pith. sign in

arxiv: 2605.25759 · v1 · pith:I5XBKMIUnew · submitted 2026-05-25 · 💻 cs.CV

Towards Anatomically Plausible Human Image Generation via Synthetic Localized Preferences

Bao Li , Yuliang Xiu , Zhen Liu This is my paper

Pith reviewed 2026-06-29 22:52 UTC · model grok-4.3

classification 💻 cs.CV
keywords anatomical fidelitytext-to-image generationpreference optimizationhuman image synthesislocalized degradationsynthetic preferencesHAF-BenchHAP dataset
0
0 comments X

The pith

A localized degradation method creates synthetic preference pairs that align text-to-image models for better human anatomy.

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

Text-to-image models generate realistic images but often fail at correct human anatomy due to ambiguous training signals from limited datasets. The paper proposes ASAP to build over 10,000 preference pairs by applying controlled anatomical errors to specific regions of high-quality images while leaving other content intact. It then uses a localized, margin-bounded version of direct preference optimization to focus learning on those error-prone areas and prevent over-optimization. Experiments show this reduces anatomical mistakes across several models while keeping overall image quality intact, supported by a new evaluation benchmark called HAF-Bench.

Core claim

The framework of Alignment via Synthetic Anatomical Preference constructs controlled preference pairs through a localized degradation mechanism that introduces explicit anatomical errors in targeted regions of high-fidelity human images. These pairs enable a localized and margin-bounded variant of DPO to prioritize optimization in anatomical regions, resulting in reduced anatomical errors across foundation models as evaluated on the HAF-Bench using the HAP dataset.

What carries the argument

The localized degradation mechanism that performs a controlled experiment by introducing explicit anatomical errors in targeted regions while preserving the remaining content.

If this is right

  • ASAP reduces anatomical errors consistently across multiple foundation models.
  • The method maintains overall image quality during alignment.
  • The HAP dataset of over 10K pairs enables effective anatomical alignment.
  • The HAF-Bench provides a systematic way to evaluate anatomical fidelity.

Where Pith is reading between the lines

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

  • This approach could be adapted to address other localized generation artifacts such as inconsistent lighting or object interactions.
  • Applying the synthetic preference construction to real user feedback data might further improve transfer to practical use cases.
  • Extending the margin-bounded optimization to other preference alignment domains could help avoid over-optimization in general.

Load-bearing premise

The anatomical errors created synthetically through localized degradation match the distribution of errors that naturally occur in text-to-image model outputs.

What would settle it

Running the ASAP alignment on a foundation model and measuring no decrease in anatomical error rates on the HAF-Bench compared to a baseline using standard DPO would falsify the effectiveness claim.

Figures

Figures reproduced from arXiv: 2605.25759 by Bao Li, Yuliang Xiu, Zhen Liu.

Figure 1
Figure 1. Figure 1: Effectiveness of ASAP. While foundation models like FLUX and SDXL frequently struggle with complex human anatomy (FLUX/SDXL-Base), especially for fingers, ASAP significantly improves anatomical plausibility (FLUX/SDXL-ASAP). anatomical structures remains a challenge [11,12,15,18,48]: existing foundation models frequently produce severe anatomical artifacts, such as malformed hands with incorrect digit coun… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the ASAP framework. Phase I (Left): [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual examples from the HAP dataset. Each column represents a pref￾erence pair conditioned on a specific text prompt. High-fidelity positive samples (xw) are shown in the top row, with correct anatomy highlighted in green boxes. Syn￾thesized degraded negative samples (xl) are in the bottom row, with anatomical arti￾facts highlighted in red boxes. HAP Dataset Construction. With the approaches described abo… view at source ↗
Figure 4
Figure 4. Figure 4: AER Evaluation on HAF￾Bench. Category-wise AER (↓) across five anatomical domains. The smaller polygons for ASAP-tuned models demonstrate consistent AER improve￾ments across all anatomical categories. Dataset Curation and Taxonomy. We curated a diverse evaluation set com￾prising 500 challenging prompts designed to capture the long-tail distribution of anatomical anomalies. To ensure both di￾versity and dif… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Comparison with Baselines. Visual comparison against SFT and DPO baselines. While standard SFT fails to resolve complex anatomical artifacts, unbounded DPO reduces these errors but inadvertently distorts background regions with checkerboard artifacts. Conversely, ASAP achieves the precise localized anatomi￾cal corrections without compromising the original global visual fidelity. these metrics. … view at source ↗
Figure 6
Figure 6. Figure 6: Impact of the Regularization Margin. The plot illustrates the trade-off between AER ↓ and HPSv3 ↑ across dif￾ferent margin parameters. Excessively large margins approximate unbounded DPO, aggressively reducing AER but severely compromising background semantics. Con￾versely, overly restrictive margins provide insufficient optimization space to correct anatomical artifacts. The optimal margin (τ = 0.01) achi… view at source ↗
Figure 7
Figure 7. Figure 7: Human Evaluation. (a) Human Assessed AER. Compared to the original FLUX and SDXL base models, our ASAP-tuned models significantly reduce human￾assessed anatomical errors. (b) Human Preference Voting. ASAP achieves a dominant win rate against the base models. This confirms that our method successfully improves anatomical correctness without compromising global image quality. engaged 11 independent raters wh… view at source ↗
Figure 8
Figure 8. Figure 8: Failure Case Analysis. (Left) Text-Image Misalignment: [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
read the original abstract

