Pith. sign in

REVIEW 4 major objections 5 minor 32 references

A three-stage Specify–Render–Inspect pipeline turns diffusion generators into auditable, curated human-centric datasets, retaining 9,401 of 11,765 candidates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Poplar-9K is a curated human-centric image dataset generated through a reproducible attribute sampling, diffusion rendering, and vision-language inspection pipeline, with 9,401 accepted pairs from 11,765 candidates.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A well-documented, reproducible pipeline for synthetic human-centric datasets with a strong audit trail; the curation quality rests on one unvalidated VLM, but that is fixable and does not sink the contribution. the 4 major comments →

arxiv 2608.00440 v1 pith:EETO4EA3 submitted 2026-08-01 cs.CV

Poplar: A Scalable Pipeline for Human-Centric Image Dataset Synthesis

Zhishan Zou This is my paper
classification cs.CV
keywords human-centric image datasetssynthetic data pipelinetext-to-image diffusionvision-language quality inspectionprompt engineeringdataset curationimage-text pairsreproducibility
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that building a useful synthetic image dataset is a collection-level engineering problem, not a matter of making each image look good. It proposes Poplar, a configurable three-stage pipeline—Specify, Render, Inspect—that samples structured human attributes under commonsense constraints, turns them into photography-oriented prompts, renders candidates with a diffusion model, and filters them with one vision-language review. The central demonstration is Poplar-9K, a curated set of 9,401 human-centric image-text pairs retained from 11,765 candidates (79.9% acceptance), with every prompt kept immutable and every rejection recorded with visible evidence. If the pipeline works as claimed, it offers a reproducible way to grow custom human-centric datasets whose composition and quality-control decisions are inspectable rather than opaque.

Core claim

Poplar's working hypothesis is that dataset synthesis succeeds only when semantic planning, image rendering, and quality control are coupled through explicit contracts. The Specify stage represents each sample as a structured attribute vector sampled from configurable distributions and checked against commonsense compatibility rules, then verbalized by an LLM into a photography-oriented prompt that includes capture conditions such as viewpoint, flash, and framing rather than just content. The Render stage treats a pretrained generator as a replaceable backend, samples composition-aware aspect ratios, and applies a conservative deterministic prefilter for blur, grayscale, and decoding failure

What carries the argument

The central mechanism is the Specify–Render–Inspect contract: a structured attribute vector a sampled under commonsense constraints C(a), verbalized into an immutable photography-oriented prompt p = f_LLM(a, t_photo, r), rendered by a replaceable diffusion backend across composition-aware aspect ratios, and finally judged by a single vision-language inspector V(x, p) that returns binary defect flags, prompt-mismatch severities, visible-person descriptions, and evidence. A candidate is admitted only when no intrinsic defect is flagged and no critical or major mismatch is reported.

Load-bearing premise

Everything hangs on one vision-language model pass being able to reliably tell a defective or prompt-mismatched image from a good one; the paper reports no human agreement study, external benchmark, or downstream task validation to verify that the inspector is correct.

