Source-linked AI summary
Poplar: A Scalable Pipeline for Human-Centric Image Dataset Synthesis
Zhishan Zou
TL;DR
Human-centric dataset synthesis must jointly achieve semantic diversity, plausible attribute combinations, visual fidelity, and image–text consistency. Poplar addresses this with a configurable specification, rendering, and inspection pipeline, producing Poplar-9K by retaining 9,401 images from 11,765 reviewed candidates.
Problem
Human-centric dataset synthesis must jointly capture semantic diversity, realistic attribute co-occurrence, visual fidelity, and image–text consistency at the collection level.
Method
Poplar connects structured, commonsense-constrained attribute sampling and photography-oriented prompts with diffusion rendering and vision–language quality inspection.
Results
9,401 curated human-centric image–text pairs were retained from 11,765 reviewed candidates to form the open-source Poplar-9K dataset.
Takeaways & Limitations
Poplar provides a reusable collection-building process and a compact dataset exposing generated, rejected, and recorded outputs.
Takeaways & Limitations
Poplar-9K is a modest-scale instantiation with one configurable sampling policy and does not claim to reproduce the full distribution of human photography.
Abstract
from arXiv · showhide
Recent image generators can synthesize convincing human-centric images, yet producing a useful collection remains different from producing a single successful image. A human-centric dataset must cover varied people and contexts, avoid implausible attribute combinations, preserve an everyday photographic character, and expose quality-control decisions at scale. We present Poplar, a reproducible Specify--Render--Inspect pipeline for human-centric image dataset synthesis. Specify samples structured attributes under commonsense constraints and verbalizes them as photography-oriented prompts. Render uses a realism-adapted image generator across composition-aware aspect ratios and retries obvious technical failures. Inspect applies a single structured vision--language review to each candidate, preserving the original prompt while rejecting intrinsic image defects or material prompt mismatches. Using Poplar, we construct Poplar-9K: 9,401 curated human-centric image--text pairs retained from 11,765 reviewed candidates (79.9\% acceptance). We release the dataset together with the pipeline, configurations, immutable generation prompts, and auditable inspection records as a compact resource for building customizable human-centric collections.
1 Introduction
Poplar addresses the gap between generating realistic individual images and synthesizing human-centric datasets with collection-level semantic coverage, structural coherence, and quality control. It unifies constrained prompt generation, image synthesis, and vision–language inspection, producing the released Poplar-9K dataset.
- Motivation: Dataset synthesis must capture a target domain’s semantic distribution and structural characteristics, not merely realistic individual samples.The work focuses on human-centric photography, where people, activities, interactions, and contexts are the primary visual content.
- Motivation: Existing pipelines often prioritize individual-generation quality or downstream utility over collection-level coverage and everyday human-photography structure.Manual prompts are costly to scale, while independent attribute sampling can produce implausible descriptions and distorted co-occurrences.
- Motivation: Errors can propagate from semantic specification through synthesis, while defects and instruction-following failures accumulate at collection scale.The paper argues that semantic coverage, commonsense consistency, rendering, and quality control must be coordinated rather than handled as isolated prompt decisions.
- Poplar pipeline: Poplar unifies knowledge-aware prompt generation, diffusion-based image synthesis, and vision–language quality inspection for scalable human-centric dataset synthesis.Its prompt producer samples structured attributes under commonsense constraints and converts them into natural photographic descriptions.
- Released resource: 9,401 curated human-centric image–text pairs were retained from 11,765 reviewed candidates in the released Poplar-9K dataset.The dataset covers diverse subjects, activities, environments, camera viewpoints, and photographic styles, and the pipeline and dataset are open-sourced.
2 Related Work
Prior work has advanced photorealistic synthesis and synthetic human-centric data through powerful generators, procedural control, and foundation models. However, constructing everyday human-photography datasets additionally requires semantic coverage, plausible co-occurrence, and scalable quality inspection, motivating Poplar’s integrated pipeline perspective.
- Photorealistic Image Synthesis: Photorealistic synthesis progressed from GANs and diffusion models to large text-conditioned generators, establishing a mature foundation for convincing individual images.StyleGAN, diffusion probabilistic models, Imagen, and latent diffusion mark this progression.
- Dataset-Level Challenges: Yet image-generation advances do not determine dataset composition, semantic factor co-occurrence, or scalable control of generation failures.The passage distinguishes producing convincing individual images from deciding which images populate a dataset and how failures are controlled.
- Synthetic Dataset Construction: Procedural engines, simulators, and physically based rendering provide scalable scene control and labels, while foundation models offer flexible semantic specification and diffusion-based rendering.Examples span Kubric, PSP-HDRI+, Gen4D, StableRep, SynthCLIP, Real-Fake, DistDiff, DatasetDM, and DiffuMask.
- Human-Centric Synthesis: Human-centric synthesis has shifted from graphics-based control toward pretrained models, but everyday photography requires long-tail variation and dependencies across people, activities, clothing, interactions, environments, viewpoints, compositions, and styles.Graphics-based resources offer geometric control but remain tied to body assets, motion libraries, and rendering environments.
- Poplar’s Perspective: Poplar addresses this dataset-level perspective by connecting semantic specification, image generation, and quality inspection within one configurable pipeline.This frames the work as an integrated approach to human-centric photography dataset construction.
