Reasoning-Informed Visual Editing

Reasoning-Informed Visual Editing. By Xue Yang and fourteen others, from Shanghai Jiao Tong University, Southeast University, and four other institutions. Published on the eighth of October.

Large multi-modality models have made significant progress at understanding images and at generating them. Editing is where they still struggle, particularly with three demands: following complex instructions, keeping the appearance of what should stay the same, and accepting flexible formats of input. To study that gap, the authors introduce RISEBench, the first benchmark for what they call reasoning-informed visual editing, or RISE, and then extend it into RISEBench++, a more comprehensive and more fine-grained benchmark for the same emerging task. The premise is that an edit can require thought before it requires pixels. An instruction may only make sense once the model has worked out what would happen over time, what would cause what, or what the scene would look like if something were otherwise.

At its centre is a hierarchical taxonomy of reasoning. It spans six dimensions. Five are single kinds of reasoning: Temporal, Causal, Spatial, Logical and Counterfactual. The sixth, Hybrid Reasoning, integrates several of those types across edits made over multiple turns. The dimensions break down further into twelve subcategories and sixty-five fine-grained task types.

Each output is judged on three criteria: Instruction Reasoning, Appearance Consistency and Visual Plausibility. The evaluation framework has been improved for this version, combining human judges with a large multi-modality model acting as a judge, aiming at judgements that are more reliable and better calibrated.

The benchmark now contains one thousand human-annotated test cases, released in both English and Chinese. Input formats have expanded to include conditioning on multiple images, not just one. Beyond measurement, the authors also introduce RISE-Agent, a training-free agentic framework that integrates reasoning-driven planning, tool-augmented execution, and refinement guided by a verifier. It outperforms most strong existing approaches across diverse RISE tasks.

The evaluation is broad. Fifty-eight visual editing approaches are tested in all. Thirty-four of them are open-source models. Nineteen are closed-source models. The remaining five are agentic methods, the category RISE-Agent belongs to. That spread means the benchmark is read across open systems, proprietary systems, and systems that plan and call tools, rather than within one family.

The results show that reasoning-based editing remains substantially challenging. Even the strongest approach evaluated, GPT-Image two point five Sunburst, achieves only fifty-six point six percent accuracy. Put plainly, the best system tested still fails on a large share of the cases, and the rest of the field sits below it.

Two numbers frame the picture: fifty-eight approaches tested, and a best score of fifty-six point six percent. A benchmark of this kind is useful in proportion to how specific it is about where things break. Splitting reasoning into dimensions, subcategories and task types means a failure can be located, not just counted. The authors present RISEBench++ as a map of the limitations of current editing models, with insights and future directions for editing that has to reason before it changes a pixel.