Multi-Reference Image Generation Benchmark with Structured Long-Context Data

MACRO is a large-scale benchmark and training dataset for multi-reference image generation tasks. It contains training splits totaling at least 100,000 samples across four task categories and four image-count brackets, plus a curated evaluation set. The dataset was created by Azily and was last updated in March 2026.

Use Cases

• Train models for text-to-image generation using structured long-context prompts from the 'Customization' task category.
• Benchmark model performance on image-to-image generation tasks across different image-count brackets (e.g., 1-3, 4-5, 6-7, ≥8 reference images).
• Evaluate multi-reference image synthesis capabilities using the curated evaluation set of 250 samples per task category.
• Develop and test models for tasks requiring conditioning on multiple image references, leveraging the dataset's structured context data.

Strengths

• At least 100,000 training samples across defined task categories.
• Structured data covering four distinct image-count brackets for multi-reference tasks.
• Includes a curated evaluation benchmark with 250 samples per task category.

Limitations

• Exact total row count, column names, and sample data details are unavailable.
• Dataset size, file formats, and specific license terms are unknown.
• Potential for task-specific class or distribution imbalances is not detailed.

Provenance

Source

Hugging Face dataset repository by author Azily.

Collection Method

Constructed as a benchmark and training dataset; specific collection/generation method unknown.

Freshness

Last updated March 2026.