Pith. sign in

REVIEW 5 major objections 5 minor 49 references

Preserve Anything: Controllable Image Synthesis with Object Preservation

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Preserve Anything claims object-preserving image synthesis with an N-channel ControlNet, reaching FID 15.26 and CLIP-S 32.85.

desk verdict Solid engineering, weak evidence: the SOTA claim rests on non-comparable baselines and possible train/eval overlap. read the letter →

arxiv 2506.22531 v2 pith:SQCBMGBL submitted 2025-06-27 cs.CV

Prasen Kumar Sharma , Neeraj Matiyali , Siddharth Srivastava , Gaurav Sharma This is my paper
classification cs.CV
keywords controlledimagesynthesisobjectpreservationtext-to-imagegenerationControlNetlightingconsistencybackgroundguidancehigh-frequencyoverlaydatasetcuration
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

Preserve Anything aims to fix a persistent failure of text-to-image models: when a user supplies a photo of an object and wants it placed in a new scene, existing pipelines either distort the object, lose fine text and patterns, or cannot control the background and lighting. The paper's central claim is that all of these can be handled by one N-channel ControlNet that takes the object image, its mask, an optional background layout, and an optional lighting direction as separate conditioning channels, so the diffusion model generates the scene around the object rather than blending the object into a scene afterward. To support this, the authors introduce PA-240K, a curated benchmark of 240K aesthetically filtered natural images and 18K 3D-rendered images with known lighting and camera metadata, and they report state-of-the-art FID 15.26 and CLIP-S 32.85, with user-study gains of roughly 25% in prompt alignment, 19% in photorealism, 13% in AI-artifact reduction, and 14% in natural aesthetics over Inpaint Anything. A sympathetic reading of the contribution is the demonstration that explicit channels for object, mask, background, and lighting make object-preserving synthesis controllable at inference without per-object fine-tuning.

What carries the argument

The load-bearing mechanism is the N-channel ControlNet conditioning vector $c_f = E(I, M, B, L)$, where $I$ is the object RGB image on a white background, $M$ is the object mask, $B$ is the background layout, and $L$ is the lighting gradient map. The argument runs through this design: object identity is carried by the RGB-plus-mask pair rather than by fine-tuned subject embeddings, background controllability comes from the HED edge layout, and lighting consistency comes from a gradient map derived from spherical angles. The high-frequency overlay then completes the pipeline, combining low frequencies from the generated scene with high frequencies from the source object via $\hat{J} = M \cdot \ell_{hf}(I) + (1 - M) \cdot \ell_{hf}(J) + \ell_{lf}(J)$, preserving text and fine patterns without external blending tools.

What would settle it

Recompute FID and CLIP-S on a held-out MS-COCO subset that is provably disjoint from PA-240K and generated with the same pipeline as the baselines; if FID rises by more than the reported 0.83 gap over AnyScene, the central performance claim fails. Separately, pixel-wise comparison of the masked object region before and after the high-frequency overlay would show whether fine text and patterns are truly preserved.

Watch

Extended reading notes

Core claim

The paper's central discovery is that object-preserving controlled image synthesis can be achieved by extending ControlNet to an N-channel design, with N=6 in the current implementation: three channels carry the target object on a white canvas, one channel carries its binary mask, one channel carries a background layout derived from holistically-nested edge detection, and one channel carries a lighting gradient map converted from spherical coordinates. The object mask enforces color and detail fidelity without requiring additional loss functions, the background layout gives explicit user control over scene structure, and the lighting channel aligns shadows and highlights with the stated illumination direction. After generation, a high-frequency overlay module re-injects fine details such as text and patterns from the source object, replacing Poisson blending with a lightweight Gaussian-blur decomposition. The paper reports that this design achieves the best FID (15.26) and CLIP-IQA (0.6869) among compared methods, competitive aesthetic scores, and generalization to unseen objects drawn from the Dreambooth dataset.

Load-bearing premise

The claim of state-of-the-art performance assumes the baseline numbers reproduced from AnyScene were measured under the same evaluation protocol and that the 3,000 test images are disjoint from the 240,000 training images, although both come from MS-COCO.

Editorial extensions

If this is right

  • A single product photo can be placed into arbitrary scenes at different sizes and positions without per-object fine-tuning.
  • Users can control background layout through an edge map and lighting through a spherical-direction gradient map, making shadows and highlights follow the stated light source.
  • Text and fine patterns on preserved objects survive generation because the high-frequency overlay re-injects them from the source image.
  • The curated PA-240K dataset with lighting, camera, and relationship annotations provides training signal that generalizes to new objects from the Dreambooth benchmark.
  • Switching the backbone from SD v1.5 to SDXL further improves FID to 14.54, so the design carries over to higher-resolution generators.

