Researchers have introduced RISEBench, described as the first benchmark for evaluating Reasoning-Informed viSual Editing (RISE), and its expanded successor RISEBench++, according to a paper posted on Hugging Face. The work targets a persistent gap: while Large Multi-modality Models (LMMs) have advanced in visual understanding and generation, they still struggle with complex editing instructions, appearance consistency, and flexible input formats.
RISEBench++ organizes the task into a hierarchical taxonomy of six reasoning dimensions: Temporal, Causal, Spatial, Logical, Counterfactual, and Hybrid Reasoning, the latter integrating multiple reasoning types across multi-turn edits. These are broken down further into 12 subcategories and 65 fine-grained task types. The benchmark scales to 1000 human-annotated test cases, released in both English and Chinese, and expands input formats to include multi-image conditioning.
The evaluation framework assesses three axes — Instruction Reasoning, Appearance Consistency, and Visual Plausibility — using human judges alongside an LMM-as-a-judge approach intended to yield more reliable and calibrated judgments.
Beyond benchmarking, the authors present RISE-Agent, a training-free agentic framework that combines reasoning-driven planning, tool-augmented execution, and verifier-guided refinement. The paper reports that it outperforms most strong existing approaches across diverse RISE tasks.
In total, the study evaluates 58 visual editing approaches: 34 open-source models, 19 closed-source models, and 5 agentic methods. The results point to substantial remaining challenges, with the strongest evaluated approach, GPT-Image-2.5 Sunburst, achieving only 56.6% accuracy. The authors position RISEBench++ as a tool for highlighting the limitations of contemporary editing models and for charting future directions in reasoning-aware visual editing.