saneval/saneval-release

SANEval Release

SANEval (Structured Attribute and Numerical Evaluation) is a benchmark dataset for evaluating text-to-image generation models across five visual attribute categories: spatial reasoning, numeracy, color, shape, and texture.

Each record pairs a text prompt with an image generated by one of six state-of-the-art text-to-image models, enabling structured evaluation of prompt-image alignment at scale.

Dataset Summary

Field	Details
Task	Text-to-image generation evaluation
Categories	Spatial, Numeracy, Color, Shape, Texture
Models	Imagen 3.0, Imagen 4.0, Imagen 4.0 Ultra, Nano Banana, Seedream 3.0, GPT Image 1 (all commercially available)
Splits	simple (standard prompts), hard (adversarial prompts)
Total samples	~30,000
License	MIT

Schema

Column	Type	Description
prompt	string	Text prompt used to generate the image
model	string	Name of the text-to-image model
split	string	simple or hard
type	string	Benchmark category: spatial, numeracy, color, shape, texture
image	image	Raw generated image (PNG)

Source Experiments

Experiment	Split	Type
saneval-numeracy	simple	numeracy
saneval-spatial	simple	spatial
saneval-color	simple	color
saneval-shape	simple	shape
saneval-texture	simple	texture
saneval-hard-numeracy	hard	numeracy
saneval-hard-spatial	hard	spatial
saneval-hard-color	hard	color
saneval-hard-shape	hard	shape
saneval-hard-texture	hard	texture

---

Responsible AI (RAI) Documentation

Data Limitations (rai:dataLimitations)

SANEval prompts are structured and template-based, covering five visual attribute categories. The dataset does not cover open-ended or compositional prompts beyond these categories. Hard-split prompts are adversarially constructed and may not reflect naturalistic user queries. Model coverage is limited to six commercially available text-to-image systems available at collection time; results may not generalize to other architectures. The simple split contains approximately 1,000 prompts per model per category; the hard split contains approximately 50. Evaluations derived from this dataset should account for these distributional constraints.

Data Biases (rai:dataBiases)

Prompts are constructed from curated templates and may reflect biases present in the underlying design choices, including Western-centric object categories and color terminology. The dataset does not systematically represent culturally diverse visual concepts or non-English linguistic structures. Model outputs inherit any biases present in the respective training corpora of the six evaluated models. Users should be aware that benchmark scores reflect performance on this specific prompt distribution and may not generalize to broader real-world usage.

Personal and Sensitive Information (rai:personalSensitiveInformation)

This dataset does not contain personal or sensitive information. All prompts are synthetically constructed and describe objects, spatial arrangements, quantities, colors, shapes, and textures. No images of real individuals, personally identifiable information, health data, political or religious content, or sensitive demographic attributes are present.

Data Use Cases (rai:dataUseCases)

This dataset is intended for benchmarking and evaluating text-to-image generation models on structured visual attribute understanding. Validated use cases include: comparative model evaluation across attribute categories, analysis of prompt difficulty (simple vs. hard), and scoring pipeline development and validation (e.g., object detection, synonym mapping, VQA-based scoring). This dataset is not validated for fine-tuning generative models, safety or toxicity evaluation, or human preference modeling.

Data Social Impact (rai:dataSocialImpact)

SANEval supports the development of more reliable and interpretable evaluation methods for text-to-image generation, which can contribute positively to model accountability and transparency. Potential risks include over-reliance on benchmark scores as a proxy for real-world model quality, or misuse of the dataset to optimize models specifically for this benchmark without broader generalization. The dataset is released under a private access model to allow for responsible use and review prior to broader dissemination.

Synthetic Data (rai:hasSyntheticData)

True. All images in this dataset are synthetically generated by text-to-image models (Imagen 3.0, Imagen 4.0, Imagen 4.0 Ultra, Nano Banana, Seedream 3.0, GPT Image 1) conditioned on structured text prompts. Prompts are constructed from curated benchmark templates designed to systematically probe specific visual attributes.

Source Datasets (prov:wasDerivedFrom)

Prompts are derived from the SANEval benchmark prompt suite, designed to systematically evaluate visual attribute understanding. Images are generated outputs from six commercially available text-to-image generation APIs. No publicly released upstream dataset was used as a direct source.

Provenance Activities (prov:wasGeneratedBy)

Collection: Prompts were carefully designed using structured templates targeting five visual attribute categories (spatial, numeracy, color, shape, texture) across two difficulty tiers (simple, hard). Images were generated via API calls to six commercially available text-to-image models. All generation runs were tracked using MLflow.

Preprocessing: Images were stored as PNG files and downloaded from S3 artifact storage. Entries with failed image generation (failed_imagegen flag) were excluded. Images are provided as-is without post-processing or filtering.

Annotation: No human annotation was performed on the images. Automated scoring was applied using object detection (YOLO), VQA-based scorers (Gemini, Llama), and synonym mapping modules as part of the SANEval evaluation pipeline. Score metadata is not included in this release dataset.

Tools and platforms: MLflow (experiment tracking), AWS S3 (artifact storage), Hugging Face datasets library (dataset packaging and upload).