MTADataset is a large-scale dataset designed for image inpainting. It contains images processed with Grounded-SAM to extract labels, bounding boxes, and masks, and uses LLaVA to generate detailed descriptions for approximately 5 masks per image. The dataset was created by huangjun12 and was last updated on October 23, 2025.
Use Cases
- Train image inpainting models based on the provided image-mask pairs.
- Develop or fine-tune vision-language models using the detailed mask descriptions generated by LLaVA.
- Benchmark object-aware image generation algorithms based on the extracted labels and bounding boxes.
- Research on mask-text alignment for conditional image synthesis.
Strengths
- Designed as a large-scale dataset for image inpainting.
- Includes approximately 5 detailed descriptions per image for its masks, generated by LLaVA.
- Provides multiple annotation types per image: labels, bounding boxes, and masks extracted via Grounded-SAM.
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- huggingface
- Collection Method
- Images processed with Grounded-SAM for annotations and LLaVA for descriptions.
- Freshness
- Last updated 2025-10-23 11:23:41; freshness should be verified.