Pith. sign in

REVIEW 4 major objections 6 minor 34 references

GENIE splits a reference image into a pure appearance map and an attribute residual, then renders that appearance onto a target while preserving structure, and reports top scores on object, garment, and person editing benchmarks.

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

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

T0 review · deepseek-v4-flash

2026-08-03 15:50 UTC pith:ZCGDLSCV

load-bearing objection A clean modular system with consistent reported gains, but the missing train/test split for AnyInsertion makes the SOTA claim unverified until clarified. the 4 major comments →

arxiv 2512.15138 v2 pith:ZCGDLSCV submitted 2025-12-17 cs.CV

Borrowing from anything: A generalizable framework for reference-guided instance editing

Shengxiao Zhou , Chenghua Li , Jianhao Huang , Qinghao Hu , Yifan Zhang This is my paper
classification cs.CV
keywords reference-guided image editingfeature disentanglementspatial alignmentadaptive residual scalingprogressive attention fusionlatent diffusionAnyInsertion datasetinstance editing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Reference-guided instance editing transfers a reference object's look onto a target image, but the reference's texture is entangled with its pose, scale, and lighting, so prior methods leak appearance or distort identity. GENIE is a diffusion-based framework built on the claim that this entanglement can be made explicit. Its three modules — a spatial alignment net, a per-channel residual scaling mask, and a three-stage progressive attention — learn what to borrow from the reference and how to apply it to the target. On the AnyInsertion dataset, the paper reports the best or second-best scores on reconstruction, perceptual, and semantic metrics across object, garment, and person categories, with the largest gains in PSNR and FID. If those numbers hold under proper held-out evaluation, the framework would be a new standard for disentanglement-based editing.

Core claim

The paper's central claim is that the editing bottleneck is semantic entanglement, and that a structured decomposition — spatial alignment, target-aware multiplicative scaling, and progressive attention fusion — achieves explicit disentanglement and thereby higher fidelity than monolithic adapters. The ARSM produces a per-channel modulation signal alpha = tanh(...) in (-1,1), applied as (1+alpha) to reference features, so positive values amplify intrinsic cues and negative values suppress extrinsic attributes. The PAF then fuses the purified features through three attentions (structural, synergistic, appearance) with learnable weights, first stabilizing the target structure, then exploring s

What carries the argument

The load-bearing mechanism is the Adaptive Residual Scaling Module (ARSM), which computes a target-aware spatial scaling map alpha = tanh(f_scale(F_r ⊕ F_t)) and applies it as F'_r = (1 + alpha) ⊙ F_r, constraining the modulation to (0,2) for continuous suppression-to-enhancement of each feature channel. The Progressive Attention Fusion (PAF) complements it: three attentions (structural, synergistic, appearance) are mixed by learnable scalars to render the purified appearance onto the target without disturbing its layout. Together they turn a generic dual U-Net diffusion model into an explicit what-to-borrow / how-to-render pipeline.

Load-bearing premise

The load-bearing premise is that the AnyInsertion evaluation is held out from training — the paper never states a train/test split — and, architecturally, that per-channel multiplicative scaling can truly separate intrinsic appearance from extrinsic attributes.

What would settle it

Re-run the AnyInsertion evaluation on a provably disjoint held-out split (with the same checkpoints and seeds) and observe whether the PSNR/SSIM/FID advantages over InsertAnything and MimicBrush persist; or probe the ARSM mask by comparing alpha values for images with altered pose/lighting but identical texture and check whether the expected channels flip sign.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the disentanglement claim holds, reference-guided editing can transfer fine textures and patterns without identity distortion or appearance leakage from the reference.
  • The reported 2 dB PSNR improvement and large FID reductions over the strongest baselines suggest that explicit feature decomposition is more effective than the implicit conditioning used by adapter-based methods.
  • The ablation indicates the modules address distinct failure modes: SAM matters most for person images, PAF for garments, and ARSM for objects, pointing to a design recipe for other editing tasks.
  • Freezing the reference U-Net and IP-Adapter while fine-tuning only the target U-Net is claimed to preserve pre-trained feature quality, a practical training-strategy takeaway.

