GenRef-1M: 1 Million Image Quality Improvement Triplets

1 million high-quality triplets of flawed images, corresponding high-quality images, and text reflections, collected across multiple domains using a scalable pipeline. The dataset was created by the diffusion-cot organization and was last updated on 2025-04-24. It was used to train a reflection tuning model for image quality improvement.

Use Cases

• Train image-to-image translation models based on the paired flawed and high-quality images.
• Develop models for generating textual feedback or instructions for image improvement based on the reflection text.
• Benchmark the performance of image restoration or enhancement algorithms.
• Research the relationship between visual flaws and corrective textual descriptions.

Strengths

• Contains 1 million data points, providing a large-scale resource.
• Triplets include both visual (image pairs) and textual (reflection) data, enabling multimodal tasks.
• Described as 'high-quality' and collected via a defined, scalable pipeline.

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

diffusion-cot organization via Hugging Face.

Collection Method

Collected across multiple domains using a scalable pipeline described in the referenced paper.

Freshness

Last updated 2025-04-24 19:22:11; freshness should be verified.