Large-scale text-to-image foundation models have achieved remarkable visual realism, yet generating human images with correct anatomical structures remains challenging. Existing approaches enforce anatomical constraints through part-specific modules or localized loss weighting during supervised fine-tuning on high-quality human photos, but such datasets are limited and often provide ambiguous optimization signals due to confounding factors such as lighting, pose, and background. Preference-based alignment offers an alternative, but standard Direct Preference Optimization (DPO) treats all pixels equally and therefore fails to exploit the localized nature of anatomical artifacts. To address this, we propose the framework of Alignment via Synthetic Anatomical Preference (ASAP), which constructs controlled preference pairs through a localized degradation mechanism applied to high-fidelity human images. This mechanism performs a controlled experiment on images by introducing explicit anatomical errors in targeted regions while preserving the remaining content. With this mechanism, we create the Human Anatomical Preference (HAP) dataset with over 10K curated pairs for effective anatomical alignment of text-to-image human image generative models. To better leverage the locality of these controlled preference pairs, we introduce a localized and margin-bounded variant of DPO that prioritizes optimization in targeted anatomical regions while enforcing a finite preference margin to prevent over-optimization and preserve global semantics. We further introduce HAF-Bench, a benchmark for systematic evaluation of anatomical fidelity. Extensive experiments demonstrate that ASAP consistently reduces anatomical errors across multiple foundation models while maintaining overall image quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes the ASAP framework to improve anatomical fidelity in text-to-image human generation. It builds the HAP dataset (>10K pairs) by applying a localized degradation mechanism to high-fidelity images that introduces explicit anatomical errors in targeted regions while preserving other content. A localized margin-bounded variant of DPO is introduced to focus optimization on those regions and avoid over-optimization. HAF-Bench is presented for systematic evaluation. The central claim is that ASAP reduces anatomical errors across foundation models while maintaining overall image quality.

Significance. If the synthetic preference pairs accurately capture the error distributions that arise naturally during T2I sampling, the approach offers a scalable route to anatomical alignment that sidesteps the scarcity and confounding factors of high-quality human photo datasets. The localized DPO formulation and the introduction of HAF-Bench are potentially reusable contributions for other localized artifact problems in generative models.

major comments (2)
  1. [Abstract] Abstract: the statement that the degradation mechanism 'performs a controlled experiment on images by introducing explicit anatomical errors in targeted regions' is used to ground the HAP dataset and the subsequent localized DPO, yet no argument or experiment is supplied showing that the introduced errors (isolated part distortions) match the frequency, co-occurrence, or causal structure of emergent T2I failures such as global limb-count inconsistencies or joint misalignments. This matching is load-bearing for the claim that the derived optimization signal addresses actual inference-time artifacts.
  2. [Abstract] Abstract (experiments paragraph): the claim of 'consistent error reduction' and 'extensive experiments' is asserted without any quantitative metrics, ablation tables, description of the degradation implementation, or definition of how HAF-Bench scores are computed, preventing verification of the central empirical claim.
minor comments (1)
  1. The abstract refers to 'over 10K curated pairs' and 'multiple foundation models' but supplies no breakdown by model, degradation severity, or region; these details belong in the experimental section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We address the major comments point by point below. Both points identify areas where the abstract can be strengthened for clarity and verifiability, and we will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the statement that the degradation mechanism 'performs a controlled experiment on images by introducing explicit anatomical errors in targeted regions' is used to ground the HAP dataset and the subsequent localized DPO, yet no argument or experiment is supplied showing that the introduced errors (isolated part distortions) match the frequency, co-occurrence, or causal structure of emergent T2I failures such as global limb-count inconsistencies or joint misalignments. This matching is load-bearing for the claim that the derived optimization signal addresses actual inference-time artifacts.