What would settle it

Manually review a random sample of 500 accepted and 500 rejected candidates from Poplar-9K with human raters using the same rubric. If raters disagree with the inspector on a substantial share—for example, accepting visibly stitched or duplicated subjects, or rejecting clearly aligned images—the 79.9% acceptance rate and the audit records stop supporting the curation claim. Alternatively, compare a downstream model trained on Poplar-9K versus the same number of unfiltered candidates: no improvement would suggest the inspection stage is not doing the claimed filtering work.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Collection composition is directly controllable: adjusting the attribute spaces, proposal distributions, and compatibility rules changes what the dataset covers, so builders can target coverage rather than accept an LLM's implicit preferences.
  • Because prompts survive inspection unchanged, the same semantic specifications can be re-rendered with different image generators without redoing the planning stage.
  • Resumable JSONL records and deterministic decoding make each run auditable and reproducible; interrupted runs continue without silently replacing samples.
  • The fixed funnel from 11,765 reviewed to 9,401 retained gives downstream users a concrete construction record rather than just a pile of 'good' images.
  • The pipeline is not tied to one generator: replacing the rendering backend only swaps the image producer while keeping prompt generation and inspection intact.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural next step the paper does not take is measuring downstream value: training a model on Poplar-9K versus an unfiltered random subset would test whether the inspector's filtering actually improves learning outcomes.
  • The rejection codes and evidence fields could be aggregated to map where current generators systematically fail on human attributes such as faces, clothing, and multi-person composition, turning the audit trail into a diagnostic instrument.
  • The same Specify–Render–Inspect skeleton should transfer to other visual domains—indoor scenes, products, documents—provided the attribute spaces and inspection rubric are redesigned; the paper keeps the components replaceable but only demonstrates human-centric photography.
  • Without a human agreement study, the 79.9% acceptance rate should be read as the inspector's own verdict on its output, not as an independent measure of image quality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents Poplar, a three-stage Specify–Render–Inspect pipeline for synthesizing human-centric image–text datasets. The Specify stage samples structured attributes under hand-authored commonsense constraints and verbalizes them into photography-oriented prompts via an LLM; the Render stage uses a commercial diffusion backbone with composition-aware aspect-ratio sampling and a conservative deterministic pre-filter; the Inspect stage applies a single Qwen3.5-27B-FP8 vision-language pass to each candidate, rejecting images with intrinsic defects or critical/major prompt mismatches (Eq. 5–6). The authors construct Poplar-9K, reporting 9,401 retained pairs from 11,765 reviewed candidates (79.9% acceptance), and release the dataset, pipeline code, configurations, prompts, and JSONL audit records. The paper frames the contribution as a reproducible, configurable construction process with explicit provenance, rather than as a new generative model.

Significance. If the pipeline performs as claimed, Poplar is a useful template for building human-centric synthetic datasets with auditable quality control. The explicit separation of semantic planning, prompt verbalization, and visual inspection, together with immutable prompt hashing, deterministic seeds, and per-item inspection records, is a genuinely reproducible and inspectable design. The release of both dataset and code with a fixed configuration (Section 4.4) is a concrete asset for the community. However, the central empirical claim—that the 79.9% acceptance funnel produces a 'curated' dataset with high visual fidelity and reliable image–text consistency—rests entirely on an unvalidated automatic inspector and on qualitative illustrations. No human agreement study, independent inspector, external benchmark, or downstream task evaluation is reported. The paper's own Section 5 concedes that automated inspection 'remains fallible' and that the compatibility rules encode design choices. Thus the significance of the contribution is conditional on additional validation of the quality gate and of the dataset's utility.

