Scone-S2I-57K: 57,000 Subject-Driven Image Generation Pairs

Scone-S2I-57K provides 57,000 records for subject-driven image generation, released by Ryann829 in early 2026. The collection features 22,000 single-candidate subject images and 35,000 multi-candidate subject images designed for training unified understanding-generation models.

Use Cases

• Training subject-driven generation models using the 22,000 single-candidate subject records
• Enhancing multi-subject compositional accuracy using the 35,000 multi-candidate subject samples
• Fine-tuning image-to-image models to maintain subject identity across varied text prompts

Strengths

• 57,000 total records for subject-driven generation tasks
• Explicit distinction between 22K single-subject and 35K multi-subject training pairs
• Apache 2.0 licensed

Limitations

• Missing documentation for specific column names and internal schema
• Potential for selection bias in the 'refined' 22K single-candidate subset

Provenance

Source

Arxiv paper 2512.12675

Freshness

Last updated January 2026.