    Authors: We agree that an explicit argument or comparison would strengthen the grounding of the HAP dataset. The degradation mechanism is intentionally designed to introduce isolated, explicit anatomical errors in targeted regions to generate clean, localized preference signals that avoid the confounding factors (lighting, pose, background) present in real human preference data. The operations target known anatomical failure modes commonly reported in T2I literature (e.g., limb distortions, joint misalignments). The full manuscript describes these operations and shows downstream error reduction on HAF-Bench. However, we did not include a direct distributional comparison between synthetic and naturally occurring errors. We will add a dedicated paragraph in the method or discussion section providing justification based on observed T2I failure modes and include qualitative side-by-side examples of synthetic versus real artifacts to address this concern. revision: yes

  2. Referee: [Abstract] Abstract (experiments paragraph): the claim of 'consistent error reduction' and 'extensive experiments' is asserted without any quantitative metrics, ablation tables, description of the degradation implementation, or definition of how HAF-Bench scores are computed, preventing verification of the central empirical claim.

    Authors: The abstract is a high-level summary and therefore omits implementation details and full metrics, which appear in the main body (degradation implementation in Section 3, HAF-Bench definition and scoring in Section 4, quantitative results and ablations in Section 5 and the appendix). To make the central empirical claims more verifiable from the abstract itself, we will revise the experiments paragraph to include key quantitative results (e.g., error reduction percentages on HAF-Bench across models) while remaining within length constraints. revision: yes

Circularity Check

0 steps flagged

No circularity: method is data construction plus optimization with external grounding

full rationale

The paper describes a procedural pipeline: synthetic localized degradation to build the HAP preference dataset from high-fidelity images, followed by a localized margin-bounded DPO variant and evaluation on the introduced HAF-Bench. These elements are presented as explicit construction steps and an empirical optimization procedure whose inputs are external images and the standard DPO objective; no derivation, prediction, or uniqueness claim reduces by the paper's own equations or self-citations to a quantity defined in terms of the target result. The central claims rest on the empirical outcomes of this pipeline rather than tautological equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The central claim rests on the untested assumption that synthetic localized errors are distributionally equivalent to natural generation failures; no free parameters are named in the abstract, and no new physical or mathematical entities are postulated.

axioms (1)
  • domain assumption High-fidelity human images exist that can serve as clean starting points for controlled degradation.
    Invoked when the authors describe applying degradation to 'high-fidelity human images' to create preference pairs.
invented entities (3)
  • ASAP framework no independent evidence
    purpose: Method for constructing synthetic anatomical preference pairs and localized DPO training.
    New named procedure introduced in the abstract; no independent evidence supplied.
  • HAP dataset no independent evidence
    purpose: Collection of over 10K curated preference pairs for anatomical alignment.
    New dataset introduced; existence asserted but no release or verification details given.
  • HAF-Bench no independent evidence
    purpose: Benchmark for systematic evaluation of anatomical fidelity.
    New evaluation benchmark introduced; no scoring details or validation provided.

pith-pipeline@v0.9.1-grok · 5783 in / 1514 out tokens · 26879 ms · 2026-06-29T22:52:08.259244+00:00 · methodology

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

49 extracted references · 17 canonical work pages · 11 internal anchors

  1. [1]

    In:InternationalConferenceonArtificialIntelligenceandStatistics.pp.4447–4455

    Azar, M.G., Guo, Z.D., Piot, B., Munos, R., Rowland, M., Valko, M., Calandriello, D.: A general theoretical paradigm to understand learning from human preferences. In:InternationalConferenceonArtificialIntelligenceandStatistics.pp.4447–4455. PMLR (2024) 8

    2024

  2. [2]

    the method of paired comparisons

    Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika39(3/4), 324–345 (1952) 6

    1952

  3. [3]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025) 6

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    In: Forty-first international conference on machine learning (2024) 1, 3

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024) 1, 3

    2024

  5. [5]

    In: European Conference on Computer Vision

    Fang, G., Yan, W., Guo, Y., Han, J., Jiang, Z., Xu, H., Liao, S., Liang, X.: Human- refiner: Benchmarking abnormal human generation and refining with coarse-to-fine pose-reversible guidance. In: European Conference on Computer Vision. pp. 201–

  6. [6]

    The Pile: An 800GB Dataset of Diverse Text for Language Modeling

    Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., et al.: The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027 (2020) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  7. [7]

    Advances in neural information processing systems33, 6840–6851 (2020) 3

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020) 3

    2020

  8. [8]