major comments (4)
  1. [§3.3, Eq. (5)–(6); §4.1] The curation claim—9,401 retained, 2,364 rejected, 79.9% acceptance—is entirely determined by a single forward pass of Qwen3.5-27B-FP8 applying the rubric in Eq. (5)–(6). No human agreement study, no second independent VLM, and no external benchmark validates these keep/reject decisions. A consistently biased inspector would still produce clean audit records, so the released 'evidence' makes the funnel inspectable but not accurate. I recommend adding a quantitative validation of the inspector: e.g., a random sample of 200–300 accepted and rejected candidates labeled by human annotators, with reported precision/recall, confusion matrix, and agreement statistics (Cohen's kappa). Without this, the 'curated' characterization is not established.
  2. [§3.1.3 + §3.3] The same model family (Qwen3.5-27B-FP8) is used both as the prompt verbalizer and as the quality inspector. Since the inspector evaluates fidelity to a prompt it effectively helped write, self-consistency may be inflated; the reported mismatch rate does not measure alignment with the structured attributes independently. The paper should address this circularity explicitly. A practical fix is to re-inspect a subset with a different model family (or a human study, as above) and report the disagreement rate, or at minimum to ablate the inspector by replacing the verbalizer with a deterministic template and measuring the change in rejection statistics.
  3. [§4, §5] The paper provides no downstream evaluation of Poplar-9K. For a dataset paper, the claimed usefulness of the collection is unsubstantiated without at least one concrete use: e.g., fine-tuning or zero-shot evaluation of image-text retrieval, representation learning (CLIP-style), human-centric perception tasks, or a comparison against existing synthetic or real datasets (SynthCLIP, StableRep, PSP-HDRI+). The four design properties in the Introduction (semantic diversity, realistic co-occurrence, visual fidelity, image–text consistency) are asserted qualitatively; statistics in Figure 6 describe the sampling specification, not the rendered images. Adding a task-based evaluation would demonstrate that the curation funnel improves downstream utility relative to unfiltered generations.
  4. [§3.1.2, Eq. (2)] The commonsense constraint evaluation is entirely qualitative. The paper does not report how many sampled configurations were rejected or repaired by the compatibility rules, nor does it validate that these rules reduce implausible combinations in the final dataset. Given that 'realistic attribute co-occurrence' is a core stated property, quantitative evidence (e.g., violation rates in the released metadata before/after constraint enforcement, or human judgments of prompt plausibility on a sample) would strengthen the claim. This is less central than the inspector validation, but it supports the Specify-stage contribution.
minor comments (5)
  1. [§3.1.2, Eq. (2)] Notation 'K^ k=1' is ambiguous; it should be written as a big conjunction (e.g., ⋀_{k=1}^K c_k(a)).
  2. [Figure 3] The qualitative comparison of content-centric vs. photography-oriented prompts is illustrative but not quantified; consider reporting a simple metric (e.g., proportion of images judged 'casual snapshot' by human raters).
  3. [§4.2] The cultural-background and age statistics are computed from prompt attributes and are explicitly not verified image labels. This caveat is stated, but it should also appear in the data card/caption for Figure 6 to avoid reader misinterpretation.
  4. [§4.1] The preliminary filter's retry behavior is described, but no statistics are given (e.g., how many candidates required retries, how many were discarded before the Inspector stage). Adding these numbers would make the funnel fully auditable as advertised.
  5. [§5] The limitation statement is honest and comprehensive, but it currently functions as a disclaimer rather than an evaluation agenda. I suggest moving some of these limitations into explicit future-work validation plans (human study, downstream tasks, external inspector).

Circularity Check

0 steps flagged

No significant circularity found; Poplar's curation result is an explicitly defined operational outcome, not a derived prediction or fitted parameter.

full rationale

Poplar does not claim to derive an external quantity or predict a benchmark. The reported funnel (11,765 reviewed, 9,401 retained, 79.9% acceptance) is defined by the pipeline's own acceptance rule in Eq. (6): a candidate is admitted only when no intrinsic defect flag is set and no critical/major prompt mismatch is reported. This is a transparent operational definition, not a hidden circular reduction. The Quality Inspector uses Qwen3.5-27B-FP8, the same model family employed as the prompt verbalizer in Section 3.1.3, so prompt–image agreement could be inflated by model self-consistency; however, the paper does not rely on that self-consistency as proof of correctness. Section 5 explicitly states that automated inspection remains fallible and that the released inspection evidence and rejected-item metadata allow users to audit or replace the policy. No parameter is fitted to a subset and then re-predicted; no uniqueness theorem is imported from the authors' prior work; no known result is renamed; and there are no load-bearing self-citations. The only self-referential element is that 'curated' means 'passed the Inspector,' but the paper states this criterion and its limitations explicitly, so the claim reduces to its own definition only in the non-circular sense of an explicit pipeline contract. Thus the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim depends on configurable design choices and on the reliability of two third-party model families. None of the free parameters are fitted to data in the classic sense, but several are hand-set thresholds and distributions that directly shape the reported acceptance rate and dataset composition.