Where Pith is reading between the lines

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

  • If the evaluation protocol lacks a disjoint train/test split (as the paper's Section 4.1.1 and 4.2.1 suggest), the reported SOTA margins could be inflated; a held-out re-run with published seeds would settle the practical gain.
  • The same per-channel residual scaling idea could transfer to video editing, where temporal extrinsic attributes like camera motion might be suppressed while intrinsic appearance is enhanced.
  • One testable prediction: probing the sign of alpha across input pairs should show negative channels aligned with pose/lighting variation and positive channels with texture/pattern; this would directly validate the disentanglement claim.
  • The framework's modularity suggests extensions to other reference-conditioned generation tasks, such as virtual try-on, where pose/scale entanglement is a known failure mode.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes GENIE, a dual U-Net diffusion framework for reference-guided instance editing. Three modules are introduced: a Spatial Alignment Module (SAM) that predicts an affine warp to canonicalize reference features; an Adaptive Residual Scaling Module (ARSM) that learns a spatial scaling map to modulate reference features; and a Progressive Attention Fusion (PAF) mechanism that fuses the modulated reference features into the target U-Net via structural, synergistic, and appearance attention. The method is trained for 10 epochs on the AnyInsertion dataset and evaluated on Object, Garment, and Person categories against AnyDoor, Paint-by-Example, MimicBrush, OOTDiffusion, and InsertAnything. The paper reports state-of-the-art or near-state-of-the-art results on PSNR, SSIM, LPIPS, CLIP, DINO, DreamSim, and FID, plus component and training-strategy ablations. A code repository link is provided.

Significance. If the reported results are valid, GENIE would be a strong empirical contribution to reference-guided instance editing, with a plausible modular decomposition of alignment, appearance purification, and fusion. The paper includes comparisons with several baselines on three categories and provides code, which are assets for reproducibility. However, the central claims — state-of-the-art fidelity and 'explicit disentanglement' — are not yet established because the evaluation protocol omits the train/test split and the disentanglement claim is not directly measured. The missing split is load-bearing: diffusion models can memorize training images, so every headline metric in Table 1 could be inflated if evaluation images overlap with training images.

major comments (4)
  1. [§4.1.1 and §4.2.1] The manuscript states that the model is trained for 10 epochs on the AnyInsertion dataset and then evaluated 'on the AnyInsertion dataset,' but it never reports a train/test split, random seed, or an overlap check between training and evaluation images. This is load-bearing for the central SOTA claim: if the evaluation set is not disjoint from the training set, the reported PSNR, SSIM, LPIPS, CLIP, DINO, DreamSim, and FID numbers — including the 2 dB PSNR gains and the 7–16 point FID reductions — can be inflated by memorization. The authors must specify the exact split, the number of evaluation samples, and confirm that no training images appear in the evaluation set. Without this, the headline comparison in Table 1 cannot be accepted.
  2. [Table 2] The ablation baseline 'B' is never defined. Its Object row (PSNR 23.96, SSIM 0.874, LPIPS 0.092, CLIP 93.40, DINO 93.48, DreamSim 0.081, FID 79.72) exactly reproduces the MimicBrush row in Table 1 (PSNR 23.9651, SSIM 0.8740, LPIPS 0.0925, CLIP 93.40, DINO 93.48, DreamSim 0.0819, FID 79.7223). If B is MimicBrush, that is a critical fact about the experimental setup and must be stated explicitly; otherwise the attribution of gains to SAM/PAF/ARSM is unclear. The same issue affects the interpretation of Table 3, whose checkmark semantics are also not defined in the caption.
  3. [§3.3 and Table 2] The paper repeatedly claims that GENIE achieves 'explicit disentanglement' of intrinsic appearance and extrinsic attributes, but no experiment directly tests this. ARSM's residual scaling operation in Eq. (3) is a learned modulation; the paper does not visualize or analyze the predicted α map, nor does it probe whether F'_r actually contains less pose, scale, or illumination information than F_r. The end-task metrics in Table 2 support overall fidelity improvements but not the specific disentanglement interpretation. To sustain the 'explicit disentanglement' claim, the authors should provide intervention or feature-level evidence, e.g., showing that varying the reference's extrinsic attributes leaves F'_r largely unchanged, or that α selectively suppresses known attribute directions.
  4. [§4.3.1 and Table 2] The ablation narrative is selective. When discussing SAM, the paper highlights the large FID improvement on Person (124.05 to 93.30) but omits that Person PSNR drops from 24.09 to 23.46 and SSIM drops from 0.837 to 0.826. When discussing ARSM, the paper reports a 'marginal' 0.14 PSNR gain on Object but does not mention that DreamSim worsens from 0.062 to 0.065 and that LPIPS and SSIM are flat. Since the modules are claimed to work synergistically, all metrics should be reported consistently in the text, and the paper should include error bars or significance tests for at least the main comparisons, given that several reported differences are small (e.g., CLIP 94.28 vs. 94.13 in Table 1).
minor comments (6)
  1. [Table 1] The row label 'A verage (Ours)' appears to be a formatting artifact; it should read 'Ours' or 'Average (Ours)' consistently.
  2. [§3.4] The weights β, γ, λ in Eq. (4) are described as 'dynamic weighted fusion,' but they are learned scalar constants, not input-dependent dynamic weights. Please rephrase to avoid overclaiming.
  3. [Figure 1] The α map is drawn with shape (B, L, 1) while F_r is (B, L, C). Clarify whether the scaling is a single scalar per spatial token broadcast across channels, or a per-channel vector. The text in §3.3 should match the figure.
  4. [§4.1.2 and §4.2.1] Provide details on the evaluation protocol: number of test samples, FID reference set and whether it is the same as the generated set, CLIP and DINO backbone/layer choices, and whether baseline numbers are re-run under identical conditions or taken from prior papers.
  5. [References] Minor typos: 'V AE' in §3.1 should be 'VAE'; 'PmLR' in reference [26] should be 'PMLR'. Also, references [6] and [8] are cited in the introduction but not discussed in Related Work; consider aligning the citations.
  6. [§4.1.1] The paper states training is 10 epochs without mentioning checkpoint selection, early stopping, or validation-based model selection. Please specify how the final model was chosen.