Reading between the lines

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

  • The random channel-pruning of background layout and lighting during training implies the same model can accept partial condition sets at inference, so lighting-only or layout-only control should work even though the paper reports those cases mainly qualitatively.
  • Because the mask is the only hard object constraint, the paper implicitly claims object identity is carried by RGB-plus-mask rather than by learned subject embeddings; a natural test is whether the method preserves objects never seen during training, which the Dreambooth generalization test only partially covers.
  • The curated dataset's lighting, camera, and relationship annotations could be reused for other tasks such as shadow estimation and depth prediction, an extension the paper mentions but does not evaluate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes Preserve Anything, an N-channel extension of ControlNet for controlled text-to-image synthesis with object preservation. The framework takes as conditioning a foreground object on a white canvas, a binary mask, an optional background layout edge map, and an optional lighting-direction map; after generation it applies a cheap high-frequency overlay (Eq. 10) to restore source details. The authors introduce PA-240K, a curated dataset of 240K aesthetic-filtered real images with GPT-4o annotations plus 18K Blender-rendered images with lighting/shadow maps, and they report state-of-the-art results: FID 15.26, CLIP-S 32.85, and user-study improvements of roughly 25% in prompt alignment, 19% in photorealism, 13% in AI artifacts, and 14% in natural aesthetics over Inpaint Anything.

Significance. If the quantitative claims were supported, this would be a practically useful contribution: the conditioning design is simple and flexible, the dataset is a potentially valuable resource for controlled image synthesis, and the high-frequency overlay in Eq. (10) is a refreshingly inexpensive alternative to Poisson blending. The paper also demonstrates interesting qualitative capabilities: size/placement-agnostic multi-object preservation, background-layout control via HED edges, and lighting-conditioned generation. However, the central state-of-the-art claim currently rests on a comparison in Table 1 that mixes baseline numbers borrowed from AnyScene with the authors' own in-house metrics, and the text does not establish that the 3K evaluation images are disjoint from the 240K training images. These are load-bearing gaps, so the paper needs major revision before the claims can be accepted.

major comments (5)
  1. [§5, Table 1 and caption] The caption states that "The SoTA numbers for baseline methods are taken from AnyScene [5]," while the Preserve Anything row is computed in-house. This mixes measurement protocols. The scale mismatch is stark: AnyScene is listed with CLIP-S 15.18, whereas Preserve Anything scores 32.85 and the authors' own RGB Only-GPT4o ablation scores 32.38 in Table 2. A CLIP-S gap of that size cannot be attributed to method quality; it indicates different CLIP models, prompt templates, or preprocessing. Consequently, the abstract's claim of significant semantic-alignment improvement over AnyScene is not supported. In addition, the same table internally contradicts a blanket prompt-alignment claim: Preserve Anything's 32.85 is below SDXL's 34.18 and Inpaint Anything's 33.30. Please recompute all baselines under one protocol, or restrict the claims to the comparisons that are actually measured.
  2. [§4 and §3.1] The text says the training set comprises 240K images and evaluation is on 3K images, and that "Following [5], these images are a subset of MS-COCO." The real-image subset of PA-240K is also drawn from MS-COCO (plus OpenImagesV7 and FFHQ). The paper never states that the 3K evaluation images were excluded from the 240K training images. If there is overlap, FID and object-preservation scores can be inflated by near-memorization of training foregrounds. Please specify the exact split, verify non-overlap, and state the FID reference distribution (real-image set and size) used for the score in Table 1.
  3. [§7, Table 3 and Figure 7] The user study uses only five annotators and thirty images, and it reports no inter-rater reliability, no confidence intervals, and no significance test. The per-user breakdown in Figure 7 is revealing: User 3 rates Inpaint Anything higher than Preserve Anything on all four criteria. The headline improvements of roughly 25%, 19%, 13%, and 14% are therefore not robustly established. Please report per-item distributions, a significance test or effect-size measure, and the number of ratings per image; otherwise soften the perceptual-claim language.
  4. [§5, Feature Space Analysis] The claim that FID 15.26 "significantly outperforms" AnyScene (16.09) and Inpaint Anything (17.17) is made without error bars, multiple seeds, or significance testing. Single-run FID differences of this magnitude can be within noise, particularly when the reference set is not precisely specified. Please provide confidence intervals or repeated-evaluation statistics, and ensure the reference distribution is identical across all compared methods.
  5. [§B, Eq. (10) and Figure 4] The text states that the high-frequency overlay is "as effective as Poisson Blending" and does not incur computational cost, but no quantitative comparison to Poisson blending is provided. Since this module is presented as a key contribution replacing prior blending, please add a direct evaluation (e.g., detail-retention metrics or a focused user study on text/pattern fidelity) rather than relying only on the qualitative Figure 4.
minor comments (5)
  1. [§3.2, Lighting Maps] The conversion of spherical lighting coordinates into gradient maps is described only verbally; please give the precise formula or normalization used for the L channel so the method is reproducible.
  2. [§3.1] The dataset section states that the LAION-Aesthetic threshold of 5.0 was "chosen based on the high-quality standards of MIT Adobe 5K," but no quantitative justification or reference to a specific distribution statistic is given; please clarify how the threshold was derived.
  3. [§2, Related Works] The reference for Replace Anything [4] is cited as "Virtualmodel," whereas the text refers to Replace Anything; please verify the citation and, if the paper is different, add the correct reference.
  4. [§3.2, Equation (1)] The conditioning vector c_f is defined as the output of encoder E, but the notation does not specify the network architecture or how c_f is fused into the ControlNet pathway; a sentence or pointer to the supplementary architecture details would improve reproducibility.
  5. [Appendix F] The GPT-4o visual-critic evaluation for lighting and shadow consistency is mentioned without reporting the actual scores or the number of images evaluated; please report the numbers or remove the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical training/evaluation pipeline, and none of its reported gains reduce by construction to its inputs.

