by InclusionAI
ConceptEdit-12M is a large image editing dataset from InclusionAI. Each sample pairs a source image with an edited image and JSON metadata.
DatasetFree
Step 1: Download the dataset from Hugging Face.
Step 2: Pick a split, such as the enhanced prompt square resolution one.
Step 3: Extract one shard with tar to inspect it, or loop over all shards in a split.
Step 4: Read each sample's JSON for the instruction, edit concept and image paths.
Step 5: Filter samples using the evaluation fields such as the keep flag and VQA score.
A source image, an edited image and a JSON file. The JSON has the instruction, edit concept, image paths, evaluation data and a caption.
Four: enhanced or original prompts, each in square or random resolution.
Each sample has short and detailed instructions in English and Chinese.
The card says they are based on images from Fine-T2I.