free parameters (5)
  • preliminary filter thresholds = not specified numerically
    Chroma, neutral-pixel fraction, and edge-variance thresholds are chosen by hand and described only as conservative (Section 3.2).
  • aspect ratio distribution q_rho = configurable, not enumerated
    The distribution over the five aspect ratios is conditioned on composition-relevant attributes and chosen by the dataset builder (Section 3.2, Eq. 4).
  • inspection severity rule = critical/major rejection
    The acceptance rule in Eq. 6 rejects any image with an intrinsic defect or a critical/major mismatch; this severity mapping is a design choice.
  • attribute sampling distributions = e.g., 13 cultural-background descriptors, 8 clothing groups
    Each attribute dimension is sampled from an explicit proposal distribution chosen by the authors to shape marginal coverage (Section 3.1.2).
  • rendering parameters = scale 1.5, 8 steps, zero guidance, max side 1024
    These are fixed configuration choices for the Krea 2 Turbo backend, reported in Section 4.4.
axioms (5)
  • domain assumption Qwen3.5-27B-FP8 can verbalize structured attributes into coherent photography prompts without altering the planned content.
    Invoked in Section 3.1.3, where the LLM acts as a verbalizer with consistency rules.
  • domain assumption Krea 2 Turbo with the Krea2-realism-V2 adapter renders prompts as everyday photographic images with sufficient fidelity.
    Invoked in Section 3.2; no quantitative evaluation of rendering fidelity is provided.
  • domain assumption The vision-language inspector's keep/reject decisions are accurate enough for dataset curation.
    Invoked in Section 3.3 and formalized in Eqs. 5 and 6; no human validation or inter-rater agreement is reported.
  • domain assumption The commonsense compatibility constraints C(a) capture enough contradictions to ensure realistic attribute co-occurrence.
    Invoked in Section 3.1.2, Eq. 2; the rule set is described qualitatively and may encode design choices.
  • domain assumption Multiple aspect ratios conditioned on composition-relevant attributes improve compositional diversity.
    Invoked in Section 3.2, Eq. 4; no compositional diversity metric is reported.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Poplar: A Scalable Pipeline for Human-Centric Image Dataset Synthesis." pith.science (2026). https://pith.science/paper/EETO4EA3

@misc{pith2026260800440,
  author       = {Pith},
  title        = {Pith review of: Poplar: A Scalable Pipeline for Human-Centric Image Dataset Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EETO4EA3}},
  note         = {Machine review of arXiv:2608.00440}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

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.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 21 canonical work pages · 1 internal anchor

  1. [1]

    Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang

    Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726–8737, 2023

    work page 2023

  2. [2]

    Promptify: Text-to-image generation through interactive prompt exploration with large language models

    Stephen Brade, Bryan Wang, Mauricio Sousa, Sageev Oore, and Tovi Grossman. Promptify: Text-to-image generation through interactive prompt exploration with large language models. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages Article 96, 1–14, 2023. doi: 10.1145/3586183. 3606725

    doi:10.1145/3586183 2023

  3. [3]

    Jerrin Bright, Zhibo Wang, Yuhao Chen, Sirisha Rambhatla, John Zelek, and David A. Clausi. Gen4D: Synthesizing humans and scenes in the wild. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2025

    work page 2025

  4. [4]

    PSP-HDRI$+$: A Synthetic Dataset Generator for Pre-Training of Human-Centric Computer Vision Models

    Salehe Erfanian Ebadi, Saurav Dhakad, Sanjay Vishwakarma, Chunpu Wang, You-Cyuan Jhang, Maciek Chociej, Adam Crespi, Alex Thaman, and Sujoy Ganguly. PSP-HDRI+: A synthetic dataset generator for pre-training of human-centric computer vision models.arXiv preprint arXiv:2207.05025, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  5. [5]

    Datasheets for datasets.Communications of the ACM, 64(12):86–92, 2021

    Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. Datasheets for datasets.Communications of the ACM, 64(12):86–92, 2021. doi: 10.1145/ 3458723

    work page 2021

  6. [6]