Circularity Check

0 steps flagged

No circular derivation; the SOTA claim is an empirical evaluation, and the unresolved train/eval overlap is a data-hygiene risk, not a circularity.

full rationale

GENIE's derivation chain is entirely empirical: Eq. (1) is a standard noise-prediction objective, Eq. (2) is a differentiable spatial warp, Eq. (3) defines a learned modulation map, and Eq. (4) is a learned weighted fusion. None of these equations is defined in terms of a target result, and no fitted parameter is renamed as a prediction. The 'explicit disentanglement' claim is supported by ablations (Table 2) and by comparisons with external baselines (Tables 1-3), not by a definitional equivalence. There are no self-citations used as load-bearing evidence, no imported uniqueness theorem, and no ansatz smuggled in via citation. The main validity concern is that Section 4.1.1 reports training for 10 epochs on AnyInsertion while Section 4.2.1 evaluates on AnyInsertion without stating a held-out split; if evaluation images overlap training images, Table 1 numbers could be inflated by memorization. This is a serious experimental-protocol omission and a correctness/evaluation risk, but it is not a circularity of the kind defined here: the paper does not show that any equation reduces to another by construction, and the absence of a split is not itself a derivation. Per the hard rules, evaluation hygiene concerns belong to correctness risk, not to the circularity score. Thus the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

GENIE's central contribution is a learned pipeline; the 'disentanglement' property is asserted from benchmark improvements rather than measured directly. Each module introduces learned parameters fitted on the same dataset used for evaluation, and no independent evidence (e.g., feature probing, perturbation tests, or a dedicated disentanglement metric) supports the disentanglement claim.

free parameters (3)
  • ARSM scaling map α (network weights and tanh output) = learned; no explicit numeric values reported
    α = tanh(f_scale(Fr ⊕ Ft)) modulates reference features per location; all parameters fitted on AnyInsertion training set with no independent constraint enforcing disentanglement.
  • PAF fusion weights β, γ, λ = learned scalars
    Eq. (4) combines structural, synergistic, and appearance attention with learnable scalar weights fitted end-to-end; no validation outside the benchmark.
  • SAM localization network parameters = learned
    Predicts a 2D affine transformation for feature warping (§3.2); trained jointly with no geometric supervision or alignment ground truth stated.
axioms (5)
  • domain assumption Pre-trained Stable Diffusion U-Net and IP-Adapter provide reliable feature priors for the reference and target branches.
    Used in §3.1 and frozen in the optimal training configuration (Table 3, row 6); no analysis of failure modes of these pretrained components.
  • domain assumption AnyInsertion provides reliable ground-truth and masks for reconstruction-based evaluation.
    Sec. 4.1 uses AnyInsertion without an explicit split or error analysis of dataset labels/masks.
  • ad hoc to paper Scalar multiplicative modulation of full-channel features can separate intrinsic from extrinsic attributes.
    ARSM (§3.3) is the core disentanglement mechanism, asserted with no formal proof or independent probing evidence.
  • standard math The standard noise-prediction diffusion loss is sufficient for the editing objective.
    Eq. (1) uses the standard LDM objective; no task-specific loss term encourages disentanglement.
  • domain assumption PSNR, SSIM, LPIPS, CLIP, DINO, DreamSim, and FID faithfully measure editing fidelity and disentanglement.
    Sec. 4.1.2 lists these metrics; no direct disentanglement metric, robustness test, or human study is reported.
invented entities (2)
  • Disentangled reference feature F'_r no independent evidence
    purpose: Represent the reference with extrinsic attributes suppressed; injected into the target U-Net for rendering.
    Produced by ARSM; no external falsifiable check that it is actually disentangled—only indirect benchmark improvements.
  • Progressive Attention Fusion (PAF) hybrid feature pool F_cat no independent evidence
    purpose: Combine structural and appearance features for progressive rendering.
    Internal architectural construct; its benefit is supported only by ablation scores, not by an independent probe.

