop that generates, checks and reuses an image editor’s own edits improved Qwen-Image-Edit-2509’s scores on three benchmarks, according to an arXiv preprint posted Oct. 8. On ImgEdit, the overall score rose from 4.36 to 4.60, while the object-isolation score rose from 3.41 to 4.26. The results come from automated scoring, not an independent human evaluation.
Instruction-guided editors can produce a convincing picture that misses the requested change or alters details the user wanted to keep. Improving them often requires human-edited examples or a separate reward model to judge outputs. Rubric-CEPR, developed by researchers at Mohamed bin Zayed University of Artificial Intelligence and Aalto University, tests whether an editor can learn from its own outputs instead.
The method uses three steps. A Planner turns an unlabeled source image into an editing instruction and a structured list of what should change and what should remain. The Editor generates four candidate images. A fixed Critic then checks whether the requested change appears, the old state is removed and unrelated content is preserved. For meaning, the Critic compares the output with prompts describing the requested change, the old state to remove and details to preserve. It also compares compressed representations of the source and output to check whether changes stayed local. Separate pass-or-fail gates reject a candidate that misses any required check.
The best candidate that passes becomes a training example for a lightweight update to the editor. The main Qwen experiment used no human-edited targets or external reward model during training, though it still used unlabeled images and multiple generated candidates. The authors report a 5.5% ImgEdit gain across three training seeds, from 4.36 to 4.60 on the benchmark’s five-point scale. The largest gain was in object isolation: separating an item, such as a shirt, from its surroundings. It was the base editor’s weakest category, and its score rose 24.9%, from 3.41 to 4.26.
The same adapted checkpoint scored 8.31 versus 7.39 on GEdit-Bench and 8.97 versus 8.77 on Complex-Edit. Those benchmarks use different scoring protocols, so their numbers should be read within each test, not compared with one another. Applying the process to a second editor, Step1X-Edit, raised its ImgEdit score from 3.86 to 4.16. The paper says these evaluations use one generated edit per instruction after adaptation.
The result is an early test of self-training, not evidence that image editors can verify every kind of change. The method requires access to a model’s internal visual features and compressed image representations, and the study covers two diffusion editors with a single adaptation pass. The authors say independent human assessment and tests across more models would strengthen the evidence.
Sources
- Original paper: “Rubric-CEPR: Self-Evolving Image Editing via Reward-Verified Self-Distillation”, posted Oct. 8, 2026; see Sections 3–6 and Tables 1–3.
- Illustrative image, not part of the study: IrinaSkvo, “An example of photo editing in Tonality,” used unmodified under CC BY-SA 4.0 via Wikimedia Commons.