Pith. sign in

REVIEW 3 major objections 6 minor 40 references

DiffDoctor: Diagnosing Image Diffusion Models Before Treating

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read DiffDoctor claims that locating artifacts first, then feeding pixel-level artifact confidence back into a text-to-image diffusion model, cuts artifact frequency sharply—FLUX.1 from 82.66% to 22.00%—while preserving image-reward and…

desk verdict Genuinely new pixel-level feedback approach, but the headline artifact-frequency numbers are circular and need independent verification. read the letter →

arxiv 2501.12382 v2 pith:5A23A6VO submitted 2025-01-21 cs.CV

classification cs.CV
keywords text-to-imagediffusionmodelsartifactdetectionpixel-levelfeedbackmodelfine-tuningclass-balanceddatahuman-in-the-loopannotationfrequencyreductiondiagnose-then-treat
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

DiffDoctor is a two-stage recipe for cleaning up text-to-image diffusion models. First it trains an artifact detector, a pixel-level segmenter that outputs a confidence map for three artifact classes (shape distortions, unreasonable content, watermarks), on a deliberately class-balanced, human-in-the-loop dataset of over a million images. Then it uses the detector's confidence maps as a per-pixel loss to fine-tune the diffusion model via LoRA, so the model learns to avoid generating those localized defects. The paper reports that on FLUX.1, SDXL, and Kolors the frequency of artifacts on unseen prompts falls from 82.66% to 22.00%, from 55.33% to 27.50%, and from 65.31% to 29.33%, respectively, while ImageReward and CLIP-T scores are preserved or slightly improved. The central thesis is that diagnosing where defects are, before treating the model, beats image-level reward feedback.

What carries the argument

The artifact map, a per-pixel confidence map produced by a SegFormer-b5 binary segmenter trained with mean-squared error on over a million images (starting from RichHF and PAL4VST annotations, augmented with real photos, hard cases, and pseudo-labels). Treating uses the pixel-level loss $L_{\text{pixel}} = \frac{1}{N_{\text{aggr}}} \sum_{i,j} M \circ C(\pi_\theta(z_T))[i,j]$, back-propagated through the detector into the diffusion denoiser with gradient truncation on the last 25% of steps, plus an optional rectified-flow offline regularization $\gamma L_{\text{offline}}$ with $\gamma = 0.25$ to delay collapse.

What would settle it

Measure artifact frequency on the same unseen prompts with an independently trained artifact detector or human labels; if the reported drops (for example, FLUX.1 from 82.66% to 22.00%) shrink substantially or vanish, the central claim is not supported and the treated model may simply be fooling the training detector.

Watch

Extended reading notes

Core claim

The paper's central discovery is that pixel-level artifact feedback is a viable and strong training signal for diffusion models. The authors build a robust artifact detector and show that the detector's false-positive behavior is decisive: when the detector has shortcut-like high confidence on faces and limbs, treating collapses the model into blurriness, while a balanced detector lowers artifact confidence smoothly. Using the artifact map with a threshold mask (pixels above 0.1) gives the best image quality, better than using all pixels or only the maximum-confidence pixel. They further show that diagnose-then-treat beats image-level reward fine-tuning with HPSv2 on artifact frequency, and that the pipeline transfers to DreamBooth-style customization.

Load-bearing premise

The load-bearing premise is that the artifact detector's confidence maps are an unbiased proxy for human-perceived artifacts; the same detector supplies the training signal and the primary evaluation metric, and the paper's own ablation notes that treating drives these confidences to near zero, so the reported artifact-frequency drop may reflect the model suppressing detector responses rather than genuinely removing defects.

