ConsistCompose3M: 3 Million Samples for Multimodal Layout-Controlled Image Composition

ConsistCompose3M provides approximately 3 million samples for layout-controllable multi-instance image composition. The dataset, created by sensenova, offers structured spatial-semantic supervision and includes identity-preserving samples filtered by CLIP/DINO similarity. It was last updated on March 31, 2026.

Use Cases

• Train models for layout-controllable image generation based on the dataset's structured spatial-semantic supervision.
• Develop identity-preserving image composition systems based on the CLIP/DINO similarity-filtered samples.
• Benchmark multi-instance scene generation models based on the dataset's scale and diversity of scenes.
• Research unified multimodal layout control based on the dataset's dedicated focus on this task.

Strengths

• Approximately 3 million samples, indicating a large-scale resource.
• Includes identity-preserving samples filtered by CLIP/DINO similarity, suggesting quality control.
• Designed for multi-instance scenes, implying diversity in scene composition.

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

sensenova via Hugging Face

Collection Method

Likely compiled and filtered from existing image-text sources, with identity-preserving samples filtered by CLIP/DINO similarity.

Freshness

Last updated 2026-03-31 06:39:03; freshness should be verified.