    ICLR1(2), 3 (2022) 10

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022) 10

    2022

  9. [9]

    arXiv preprint arXiv:2505.22002 (2025) 5

    Hu, Z., Zhang, F., Kuang, K.: D-fusion: Direct preference optimization for aligning diffusion models with visually consistent samples. arXiv preprint arXiv:2505.22002 (2025) 5

    work page arXiv 2025

  10. [10]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Huang, Q., Chan, L., Liu, J., He, W., Jiang, H., Song, M., Song, J.: Patchdpo: Patch-level dpo for finetuning-free personalized image generation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 18369–18378 (2025) 4

    2025

  11. [11]

    ACM Computing Surveys56(11), 1–39 (2024) 2, 3 16 B

    Jia, Z., Zhang, Z., Wang, L., Tan, T.: Human image generation: A comprehensive survey. ACM Computing Surveys56(11), 1–39 (2024) 2, 3 16 B. Li et al

    2024

  12. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Ju, X., Zeng, A., Zhao, C., Wang, J., Zhang, L., Xu, Q.: Humansd: A native skeleton-guided diffusion model for human image generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15988–15998 (2023) 2

    2023

  13. [13]

    Advances in neural information processing systems36, 36652–36663 (2023) 4, 10

    Kirstain,Y.,Polyak,A.,Singer,U.,Matiana,S.,Penna,J.,Levy,O.:Pick-a-pic:An open dataset of user preferences for text-to-image generation. Advances in neural information processing systems36, 36652–36663 (2023) 4, 10

    2023

  14. [14]

    Labs, B.F.: Flux.https://github.com/black-forest-labs/flux(2024) 1, 3, 5, 9

    2024

  15. [15]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, S., Fu, J., Liu, K., Wang, W., Lin, K.Y., Wu, W.: Cosmicman: A text-to-image foundation model for humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6955–6965 (2024) 2, 4

    2024

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liang, Y., He, J., Li, G., Li, P., Klimovskiy, A., Carolan, N., Sun, J., Pont-Tuset, J., Young, S., Yang, F., et al.: Rich human feedback for text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19401–19411 (2024) 4

    2024

  17. [17]

    Flow Matching for Generative Modeling

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022) 3

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  18. [18]

    arXiv preprint arXiv:2310.08579 (2023) 2

    Liu, X., Ren, J., Siarohin, A., Skorokhodov, I., Li, Y., Lin, D., Liu, X., Liu, Z., Tulyakov, S.: Hyperhuman: Hyper-realistic human generation with latent struc- tural diffusion. arXiv preprint arXiv:2310.08579 (2023) 2

    work page arXiv 2023

  19. [19]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022) 3, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    In: The Fourteenth International Conference on Learning Representations 8

    Liu, X., Li, M., Lyu, Z., Shang, Y., Chen, C.: Learning from noisy preferences: A semi-supervised learning approach to direct preference optimization. In: The Fourteenth International Conference on Learning Representations 8

  21. [21]

    arXiv preprint arXiv:2508.03789 (2025) 4, 5, 10

    Ma, Y., Shui, Y., Wu, X., Sun, K., Li, H.: Hpsv3: Towards wide-spectrum human preference score. arXiv preprint arXiv:2508.03789 (2025) 4, 5, 10

    work page arXiv 2025

  22. [22]

    IEEE Robotics & automation magazine19(2), 98–100 (2012) 4

    Mori, M., MacDorman, K.F., Kageki, N.: The uncanny valley [from the field]. IEEE Robotics & automation magazine19(2), 98–100 (2012) 4

    2012

  23. [23]

    In: Proceedings of the Computer Vision and Pattern Recog- nition Conference

    Na, S., Kim, Y., Lee, H.: Boost your human image generation model via direct pref- erence optimization. In: Proceedings of the Computer Vision and Pattern Recog- nition Conference. pp. 23551–23562 (2025) 2

    2025

  24. [24]

    Advances in neural information processing sys- tems35, 27730–27744 (2022) 4

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in neural information processing sys- tems35, 27730–27744 (2022) 4

    2022

  25. [25]

    The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only

    Penedo, G., Malartic, Q., Hesslow, D., Cojocaru, R., Cappelli, A., Alobeidli, H., Pannier, B., Almazrouei, E., Launay, J.: The refinedweb dataset for falcon llm: outperforming curated corpora with web data, and web data only. arXiv preprint arXiv:2306.01116 (2023) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) 1, 3, 9

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  27. [27]

    Advances in neural information processing systems36, 53728–53741 (2023) 2, 4, 5, 6

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems36, 53728–53741 (2023) 2, 4, 5, 6

