Instruction-Based Image Editing Dataset for Unified Multimodal Models

ScaleEdit-12M is a large-scale dataset for instruction-based image editing, created by InternVL-U to address challenges in constructing diverse, high-quality data without proprietary APIs. It was last updated on March 24, 2026. The dataset aims to overcome limitations of previous datasets that relied on closed-source models or fixed synthetic pipelines.

Use Cases

• Train unified multimodal models on instruction-based image editing tasks using the dataset's editing instructions and corresponding image pairs.
• Benchmark the generalizability of image editing models against the dataset's diverse editing scenarios.
• Develop cost-effective scaling methods for image editing datasets, as referenced in the dataset's abstract addressing previous reliance on closed-source models.

Strengths

• Dataset is designed for large-scale applications, as indicated by its name 'ScaleEdit-12M' suggesting a target of 12 million entries.
• Created to address specific research challenges in dataset construction, focusing on diversity and quality without proprietary APIs.
• Associated with research papers, as indicated by the 'Arxiv260309877' and 'Arxiv260320644' tags.

Limitations

• Specific details on dataset size, columns, sample data, and file formats are unavailable from the provided input.
• The actual scale, diversity, and quality of the image-editing pairs cannot be verified from the given description.
• License information is not explicitly confirmed, only suggested by a tag.

Provenance

Source

Hugging Face platform, uploaded by author InternVL-U.

Collection Method

Constructed to create a large-scale, diverse, and high-quality image editing dataset without costly proprietary APIs, as per the abstract.

Freshness

Last updated on 2026-03-24.