    Geneval: An object-focused framework for evaluating text-to-image alignment

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. InAdvances in Neural Information Processing Systems, volume 36, 2023

    work page 2023

  7. [7]

    Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al

    Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    work page 2022

  8. [8]

    Synthclip: Are we ready for a fully synthetic clip training?arXiv preprint arXiv:2402.01832, 2024

    Hasan Abed Al Kader Hammoud, Hani Itani, Fabio Pizzati, Philip Torr, Adel Bibi, and Bernard Ghanem. Synthclip: Are we ready for a fully synthetic clip training?arXiv preprint arXiv:2402.01832, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Jain, and Pieter Abbeel

    Jonathan Ho, Ajay N. Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems, volume 33, 2020

    work page 2020

  10. [10]

    Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A. Smith. TIFA: Accurate and interpretable text-to-image faithfulness evaluation with question answering.arXiv preprint arXiv:2303.11897, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation.arXiv preprint arXiv:2307.06350, 2023

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation.arXiv preprint arXiv:2307.06350, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019

    work page 2019

  13. [13]

    Krea 2 Turbo

    Krea.ai. Krea 2 Turbo. Hugging Face model card, 2026. URLhttps://huggingface.co/krea/Krea-2-Turbo

    work page 2026

  14. [14]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations, 2023

    2023

  15. [15]

    Huang, Joachim Tesch, David T

    Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13468–13478, 2021

    work page 2021

  16. [16]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023

    2023

  17. [17]

    SDXL: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Data cards: Purposeful and transparent dataset documentation for responsible AI

    Mahima Pushkarna, Andrew Zaldivar, and Oddur Kjartansson. Data cards: Purposeful and transparent dataset documentation for responsible AI. InProceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1776–1826, 2022. doi: 10.1145/3531146.3533231. 11

    work page arXiv 2022

  19. [19]

    Qwen3.5-27B-FP8

    Qwen Team. Qwen3.5-27B-FP8. Hugging Face model card, 2026. URLhttps://huggingface.co/Qwen/Qwen3. 5-27B-FP8

    work page 2026

  20. [20]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    work page 2022

  21. [21]

    Krea2-realism-V2

    RudySen. Krea2-realism-V2. Hugging Face model card, 2026. URL https://huggingface.co/RudySen/ Krea2-realism-V2

    work page 2026

  22. [22]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  23. [23]

    Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. InAdvances in Neural Information Pro...

    work page 2022

  24. [24]

    Stablerep: Synthetic images from text-to-image models make strong visual representation learners

    Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. InAdvances in Neural Information Processing Systems, volume 36, 2023

    work page 2023

  25. [25]

    Black, Ivan Laptev, and Cordelia Schmid

    Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 109–117, 2017

    work page 2017

  26. [26]

    Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau

    Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 893–911, 2023

    work page 2023

  27. [27]

    Datasetdm: Synthesizing data with perception annotations using diffusion models

    Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annotations using diffusion models. InAdvances in Neural Information Processing Systems, volume 36, 2023

    work page 2023

  28. [28]

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1206–1217, 2023

    2023

  29. [29]

    LLaVA-Critic: Learning to evaluate multimodal models.arXiv preprint arXiv:2410.02712, 2024

    Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. LLaVA-Critic: Learning to evaluate multimodal models.arXiv preprint arXiv:2410.02712, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Real-fake: Effective training data synthesis through distribution matching

    Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through distribution matching. InInternational Conference on Learning Representations, 2024

    work page 2024

  31. [31]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, 2023

    2023

  32. [32]

    Distribution-aware data expansion with diffusion models

    Haowei Zhu, Ling Yang, Jun-Hai Yong, Hongzhi Yin, Jiawei Jiang, Meng Xiao, Wentao Zhang, and Bin Wang. Distribution-aware data expansion with diffusion models. InAdvances in Neural Information Processing Systems, volume 37, 2024. doi: 10.52202/079017-3264. 12

    work page doi:10.52202/079017-3264 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.