pith-pipeline@v1.3.0-alltime-deepseek · 8030 in / 12202 out tokens · 121289 ms · 2026-08-03T15:50:57.504210+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Borrowing from anything: A generalizable framework for reference-guided instance editing." pith.science (2026). https://pith.science/paper/ZCGDLSCV

@misc{pith2026251215138,
  author       = {Pith},
  title        = {Pith review of: Borrowing from anything: A generalizable framework for reference-guided instance editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZCGDLSCV}},
  note         = {Machine review of arXiv:2512.15138}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reference-guided instance editing is fundamentally limited by semantic entanglement, where a reference's intrinsic appearance is intertwined with its extrinsic attributes. The key challenge lies in disentangling what information should be borrowed from the reference, and determining how to apply it appropriately to the target. To tackle this challenge, we propose GENIE, a Generalizable Instance Editing framework capable of achieving explicit disentanglement. GENIE first corrects spatial misalignments with a Spatial Alignment Module (SAM). Then, an Adaptive Residual Scaling Module (ARSM) learns what to borrow by amplifying salient intrinsic cues while suppressing extrinsic attributes, while a Progressive Attention Fusion (PAF) mechanism learns how to render this appearance onto the target, preserving its structure. Extensive experiments on the challenging AnyInsertion dataset demonstrate that GENIE achieves state-of-the-art fidelity and robustness, setting a new standard for disentanglement-based instance editing.

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

34 extracted references · 12 linked inside Pith

  1. [1]

    However, this task faces a fundamental challenge: semantic entanglement

    INTRODUCTION The essence of reference-guided instance editing [1–8] is to transfer the intrinsic appearance of a reference object, such as texture and patterns, to achieve fine-grained manipulation of a target. However, this task faces a fundamental challenge: semantic entanglement. In the reference features, the desired intrinsic appearance is inherently...

  2. [2]

    RELA TED WORK Recent years have witnessed rapid progress in reference-guided im- age editing using diffusion models. Early methods such as Paint-by- Example [2] and MimicBrush [1] explored example-based or zero- shot semantic editing, enabling object replacement, style imitation, and seamless blending. For object insertion and composition, re- cent works ...

  3. [3]

    Overall Framework As illustrated in Fig

    METHOD 3.1. Overall Framework As illustrated in Fig. 1, we present GENIE, a dual U-Net latent diffu- sion model [15], comprising a reference branch for appearance dis- entanglement and a target branch for denoising synthesis. It takes a reference imageI ref ∈R H×W×3 , a target imageI tar ∈R H×W×3 , and a binary maskM∈ {0,1} H×W as input, aiming to extract...

    Pith/arXiv arXiv 2025

  4. [4]

    Implementation Details 4.1.1

    EXPERIMENTS 4.1. Implementation Details 4.1.1. Hyperparameters The model is trained for 10 epochs on the AnyInsertion [3] dataset, with all images resized to a resolution of512×512. We employ the Adam optimizer [21] with a learning rate of1×10 −5. The training is conducted on 8 NVIDIA A100 GPUs, utilizing a batch size of 14 per GPU for an effective total ...

    arXiv

  5. [5]

    With mod- ules for spatial alignment, appearance purification, and high-fidelity fusion, our method effectively disentangles reference information

    CONCLUSION We propose a framework to address the fundamental challenge of feature entanglement in reference-guided image editing. With mod- ules for spatial alignment, appearance purification, and high-fidelity fusion, our method effectively disentangles reference information. Extensive experiments demonstrate the superiority of our method, which achieves...