Editorial extensions

If this is right

  • If the central claim holds, pixel-level artifact feedback can be used as a post-training step for any diffusion model that produces localized defects, without requiring paired reference images.
  • The reported artifact frequencies fall from 82.66% to 22.00% on FLUX.1, from 55.33% to 27.50% on SDXL, and from 65.31% to 29.33% on Kolors, while ImageReward and CLIP-T scores stay flat or improve.
  • Class-balancing the detector data is necessary: a detector trained on imbalanced annotations drives treating into collapse, so the diagnosis stage is what makes the treatment work.
  • The pipeline transfers to DreamBooth-style customization, reducing artifacts on unseen prompts while preserving subject identity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same diagnose-then-treat loop should transfer to other generative models with localized defects, such as video diffusion or 3D generation, as long as a pixel-level artifact detector exists for that modality.
  • Editorial inference: the artifact maps could also serve as editable masks for targeted regeneration or inpainting, rather than only as a training loss, potentially avoiding the layout shift that full-chain backpropagation can introduce.
  • Editorial inference: because the detector is trained on three artifact types (distortion, unreasonable content, and watermarks), the method's reach is bounded by the detector's vocabulary; extending to semantic or reasoning-dependent artifacts would require a detector that can reason, a limitation the paper acknowledges.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes DiffDoctor, a two-stage pipeline for reducing artifacts in text-to-image diffusion models. In the first stage, the authors train an artifact detector (SegFormer-b5) on a large dataset of flawed synthesized images, using a class-balancing strategy, human-in-the-loop labeling of hard cases, and semi-supervised pseudo-labeling. In the second stage, the trained detector is used as a differentiable pixel-level reward: the model is fine-tuned with a loss that minimizes detector confidence on artifact-prone pixels, combined with an offline diffusion-loss regularization term. Experiments on FLUX.1, SDXL, and Kolors report large reductions in 'mean artifact frequency' (e.g., FLUX.1 from 82.66% to 22.00%) and small improvements in ImageReward and CLIP-T scores, plus a 24-user study asking participants to choose between images before and after treating.

Significance. If the reported artifact reductions are genuine, DiffDoctor introduces a useful and novel 'diagnose-then-treat' paradigm, and it is, to my knowledge, the first method to use pixel-level artifact feedback for fine-tuning diffusion models. The paper's contribution also includes a large-scale human-in-the-loop dataset collection effort and a careful discussion of data imbalance in existing artifact annotations. The method is demonstrated on multiple backbones and on a downstream DreamBooth task, and the qualitative results show plausible artifact suppression on unseen prompts. However, the central quantitative evidence for the headline artifact-frequency drop is undermined by a circularity between the training signal and the evaluation metric, and the independent metrics (ImageReward, CLIP-T, and the small user study) do not by themselves establish the claimed magnitude. The contribution is promising, but the current evidence needs substantial strengthening before the central claim can be accepted.

major comments (3)
  1. [Sec. 4.2/4.3, Eq. (1), Tab. 4] The primary outcome 'Mean Artifact Freq.' is computed by thresholding the max artifact confidence of the same detector that is minimized by L_pixel in Eq. (1). Because L_pixel directly suppresses the detector's confidence values, and Sec. 4.2 states that 'the artifact confidences are all optimized to low values,' the large drops in Table 4 (e.g., FLUX.1 from 82.66% to 22.00%) are expected by construction and do not establish the removal of human-visible artifacts. The authors should re-evaluate artifact frequency using an independent artifact detector, human annotations on a held-out set, or an artifact-specific perceptual metric that is not used in training.
  2. [Sec. 4.1, detector benchmark] The artifact detector is validated only on a self-constructed benchmark of 771 images, and the RichHF test set is rejected because its annotations are fixed-radius circles. Without an external benchmark, there is no independent anchor for the detector's accuracy. Since the detector is the core supervisory signal for treating, I recommend reporting agreement with human annotators (e.g., pixel-level IoU or AUC) on a sample of images and, if feasible, also providing results on the RichHF test set as a secondary benchmark despite its annotation limitations.
  3. [Sec. 4.3, Tab. 5] The user study has only 24 participants and reports artifact-winning rates of 63.64%, 68.18%, and 60.98% for DiffDoctor over the three backbones, but no significance testing, confidence intervals, or inter-annotator agreement are provided. With a forced binary choice and n=24, a 60% win rate is not statistically significant; the study needs formal testing (e.g., a binomial test or exact confidence intervals) and ideally more participants to support the claim that DiffDoctor reduces artifacts as judged by humans.
minor comments (6)
  1. [Sec. 4.1] The sentence 'We conduct most experiments o FLUX.1 Schnell' contains a typo: 'o' should be 'on'.
  2. [Fig. 4] The y-axis label 'Mas Artifact Confidence' should be 'Max Artifact Confidence'.
  3. [Algorithm 1] Line 16 'Lpixel aggregate ← − − − −C(x)' is unclear; the aggregation over selected pixels should be written with a proper sum or an explicit symbol (e.g., Σ_{selected} C(x)[i,j]).
  4. [Sec. 4.2, Tab. 2 vs. Tab. 4] The row 'Threshold + Best Detector' in Table 2 reports the same ImageReward/CLIP-T values as 'FLUX.1 + DiffDoctor' in Table 4; the text should clarify whether these are the same checkpoint and whether the offline regularization is included in both.
  5. [Sec. 3.2] The selection of the 2k hard cases via 'thresholding the images detected with extremely high or low artifact confidence' is described briefly; please specify the thresholds and the number of active-learning rounds used in practice.
  6. [General] The paper does not state whether the collected dataset, the artifact detector weights, or the code will be released; providing this information would improve reproducibility.