3 Poplar
Poplar is a reproducible Specify–Render–Inspect pipeline that converts structured, commonsense-valid dataset specifications into auditable human-centric image–text pairs. It separates semantic selection and photography-oriented verbalization from replaceable rendering and structured quality inspection.
- Inspect: Inspect consumes exactly one image–prompt pair and returns a structured keep/reject decision with visible evidence, making filtering configurable and auditable without a single overall quality score.The pipeline preserves the original prompt and returns structured inspection records alongside curated outputs.
- Specify: Photography-oriented prompts combine semantic content with capture conditions to encourage casual, everyday compositions rather than polished, staged portraits.The capture layer specifies photographer perspective, camera or phone characteristics, viewpoint, framing, lighting, and plausible imperfections.
- Specify: Specify samples configurable attribute spaces and applies commonsense compatibility constraints before generating immutable prompts and structured metadata.Attributes span identity, clothing, action, scene, viewpoint, lighting, composition, and photographic style; constraints address feasibility, consistency, and distinctness in multi-person scenes.
- Specify: Poplar separates semantic selection from verbalization, using Qwen3.5-27B-FP8 to turn validated configurations into coherent photography-oriented prompts while preserving selected fields and metadata.The stored prompt combines a natural-language condition with the structured record, allowing the same specification to be rendered by different image generators.
- Render: Render treats a pretrained image generator as a replaceable backend and samples one composition-aware aspect ratio per specification from five supported formats.The implementation uses Krea 2 Turbo with the Krea2-realism-V2 adapter, while the same prompts can condition other generators.
4 Poplar-9K
Poplar-9K demonstrates Poplar as a dataset-level construction process that samples semantic configurations, renders candidates at scale, and applies technical and semantic quality control. The resulting curated collection provides inspectable human-centric image–text pairs alongside prompts, structured attributes, code, and construction records.
- Construction and curation: The Quality Inspector removes technical failures, prompt mismatches, synthetic-looking outputs, artifacts, implausible relationships, anatomy errors, and sensitive content.Candidates are rendered with photography-oriented prompts and structured attributes at one of five composition-aware aspect ratios, after preliminary retries for grayscale, severe blur, or invalid files.
- Construction and curation: Poplar-9K retains 9,401 images from 11,765 reviewed candidates, rejecting 2,364 for a 79.9% acceptance rate.Every retained image is paired with its unchanged generation prompt and structured semantic attributes.
- Dataset composition: The 9,401 images depict 10,877 people, with coverage spanning younger adults through subjects in their 70s and 13 cultural-background descriptors representing approximately 6.9%–8.5% each.These statistics come from structured prompt attributes and describe synthesis specifications rather than identities inferred from generated images.
- Dataset composition: Clothing spans eight listed categories, led by casual clothing at 18.8%, trendy feminine styles at 15.5%, and homewear at 12.3%.Scene frequencies have a long tail, with the most frequent listed setting accounting for only 3.8% of the collection.
- Release and provenance: The release includes the 9,401 curated images, prompts, structured attributes, complete Poplar codebase, executable configuration, and provenance records for retained and rejected candidates.Provenance covers immutable prompts, attributes, rendering and filtering history, inspection records, and rejected-item issue evidence, making the 11,765-to-9,401 funnel inspectable.
5 Scope and Limitations
Poplar is a configurable dataset-construction pipeline with intentionally modest scope, not a new generator, language model, or learned quality metric. Its synthetic data and fallible inspection require careful interpretation, auditing, and caution in identity, demographic, and high-stakes applications.
- Scope: Poplar constructs datasets rather than introducing a new image generator, language model, or learned quality metric.Its intentionally modest scale demonstrates one configurable sampling policy instead of reproducing the full distribution of human photography.
- Scope: Requested age and cultural-background attributes are not verified demographic labels or inferred identities.The compatibility rules reduce obvious contradictions but do not establish real-world demographic identities.
- Inspection limitations: Automated inspection can miss subtle artifacts or reject acceptable images, while deterministic pixel checks cover only narrow technical failures.Released inspection evidence and rejected-item metadata allow users to audit or replace the inspection policy.
- Use limitations: Generated people are synthetic and should not be treated as records of real individuals.Applications involving identity, demographic measurement, or high-stakes decisions require caution.
6 Conclusion
Poplar frames human-centric image dataset synthesis as a collection-level problem requiring joint control of diversity, plausible attributes, visual fidelity, and image–text consistency. Its configurable Specify–Render–Inspect pipeline produces the open-source Poplar-9K dataset with auditable generation and curation records.
- Conclusion: Poplar treats semantic diversity, realistic attribute co-occurrence, visual fidelity, and image–text consistency as joint dataset-construction goals.The work distinguishes collection-level synthesis from generating isolated successful images.
- Conclusion: Poplar connects structured attribute sampling, photography-oriented prompting, diffusion-based rendering across formats, and vision–language inspection.Its Prompt Producer controls composition, Image Producer renders specifications, and Quality Inspector rejects technically valid but semantically inconsistent images.
- Conclusion: 11,765 candidate images were reviewed and 9,401 retained, forming the open-source Poplar-9K dataset.The dataset is released with prompts, structured attributes, rendering provenance, and inspection records as a reusable resource for growing human-centric collections.