MuLAn: 44,000 Multi-Layer RGBA Image Decompositions

MuLAn provides over 44,000 multi-layer RGBA decompositions and 100,000 instance images derived from COCO and LAION sources. Released by the mulan-dataset team in late 2024, it serves as a photorealistic resource for instance-wise image manipulation and generation.

Use Cases

• Controllable text-to-image generation using RGBA instance-wise layers
• Layer-aware image editing by manipulating individual instance images
• Training models for automated image decomposition into constituent layers

Strengths

• 44,000 multi-layer annotations
• 100,000 individual instance images
• RGBA format providing native transparency for instance-wise layers
• Dual-source composition from COCO and LAION for stylistic variety

Limitations

• Restricted to non-commercial use cases
• Lack of explicit column-level metadata documentation in the initial release

Provenance

Source

mulan-dataset via Hugging Face

Collection Method

Annotated decomposition of existing images from COCO and LAION datasets

Freshness

Last updated in December 2024.