FudanCVL/Unison

Unison is a comprehensive benchmark comprising 2,169 high-quality unified task samples, designed to evaluate joint understanding and generation in unified multimodal models.

Tasks

Task	Name	What it measures
IC	Internal Consistency	Internal alignment between understanding and generation.
UGG	Understanding Guided Generation	Model's ability to leverage comprehension guiding contextually generation.
GGU	Generation Guided Understanding	How synthesized outputs assist understanding tasks.
ME	Mutual Enhancement	Multi-turn iterative synergy where understanding identifies generation errors and vice versa, enabling mutual refinement.

Layout

├── Internal_Consistency/          # IC
│   ├── prompts.txt                # 556 generation prompts
│   ├── questions.json             # yes/no VQA keyed by prompt_<N>
│   └── images/                    # 556 reference images
├── Und_Guided_Gen/                # UGG
│   ├── UGG.json                   # 543 grounded generation items
│   └── images/                    # 539 source images
├── Gen_Guided_Und/                # GGU (three spatial sub-categories)
│   ├── 2D_Spatial/                # 182 2D spatial items
│   ├── 3D_Spatial/                # 180 3D spatial items
│   └── Complex_Relation/          # 180 relation items
└── Mutual_Enhancement/            # ME
    ├── ME.json                    # 264 rows, 528 samples
    └── images/                    # 264 source images

Field notes

• IC — questions.json is keyed prompt_<N> matching the 0-indexed line N in prompts.txt; each value maps question numbers to yes/no questions about the generated image.
• UGG — image_path is relative to Und_Guided_Gen/ (e.g. images/1.jpg); bbox/mask give the grounded edit region, operation the edit type.
• GGU — image_path is relative to each sub-category dir: 2D_Spatial/matrices/..., 3D_Spatial/cubes/.... Complex_Relation items carry no source image and are reconstructed from their description / validation questions.
• ME — image_path is a bare filename resolving under Mutual_Enhancement/images/.