full rationale

The paper makes no first-principles derivation whose conclusion is equivalent to an input; it presents an N-channel ControlNet variant trained on a curated dataset and evaluated with standard metrics. The foreground image and mask are input conditioning channels, but object preservation is an optimization goal, not a claimed prediction derived from those channels by definition. The headline comparisons borrow baseline numbers from AnyScene (Table 1 caption: "The SoTA numbers for baseline methods are taken from AnyScene [5]") and the evaluation set is drawn from MS-COCO like the training set, but these are protocol and comparability risks: they do not make the reported FID, CLIP-S, or user-study numbers equal to any fitted parameter or input by construction. The use of LAION-Aesthetic both as a training-data filter and as an evaluation metric is a self-referential evaluation choice, but it does not force the reported aesthetic scores, since the metric is an external pretrained predictor rather than a function of the training filter. The paper also contains no load-bearing self-citation; its cited baselines and tools ([5], [45], [47], [49], etc.) are external to the authors. The acknowledged lack of public code for AnyScene (Section 7) is a limitation on comparison breadth, not a circular step. No specific reduction can be exhibited, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several modeling assumptions (HED edges as background layout, gradient map as lighting, high-frequency overlay as detail preservation) and on evaluation assumptions that are not fully validated. The model itself is a trained neural network, so the fitted weights are not counted as free parameters; only manually chosen thresholds and hyperparameters are listed.

free parameters (4)
  • LAION-Aesthetic filtering threshold = 5.0
    Chosen based on MIT Adobe 5K average; controls which training images are retained and influences reported aesthetic and FID scores.
  • Gaussian blur kernel size for low/high frequency split = 17x17
    Manually chosen for Eqs. (9) and (10); determines what counts as high-frequency detail in the overlay module.
  • Classifier-free guidance condition drop rate = 0.05
    Hyperparameter used during training to enable conditioning dropout.
  • Learning rate and batch size = 1e-4, batch 16
    Training hyperparameters reported in Section 4.
assumptions (5)
  • domain assumption The LAION aesthetic predictor provides a valid measure of image quality for filtering and evaluation.
    Invoked in Section 3.1 and in the quality assessment; if this score is not a reliable aesthetic proxy, the dataset curation and aesthetic metrics are compromised.
  • domain assumption A HED edge map of a reference image is a sufficient structural layout condition for background generation.
    Used in Eq. (2) in Section 3.2; the method relies on edges to convey the scene layout to the ControlNet.
  • domain assumption A single lighting direction, converted to a 2D gradient map, is sufficient to enforce consistent illumination and shadows in the synthesized scene.
    Section 3.2; no physical rendering model is used, and the paper does not quantify lighting accuracy beyond a GPT-4o critic that yields no reported numbers.
  • ad hoc to paper The high-frequency overlay defined by Eq. (10) transfers source details onto the generated object without introducing artifacts.
    Appendix B; the paper calls this a heuristic and only shows qualitative examples, not a quantitative comparison with Poisson blending.
  • standard math Diffusion model training objective and ControlNet conditioning are valid standard tools.
    Appendix A; used without proof as background methodology.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Preserve Anything: Controllable Image Synthesis with Object Preservation." pith.science (2026). https://pith.science/paper/SQCBMGBL