Circularity Check

1 steps flagged · score 8.0 of 10

Headline artifact-frequency drop is measured with the same detector used as the training loss (Eq. 1 vs. Tab. 4); Sec. 4.2 concedes confidences are driven to low values.

  1. fitted input called prediction [Sec. 3.3 Eq. (1); Sec. 4.1 'Model treating metrics'; Sec. 4.2; Tab. 4]
    "Lpixel(θ) = 1/Naggr Σ M ◦ C(πθ(zT ))[i, j] ... This loss suppresses the diffusion model for generating high artifact confidence areas ... Furthermore, we threshold max artifact confidences of images(0.5) to measure whether an image contains artifacts, based on which we calculate the 'mean artifact frequencies' of the test set. ... Note that the artifact frequency metric is not feasible here, for the artifact confidences are all optimized to low values."

    The headline metric (Tab. 4 'Mean Artifact Freq.') is the thresholded output of the same artifact detector confidence map C that Eq. (1) minimizes per-pixel in Lpixel. The diffusion model is therefore optimized to suppress exactly the signal used to measure improvement, so the reported drop (e.g., FLUX.1 82.66% to 22.00%) is in part a training curve of the objective rather than an independent measure of human-visible artifacts. The paper itself concedes in Sec. 4.2 that 'the artifact confidences are all optimized to low values,' rendering the metric 'not feasible' in its own mask ablation; that is an admission that the detector response, not an externally validated artifact quantity, is being driven down.

full rationale

The central quantitative claim of DiffDoctor is that treating reduces artifact frequency by large margins (82.66% to 22.00% on FLUX.1, etc., Tab. 4). That metric is defined by thresholding the artifact detector's confidence maps, and the treating loss (Eq. 1) minimizes exactly those confidence maps. This makes the main reported outcome circular: the model is rewarded for suppressing the detector, and the same detector is then used to score the result. The paper's own Sec. 4.2 states that after treating 'the artifact confidences are all optimized to low values' and that the artifact frequency metric is therefore 'not feasible' in the mask ablation, confirming the detector output is the quantity being optimized. The independent signals (ImageReward, CLIP-T, and the 24-user preference study) are not adequate substitutes: they are global text-alignment/quality scores or a small unpowered user study that does not report significance or inter-annotator agreement, and they do not measure artifact localization. The artifact detector itself is grounded in human labels and a 771-image author-constructed benchmark, so the detector is not vacuous; this prevents a score of 10. No load-bearing self-citation chain is present: citations to the authors' Depth Anything work support the pseudo-labeling data pipeline but are not the basis of the headline claim. The central reported magnitude, however, is substantially a self-measurement, so a high circularity score is warranted.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method relies on a learned artifact detector as the source of both training signal and evaluation metric. Its validity is the key unproven premise. The data collection involves hand-chosen thresholds and empirical weights (gamma, mask threshold, truncation), all of which affect the outcome. No new physical or conceptual entities are introduced.

free parameters (7)
  • gamma = 0.25
    Empirically chosen weight for the rectified-flow regularization loss in L = L_pixel + gamma * L_offline.
  • mask_threshold = 0.1
    Pixels with artifact confidence > 0.1 are included in L_pixel (best in Tab. 2 ablation).
  • artifact_frequency_threshold = 0.5
    Images with max artifact confidence > 0.5 are labeled as containing artifacts in the primary metric.
  • gradient_truncation_timestep = last 25% of denoising steps
    Gradients are only tracked through the final 25% of the denoising chain to save memory.
  • LoRA_rank = 16
    Rank of LoRA adaptation during treating.
  • learning_rate = 1e-4
    Learning rate used for treating experiments.
  • hard_case_thresholds = unspecified
    The 2k hard cases are selected by thresholding artifact confidence extremes, but the thresholds are not given.
assumptions (4)
  • domain assumption Artifact confidence maps from the trained detector are a reliable supervision signal for artifact reduction
    The entire pixel-aware treating step back-propagates detector confidences; if the detector is biased or foolable, the fine-tuned model may reduce detector responses without reducing real artifacts.
  • domain assumption Human annotations and existing datasets (RichHF, PAL4VST) provide accurate artifact labels
    The detector is trained with these labels plus 2k hard-case human labels; label noise degrades the detector.
  • domain assumption Gradient truncation through the last 25% of denoising steps yields useful gradients
    The paper truncates gradients for memory; this assumes artifact formation is captured in the final denoising steps.
  • domain assumption Rectified flow loss can act as a regularizer to prevent model collapse
    L_offline is added with gamma=0.25 to keep the fine-tuned model close to the real distribution; this assumes the standard diffusion loss has the intended anchoring effect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffDoctor: Diagnosing Image Diffusion Models Before Treating." pith.science (2026). https://pith.science/paper/5A23A6VO

