Poplar: A Scalable Pipeline for Human-Centric Image Dataset Synthesis
Poplar turns human-image generation into an auditable dataset pipeline, not a one-off prompt result.
The paper describes a Specify-Render-Inspect workflow for synthesizing human-centric image datasets with structured attributes, commonsense constraints, and photography-oriented prompts. Its inspection step uses a structured vision-language review to reject image defects or material prompt mismatches while keeping the original prompt record. The author reports Poplar-9K, with 9,401 curated image-text pairs retained from 11,765 reviewed candidates, a 79.9% acceptance rate. The release includes the dataset, pipeline, configurations, immutable prompts, and inspection records. HF Daily Papers' note
The paper describes a Specify-Render-Inspect workflow for synthesizing human-centric image datasets with structured attributes, commonsense constraints, and photography-oriented prompts. Its inspection step uses a structured vision-language review to reject image defects or material prompt mismatches while keeping the original prompt record. The author reports Poplar-9K, with 9,401 curated image-text pairs retained from 11,765 reviewed candidates, a 79.9% acceptance rate. The release includes the dataset, pipeline, configurations, immutable prompts, and inspection records. HF Daily Papers' note
score 4