    2023

  28. [28]

    Journal of machine learning research21(140), 1–67 (2020) 4 ASAP 17

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research21(140), 1–67 (2020) 4 ASAP 17

    2020

  29. [29]

    Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024) 7

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022) 3, 5

    2022

  31. [31]

    Advances in neural information processing systems35, 25278–25294 (2022) 4, 10

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large- scale dataset for training next generation image-text models. Advances in neural information processing systems35, 25278–25294 (2022) 4, 10

    2022

  32. [32]

    1-dev-ControlNet-Union-Pro-2.0(2025) 6, 7

    Shakker-Labs: Controlnet-union.https://huggingface.co/Shakker-Labs/FLUX. 1-dev-ControlNet-Union-Pro-2.0(2025) 6, 7

    2025

  33. [33]

    Denoising Diffusion Implicit Models

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 3

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  34. [34]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) 3

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  35. [35]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Purushwalkam, S., Ermon, S., Xiong, C., Joty, S., Naik, N.: Diffusion model alignment using direct preference optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8228–8238 (2024) 2, 4, 5, 6

    2024

  36. [36]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, B., Zhou, J., Bai, J., Yang, Y., Chen, W., Wang, F., Lei, Z.: Realishuman: A two-stage approach for refining malformed human parts in generated images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 7509–7517 (2025) 4

    2025

  37. [37]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Wang, J., Sun, Z., Tan, Z., Chen, X., Chen, W., Li, H., Zhang, C., Song, Y.: To- wards effective usage of human-centric priors in diffusion models for text-based human image generation. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 8446–8455 (2024) 2, 4, 8

    2024

  38. [38]

    arXiv preprint arXiv:2502.06812 (2025) 4

    Wang, S., Tang, H., Dou, Z., Xiong, C.: Harness local rewards for global benefits: Effective text-to-video generation alignment with patch-level reward models. arXiv preprint arXiv:2502.06812 (2025) 4

    work page arXiv 2025

  39. [39]

    arXiv preprint arXiv:2507.02714 (2025) 2, 4

    Wang, Y., Cao, T., Zhang, H., He, Z., Liang, K., Ma, Z.: Fairhuman: Boosting hand and face quality in human image generation with minimum potential delay fairness in diffusion models. arXiv preprint arXiv:2507.02714 (2025) 2, 4

    work page arXiv 2025

  40. [40]

    Qwen-Image Technical Report

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025) 1, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis

    Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 (2023) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  42. [42]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2096–2105 (2023) 4, 10

    2096

  43. [43]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xing, X., Saha, A., He, J., Hao, S., Vicol, P., Ryu, M., Li, G., Singla, S., Young, S., Li, Y., et al.: Focus-n-fix: Region-aware fine-tuning for text-to-image generation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 18486–18496 (2025) 4 18 B. Li et al

    2025

  44. [44]

    Advances in Neural Information Processing Systems36, 15903–15935 (2023) 4, 5, 10

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere- ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems36, 15903–15935 (2023) 4, 5, 10

    2023

  45. [45]

    In: The Fourteenth International Conference on Learning Representations 8

    Yang, X., Yang, M., JIA, G., Qin, L., Tan, Z., Li, H.: Dual-ipo: Dual-iterative pref- erence optimization for text-to-video generation. In: The Fourteenth International Conference on Learning Representations 8

  46. [46]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yang, Z., Zeng, A., Yuan, C., Li, Y.: Effective whole-body pose estimation with two-stages distillation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4210–4220 (2023) 6

    2023

  47. [47]

    In: Proceedings of the IEEE international conference on computer vi- sion

    Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang, X., Metaxas, D.N.: Stack- gan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In: Proceedings of the IEEE international conference on computer vi- sion. pp. 5907–5915 (2017) 4

    2017

  48. [48]

    Advances in Neural Information Processing Systems37, 29354–29386 (2024) 2, 4

    Zhu, J., Chen, Y., Ding, M., Luo, P., Wang, L., Wang, J.: Mole: Enhancing human- centric text-to-image diffusion via mixture of low-rank experts. Advances in Neural Information Processing Systems37, 29354–29386 (2024) 2, 4

    2024

  49. [49]

    arXiv preprint arXiv:2512.10264 (2025) 4, 5, 6

    Ziv, A., Chen, S., Tjandra, A., Adi, Y., Hsu, W.N., Shi, B.: Mr-flowdpo: Multi- reward direct preference optimization for flow-matching text-to-music generation. arXiv preprint arXiv:2512.10264 (2025) 4, 5, 6

    work page arXiv 2025