@misc{pith2026250622531,
  author       = {Pith},
  title        = {Pith review of: Preserve Anything: Controllable Image Synthesis with Object Preservation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQCBMGBL}},
  note         = {Machine review of arXiv:2506.22531}
}
abstract

We introduce \textit{Preserve Anything}, a novel method for controlled image synthesis that addresses key limitations in object preservation and semantic consistency in text-to-image (T2I) generation. Existing approaches often fail (i) to preserve multiple objects with fidelity, (ii) maintain semantic alignment with prompts, or (iii) provide explicit control over scene composition. To overcome these challenges, the proposed method employs an N-channel ControlNet that integrates (i) object preservation with size and placement agnosticism, color and detail retention, and artifact elimination, (ii) high-resolution, semantically consistent backgrounds with accurate shadows, lighting, and prompt adherence, and (iii) explicit user control over background layouts and lighting conditions. Key components of our framework include object preservation and background guidance modules, enforcing lighting consistency and a high-frequency overlay module to retain fine details while mitigating unwanted artifacts. We introduce a benchmark dataset consisting of 240K natural images filtered for aesthetic quality and 18K 3D-rendered synthetic images with metadata such as lighting, camera angles, and object relationships. This dataset addresses the deficiencies of existing benchmarks and allows a complete evaluation. Empirical results demonstrate that our method achieves state-of-the-art performance, significantly improving feature-space fidelity (FID 15.26) and semantic alignment (CLIP-S 32.85) while maintaining competitive aesthetic quality. We also conducted a user study to demonstrate the efficacy of the proposed work on unseen benchmark and observed a remarkable improvement of $\sim25\%$, $\sim19\%$, $\sim13\%$, and $\sim14\%$ in terms of prompt alignment, photorealism, the presence of AI artifacts, and natural aesthetics over existing works.

Figures

Figures reproduced from arXiv: 2506.22531 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. (a) Preserve Anything builds upon the ControlNet framework [47], extending it to an N-channel design with N=6 in the current implementation. The first three channels are dedicated to the RGB image of the target object, which is to be preserved against a white background, referred to as FG. The fourth channel, labeled Mask, encodes the binary mask of the object. The fifth channel r… view at source ↗
Figure 2
Figure 2. Qualitative comparison of the proposed method (Preserve Anything) against the SD/ControlNet baselines and the state-of-the-art [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. High-frequency overlay module recovers complex text [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Our method allows control over the background appear [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: With explicit lighting direction control via a gradient [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Visual demonstration of user-wise responses across four parameters. Top-left shows the preferences of users in terms of prompt alignment. Top-right shows the preferences in terms of photorealism. Bottom-left indicates the preferences in terms of AI artifacts. Whereas t…
Figure 9
Figure 9. Figure 9: An illustration of our large-scale dataset. (Left) shows an image from the real-world subset with its derived annotations (prompt and lighting cues for demonstration) from GPT4o. (Right) shows two samples from the curated 3D-rendered synthetic subset with their corresp…
Figure 10
Figure 10. Figure 10: Consistency in natural aesthetics w.r.t. prompts. Given a fixed background layout, observe how well the proposed method maintains the natural aesthetics such as lighting, casting shadows, and reflection w.r.t. input foreground and textual prompt, across various genera…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages

  1. [5]

    R. Chen, L. Wang, W. Nie, Y . Zhang, and A. Liu. Anyscene: Customized image synthesis with composited foreground. In 2024 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 8724–8733, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 1, 2, 3, 4, 5, 6, 7 8https://github.com/chaofengc/IQA-PyTorch

    work page 2024

  2. [1]

    Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  3. [2]

    From colouring-in to pointillism: revisiting semantic segmentation supervision

    Rodrigo Benenson and Vittorio Ferrari. From colouring-in to pointillism: revisiting semantic segmentation supervision. CoRR, abs/2210.14142, 2022. 3, 9

    arXiv 2022

  4. [3]