@misc{pith2026250112382,
  author       = {Pith},
  title        = {Pith review of: DiffDoctor: Diagnosing Image Diffusion Models Before Treating},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5A23A6VO}},
  note         = {Machine review of arXiv:2501.12382}
}
read the original abstract

In spite of recent progress, image diffusion models still produce artifacts. A common solution is to leverage the feedback provided by quality assessment systems or human annotators to optimize the model, where images are generally rated in their entirety. In this work, we believe problem-solving starts with identification, yielding the request that the model should be aware of not just the presence of defects in an image, but their specific locations. Motivated by this, we propose DiffDoctor, a two-stage pipeline to assist image diffusion models in generating fewer artifacts. Concretely, the first stage targets developing a robust artifact detector, for which we collect a dataset of over 1M flawed synthesized images and set up an efficient human-in-the-loop annotation process, incorporating a carefully designed class-balance strategy. The learned artifact detector is then involved in the second stage to optimize the diffusion model by providing pixel-level feedback. Extensive experiments on text-to-image diffusion models demonstrate the effectiveness of our artifact detector as well as the soundness of our diagnose-then-treat design.

Figures

Figures reproduced from arXiv: 2501.12382 by the authors.

Figure 1
Figure 1. Illustrations of DiffDoctor. We train a robust de￾tector to localize the artifacts (diagnosing) and provide pixel-level feedback to optimize the diffusion model (treating). After tun￾ing on limited samples, the diffusion model generates significantly fewer artifacts on unseen prompts while maintaining the quality. but also inadvertently learn from the noise, which manifests as artifacts during image generation. Rega… view at source ↗
Figure 2
Figure 2. Pipeline of DiffDoctor. The first part shows the training of an artifact detector – the doctor. Starting with the initial dataset, the artifact detector is trained in a humans-in-a-loop manner. The second part shows our diagnose-then-treat design, where the patient – a trainable diffusion model, is prompted to synthesize images. Then the frozen artifact detector diagnoses its result by predicting the artifact maps, … view at source ↗
Figure 3
Figure 3. Qualitative ablation study of artifact detectors. We visualize the artifact maps predicted by the artifact detector on our hard benchmark by headmaps [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Mode collapse due to naive artifact detector. The treating process collapses into blurriness. We further visualize the training and evaluation time artifact confidence curve [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: DiffDoctor on FLUX.1. All images are synthesized based on randomly generated unseen prompts not involved in training, and on the same seeds for the corresponding images before and after DiffDoctor. After treating, artifacts in images are reduced, but the content and la…
Figure 7
Figure 7. Figure 7: Qualitative comparisons on FLUX.1, SDXL, Kolors. We compare DiffDoctor with using HPSv2 to optimize diffusion models and in-painting methods that redraw artifacts using masks. and Kolors. We use HPSv2 [29] to measure human pref￾erence and tune the diffusion model to ma…
Figure 8
Figure 8. Figure 8: DiffDoctor on DreamBooth. We perform DiffDoc￾tor on the instance prompt and evaluate other various prompts. on customized generation. Specifically, we train Dream￾Booth [23] on FLUX.1 Dev, with an inference step of 10. Then we further optimize these LoRA layers using D…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 26 canonical work pages

  1. [1]

    https : / / civitai

    Hands xl. https : / / civitai . com / models / 200255/hands-xl-sd-15-flux1-dev , 2024. 7

  2. [2]

    Qwen technical report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xi- aodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv:2309.16609, 2023. 4, 5

  3. [3]

    Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jin- gren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 4

  4. [4]

