prompt stringclasses 3
values | expected_behavior stringclasses 3
values | failure_modes stringclasses 3
values | deliverable_text stringclasses 3
values | deliverable_files stringclasses 1
value |
|---|---|---|---|---|
Generate a single image containing exactly 8 different solar-powered terraforming vehicles. Each vehicle must be a different type. All vehicles must be fully visible and not overlapping. | The model should generate exactly 8 unique vehicles with no duplicates or missing objects. | The model often generates 7 vehicles, merges similar vehicles, or duplicates one type while omitting another. | One image with exactly 8 distinct vehicles. | ["README.md"] |
Generate a storyboard composed of 5 cubes. Each cube must show 3 visible faces. Each face represents a different scene. Total scenes: 15. No repetition. | Each cube face should be visually and semantically unique. | The model reuses visual assets across cube faces or collapses multiple faces into similar renders. | A storyboard with 15 unique scenes, one per cube face. | ["README.md"] |
Generate a multi-panel image with strict object and layout constraints in a single pass. | The model should plan before rendering to satisfy all constraints. | The model renders without planning, breaking constraints and requiring retries. | A constrained image satisfying all object and layout rules. | ["README.md"] |
Multimodal Constraint Evaluation Dataset
This dataset contains human-designed evaluation cases for multimodal image generation models.
Purpose
The goal of this dataset is to expose repeatable failure modes related to:
- object counting under strict constraints,
- loss of uniqueness across generated entities,
- layout and panel consistency,
- multi-step and multi-surface reasoning,
- planning vs rendering behavior in single-pass generation.
Dataset Structure
The dataset consists of prompt-based evaluation cases with:
- prompt description,
- expected behavior,
- commonly observed failure patterns.
No images are included by design. The dataset evaluates model behavior, not visual style.
Intended Use
- Human-in-the-loop model evaluation
- Multimodal reasoning benchmarking
- Analysis of structural reliability in generative models
License
This dataset is licensed under CC BY 4.0.
- Downloads last month
- 23