    Learning photographic global tonal adjustment with a database of input / output image pairs

    Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fr ´edo Durand. Learning photographic global tonal adjustment with a database of input / output image pairs. InThe Twenty- Fourth IEEE Conference on Computer Vision and Pattern Recognition, 2011. 3, 9

    work page 2011

  5. [4]

    Virtualmodel: Gener- ating object-id-retentive human-object interaction image by diffusion model for e-commerce marketing.arXiv preprint arXiv:2405.09985, 2024

    Binghui Chen, Chongyang Zhong, Wangmeng Xiang, Yifeng Geng, and Xuansong Xie. Virtualmodel: Gener- ating object-id-retentive human-object interaction image by diffusion model for e-commerce marketing.arXiv preprint arXiv:2405.09985, 2024. 1, 2, 5

    arXiv 2024

  6. [6]

    Blender - a 3d modelling and rendering package, 2018

    Blender Online Community. Blender - a 3d modelling and rendering package, 2018. 4, 10

    work page 2018

  7. [7]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Shhamir Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InAd- vances in Neural Information Processing Systems (NeurIPS),

    work page

  8. [8]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. InProceedings of the 31st International Conference on Neural Information Processing Systems, page 6629–6640, Red Hook, NY , USA, 2017. Curran Associates Inc. 2, 4, 10

    work page 2017

Show all 49 references
  1. [9]

    Cock- tail: Mixing multi-modality control for text-conditional im- age generation

    Minghui Hu, Jianbin Zheng, Daqing Liu, Chuanxia Zheng, Chaoyue Wang, Dacheng Tao, and Tat-Jen Cham. Cock- tail: Mixing multi-modality control for text-conditional im- age generation. InThirty-seventh Conference on Neural In- formation Processing Systems, 2023. 2

    2023

  2. [10]

    Imagen 3, 2024

    Imagen-Team-Google. Imagen 3, 2024. 2

    2024

  3. [11]

    Karras, S

    T. Karras, S. Laine, and T. Aila. A style-based generator ar- chitecture for generative adversarial networks.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 43 (12):4217–4228, 2021. 3, 9

    2021

  4. [12]

    Multi-concept customization of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3

    2023

  5. [13]

    BLIP-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing

    Dongxu Li, Junnan Li, and Steven Hoi. BLIP-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing. InThirty-seventh Conference on Neural Information Processing Systems, 2023. 3

    2023

  6. [14]

    Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models. InPro- ceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 3

    2023

  7. [15]

    Controlnet++: Improving conditional controls with efficient consistency feedback

    Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaon- ing Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. InEuropean Conference on Computer Vision,

  8. [16]

    Gligen: Open-set grounded text-to-image generation.CVPR,

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jian- wei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation.CVPR,

  9. [17]

    Belongie, Lubomir D

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context.CoRR, abs/1405.0312,

    arXiv

  10. [18]

    Structure matters: Tackling the semantic discrepancy in diffusion models for image inpainting

    Haipeng Liu, Yang Wang, Biao Qian, Meng Wang, and Yong Rui. Structure matters: Tackling the semantic discrepancy in diffusion models for image inpainting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8038–8047, 2024. 2

    2024

  11. [19]

    Shadow Generation for Composite Im- age Using Diffusion Model

    Qingyang Liu, Junqi You, Jianting Wang, Xinhao Tao, Bo Zhang, and Li Niu. Shadow Generation for Composite Im- age Using Diffusion Model . In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8121–8130, Los Alamitos, CA, USA, 2024. IEEE Computer S...

    2024

  12. [20]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations, 2019. 4

    2019

  13. [21]

    Learning a no-reference quality metric for single-image super-rolution.Computer Vision and Image Understanding, pages 1–16, 2017

    Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming- Hsuan Yang. Learning a no-reference quality metric for single-image super-rolution.Computer Vision and Image Understanding, pages 1–16, 2017. 5, 11

    2017

  14. [22]

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. InAAAI, pages 4296–4304. AAAI Press,

  15. [23]

    Insertdiffusion: Identity preserving visualization of ob- jects through a training-free diffusion architecture, 2024

    Phillip Mueller, Jannik Wiese, Ioan Craciun, and Lars Mikel- sons. Insertdiffusion: Identity preserving visualization of ob- jects through a training-free diffusion architecture, 2024. 1, 2

    2024

  16. [24]