    Improving image generation with bet- ter captions

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with bet- ter captions. https://cdn.openai.com/papers/ dall-e-3.pdf, 2023. 1

  5. [5]

    Training diffusion models with reinforce- ment learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforce- ment learning. In ICLR, 2024. 1, 2

  6. [6]

    Synartifact: Classifying and alleviat- ing artifacts in synthetic images via vision-language model

    Bin Cao, Jianhao Yuan, Yexin Liu, Jian Li, Shuyang Sun, Jing Liu, and Bo Zhao. Synartifact: Classifying and alleviat- ing artifacts in synthetic images via vision-language model. arXiv:2402.18068, 2024. 1, 2

  7. [7]

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable re- wards. In ICLR, 2024. 1, 2, 4, 7

  8. [8]

    Scaling recti- fied flow transformers for high-resolution image synthesis, march 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis, march 2024. URL http://arxiv. org/abs/2403.03206, 2024. 4

Show all 40 references
  1. [9]

    Re- inforcement learning for fine-tuning text-to-image diffusion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Moham- mad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Re- inforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2024. 1, 2

  2. [10]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. In NeurIPS, 2020. 1

  3. [11]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 5

  4. [12]

    Noise-aware learning from web-crawled image-text data for image captioning

    Wooyoung Kang, Jonghwan Mun, Sungjun Lee, and Byungseok Roh. Noise-aware learning from web-crawled image-text data for image captioning. In ICCV, 2023. 1

  5. [13]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Ma- tiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeurIPS, 2023. 1

  6. [14]

    Black Forest Labs. Flux.1. https : / / blackforestlabs.ai, 2024. 1, 5

  7. [15]

    Rich human feedback for text-to-image generation

    Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In CVPR, 2024. 1, 2, 3, 4

  8. [16]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5

  9. [17]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 1, 4

  10. [18]

    Training lan- guage models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training lan- guage models to follow instructions with human feedback. In NeurIPS, 2022. 2

  11. [19]

    SDXL: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 5

  12. [20]

    Aligning text-to-image diffusion models with reward backpropagation

    Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Ka- terina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv:2310.03739, 2023. 1, 2, 4

  13. [21]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2024. 2

  14. [22]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 1

  15. [23]

    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. In CVPR, 2023. 2, 8

  16. [24]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 5

  17. [25]

    Proximal policy optimization algo- rithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Rad- ford, and Oleg Klimov. Proximal policy optimization algo- rithms. arXiv:1707.06347, 2017. 2

  18. [26]

    Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis

    Kolors Team. Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint,

  19. [27]

    Diffusion model align- ment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model align- ment using direct preference optimization. In CVPR, 2024. 1, 2

  20. [28]

    Detecting photoshopped faces by scripting photoshop

    Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In CVPR, 2019. 2

  21. [29]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv:2306.09341, 2023. 1, 7

  22. [30]

    Segformer: Simple and ef- ficient design for semantic segmentation with transformers

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and ef- ficient design for semantic segmentation with transformers. In NeurIPS, 2021. 4 9

  23. [31]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation. In NeurIPS, 2024. 1, 2

  24. [32]

    Fakeshield: Explainable image forgery detection and localization via multi-modal large lan- guage models

    Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large lan- guage models. arXiv:2410.02761, 2024. 2

  25. [33]

    St++: Make self-training work better for semi-supervised se- mantic segmentation

    Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised se- mantic segmentation. In CVPR, 2022. 4

  26. [34]

    Revisiting weak-to-strong consistency in semi-supervised semantic segmentation

    Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In CVPR, 2023. 4

  27. [35]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2

  28. [36]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. In NeurIPS, 2024. 4

  29. [37]

    Diffforensics: Leveraging diffusion prior to image forgery detection and localization

    Zeqin Yu, Jiangqun Ni, Yuzhen Lin, Haoyi Deng, and Bin Li. Diffforensics: Leveraging diffusion prior to image forgery detection and localization. In CVPR, 2024. 2

  30. [38]

    Per- ceptual artifacts localization for inpainting

    Lingzhi Zhang, Yuqian Zhou, Connelly Barnes, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Per- ceptual artifacts localization for inpainting. In ECCV, 2022. 2

  31. [39]

    Perceptual artifacts local- ization for image synthesis tasks

    Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts local- ization for image synthesis tasks. In ICCV, 2023. 1, 2, 3, 4

  32. [40]

    Large-scale reinforcement learning for diffusion models

    Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. In NeurIPS, 2024. 1, 2 10

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.