Reproduce “Poplar: A Scalable Pipeline for Human-Centric Image Dataset Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhishan Zou
Poplar constructs human-photography datasets by specifying structured people, scenes, and capture conditions, rendering images with a pretrained generator, and inspecting each image against its unchanged prompt. Its photography-oriented prompts produce qualitatively more casual examples than content-centric prompts across three generators. Poplar-9K retained 9,401 of 11,765 reviewed candidates, but this funnel does not validate inspector accuracy or full photographic coverage.
Key insights
- Separating content planning from linguistic realization in prompt generation, combined with commonsense compatibility constraints, effectively reduces implausible attribute combinations in synthetic datasets.
- Integrating a preliminary, image-only visual filter prior to resource-intensive vision-language inspection significantly improves the efficiency of the synthetic data pipeline by identifying technical failures early.
- A structured, auditable pipeline with detailed provenance for both accepted and rejected samples is crucial for transparency, reproducibility, and understanding biases in synthetic data generation processes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.