    Ava: A large-scale database for aesthetic visual analysis

    Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408–2415, 2012. 11

    2012

  17. [25]

    Gpt-4 technical report.ArXiv, abs/2303.08774,

    OpenAI. Gpt-4 technical report.ArXiv, abs/2303.08774,

    arXiv

  18. [26]

    Pois- son image editing

    Patrick P ´erez, Michel Gangnet, and Andrew Blake. Pois- son image editing. InACM SIGGRAPH 2003 Papers, page 313–318, New York, NY , USA, 2003. Association for Com- puting Machinery. 1, 2, 9

    2003

  19. [27]

    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. InThe Twelfth Interna- tional Conference on Learning Representations, 2024. 1, 2, 3, 4

    2024

  20. [28]

    Unicontrol: A unified diffusion model for controllable visual generation in the wild

    Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. Unicontrol: A unified diffusion model for controllable visual generation in the wild. InThirty-seventh Conferenc...

    2023

  21. [29]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the...

    2021

  22. [30]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InProceedings of the 38th International Conference on Machine Learning, pages 8821–8831. PMLR, 2021. 2

    2021

  23. [31]

    Generative ad- versarial text to image synthesis

    Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Lo- geswaran, Bernt Schiele, and Honglak Lee. Generative ad- versarial text to image synthesis. InProceedings of The 33rd International Conference on Machine Learning, pages 1060–1069, New York, New York, USA, 2016. PMLR. 2

    2016

  24. [32]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 1, 2, 5, 9

    2022

  25. [33]

    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. 1, 3, 7

    2023

  26. [34]

    Sara Mahdavi, Raphael Gontijo-Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Lit, Jay Whang, Emily 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 ...

    2024

  27. [35]

    LAION-5b: An open large-scale dataset for train- ing next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev....

    2022

  28. [36]

    Anycontrol: Create your artwork with versatile con- trol on text-to-image generation, 2024

    Yanan Sun, Yanchen Liu, Yinhao Tang, Wenjie Pei, and Kai Chen. Anycontrol: Create your artwork with versatile con- trol on text-to-image generation, 2024. 1, 2

    2024

  29. [37]

    Going deeper with convo- lutions

    Christian Szegedy, Wei Liu, Yang Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Van- houcke, and Andrew Rabinovich. Going deeper with convo- lutions. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9,...

    2015

  30. [38]

    Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018

    Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018. 5, 11

    2018

  31. [39]

    Ai-generated image quality as- sessment in visual communication.Proceedings of the AAAI Conference on Artificial Intelligence, 39(7):7392– 7400, 2025

    Yu Tian, Yixuan Li, Baoliang Chen, Hanwei Zhu, Shiqi Wang, and Sam Kwong. Ai-generated image quality as- sessment in visual communication.Proceedings of the AAAI Conference on Artificial Intelligence, 39(7):7392– 7400, 2025. 12

    2025

  32. [40]

    Ex- ploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring clip for assessing the look and feel of images. In AAAI, 2023. 11

    2023

  33. [41]

    Holistically-nested edge detection

    Saining ”Xie and Zhuowen” Tu. Holistically-nested edge detection. InProceedings of IEEE International Conference on Computer Vision, 2015. 4

    2015

  34. [42]

    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. InCVPR, 2024. 2

    2024

  35. [43]

    Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models, 2023

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models, 2023. 2

    2023

  36. [44]

    Shadow- enlightened image outpainting

    Hang Yu, Ruilin Li, Shaorong Xie, and Jiayan Qiu. Shadow- enlightened image outpainting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7850–7860, 2024. 2

    2024

  37. [45]

    Inpaint anything: Segment anything meets image inpainting.arXiv preprint arXiv:2304.06790, 2023

    Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting.arXiv preprint arXiv:2304.06790, 2023. 1, 2, 5, 7

    2023 arXiv

  38. [46]

    Stack- gan: Text to photo-realistic image synthesis with stacked generative adversarial networks

    Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiao- gang Wang, Xiaolei Huang, and Dimitris Metaxas. Stack- gan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In2017 IEEE International Conference on Computer Vision (ICCV), pages 5908–5916,

  39. [47]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836–3847, 2023. 1, 2, 3, 4, 5, 9

    2023

  40. [48]

    Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models.Advances in Neural Information Processing Sys- tems, 2023. 2

    2023

  41. [49]

    Bilateral refer- ence for high-resolution dichotomous image segmentation

    Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral refer- ence for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 3, 7

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.