SciDraw-6K: 6,291 Multilingual Scientific Illustrations Generated by Google Gemini

SciDraw-6K contains 6,291 scientific illustrations synthesized by Google Gemini image-generation models, paired with prompts in eleven languages. The dataset spans eight scientific categories, with biomedical illustrations making up 44.9% of the collection. It was created by Chen, Davie and published via Harvard Dataverse in April 2026.

Use Cases

• Multilingual text-to-image research based on the aligned prompts in eleven languages.
• Domain-adapted diffusion fine-tuning based on the scientific-illustration genre of schematic diagrams and mechanism figures.
• Prompt-engineering studies for scientific visualization based on the domain-specific prompt taxonomy.
• Benchmarking frontier image-generation models based on the specialized visual grammar of science.

Strengths

• 6,291 images provide a substantial corpus for model training and evaluation.
• Multilingual prompts in eleven languages, including English, Chinese, and Spanish.
• Clear category distribution, with biomedical (44.9%), materials (13.4%), and AI systems (11.2%) as the largest groups.
• Purpose-built for scientific illustrations, unlike general-purpose image corpora.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Data is AI-generated, which may introduce artifacts or biases inherent to the Gemini models.
• The 'other' category (6.3%) aggregates long-tail disciplines, which may have limited examples per sub-field.

Provenance

Source

Generated by Google Gemini image-generation models (primarily gemini-3-pro-image-preview and gemini-2.5-flash-image).

Collection Method

Constructed via a domain-specific prompt taxonomy, Gemini image generation, LLM-based translation, and lightweight quality control.

Time Range

Creation timestamps are included in the metadata, but a specific overall temporal coverage is not stated.

Freshness

Last updated 2026-04-20 14:12:58; freshness should be verified.

Geography

null