Text-to-Image Rich Human Feedback: 3.7 Million Responses from 307,415 Humans

Over 3.7 million human preference responses evaluating AI-generated images, collected by Rapidata via its Python API. The data builds upon Google's research on Rich Human Feedback for Text-to-Image Generation and a previous smaller version. Collection involved 307,415 individual humans and was completed in less than two weeks.

Use Cases

• Training reward models for text-to-image generation based on human preference data.
• Benchmarking the alignment of AI-generated images with human aesthetic preferences.
• Analyzing patterns in human feedback on AI-generated visual content.
• Fine-tuning text-to-image models using direct human evaluation signals.

Strengths

• Large scale with over 3.7 million individual human responses.
• Involves a substantial participant pool of 307,415 individual humans.
• Data collection was rapid, completed in less than two weeks.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Rapidata, building upon Google's research on Rich Human Feedback for Text-to-Image Generation.

Collection Method

Collected via the Rapidata Python API from human evaluators.

Freshness

Last updated 2025-04-29 11:28:30; freshness should be verified.