  6. [6]

    ACKNOWLEDGMENTS This work was supported by the Natural Science Foundation of Jiangsu Province under Grant BK20243051

  7. [7]

    Zero-shot image editing with refer- ence imitation,

    X. Chen, Y . Feng, M. Chen, Y . Wang, S. Zhang, Y . Liu, Y . Shen, and H. Zhao, “Zero-shot image editing with refer- ence imitation,”Advances in Neural Information Processing Systems, vol. 37, pp. 84 010–84 032, 2024

    2024

  8. [8]

    Paint by example: Exemplar-based image editing with diffusion models,

    B. Yang, S. Gu, B. Zhang, T. Zhang, X. Chen, X. Sun, D. Chen, and F. Wen, “Paint by example: Exemplar-based image editing with diffusion models,” inProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 2023, pp. 18 381–18 391

    2023

  9. [9]

    Insert anything: Image insertion via in-context editing in dit,

    W. Song, H. Jiang, Z. Yang, R. Quan, and Y . Yang, “Insert anything: Image insertion via in-context editing in dit,”arXiv preprint arXiv:2504.15009, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on,

    Y . Xu, T. Gu, W. Chen, and C. Chen, “Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2025

    2025

  11. [11]

    Anydoor: Zero-shot object-level image customization,

    X. Chen, L. Huang, Y . Liu, Y . Shen, D. Zhao, and H. Zhao, “Anydoor: Zero-shot object-level image customization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 6593–6602

    2024

  12. [12]

    Paste, inpaint and harmonize via denoising: Subject-driven image editing with pre-trained diffusion model,

    X. Zhang, J. Guo, P. Yoo, Y . Matsuo, and Y . Iwasawa, “Paste, inpaint and harmonize via denoising: Subject-driven image editing with pre-trained diffusion model,”arXiv preprint arXiv:2306.07596, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Muse: Multi-subject unified synthesis via explicit layout semantic ex- pansion,

    F. Peng, J. Wu, Y . Li, T. Gao, D. Zhang, and H. Fu, “Muse: Multi-subject unified synthesis via explicit layout semantic ex- pansion,” inProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), 2025

    2025

  14. [14]

    Describe, don’t dictate: Semantic image editing with natural language intent,

    E. Ci, S. Guan, Y . Ge, Y . Zhang, W. Li, Z. Zhang, J. Yang, and Y . Tai, “Describe, don’t dictate: Semantic image editing with natural language intent,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025

    2025

  15. [15]

    Objectstitch: Generative object com- positing,

    Y . Song, Z. Zhang, Z. Lin, S. Cohen, B. Price, J. Zhang, S. Y . Kim, and D. Aliaga, “Objectstitch: Generative object com- positing,”arXiv preprint arXiv:2212.00932, 2022

    Pith/arXiv arXiv 2022

  16. [16]

    Completeme: Reference-based human image completion,

    Y .-J. Tsai, B. Price, Q. Liu, L. Figueroa, D. Pakhomov, Z. Ding, S. Cohen, and M.-H. Yang, “Completeme: Reference-based human image completion,”arXiv preprint arXiv:2504.20042, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Freeedit: Mask-free reference-based image editing with multi-modal instruction,

    R. He, K. Ma, L. Huang, S. Huang, J. Gao, X. Wei, J. Dai, J. Han, and S. Liu, “Freeedit: Mask-free reference-based image editing with multi-modal instruction,”arXiv preprint arXiv:2409.18071, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Imagdressing-v1: Customizable virtual dressing,

    F. Shen, X. Jiang, X. He, H. Ye, C. Wang, X. Du, Z. Li, and J. Tang, “Imagdressing-v1: Customizable virtual dressing,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2025

    2025

  19. [19]

    Improving diffusion models for authentic virtual try-on in the wild,

    Y . Choi, S. Kwak, K. Lee, H. Choi, and J. Shin, “Improving diffusion models for authentic virtual try-on in the wild,”arXiv preprint arXiv:2403.05139, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Efficientmt: Ef- ficient temporal adaptation for motion transfer in text-to-video diffusion models,

    Y . Cai, H. Han, Y . Wei, S. Shan, and X. Chen, “Efficientmt: Ef- ficient temporal adaptation for motion transfer in text-to-video diffusion models,”arXiv preprint arXiv:2503.19369, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Om- mer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  22. [22]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” inProceedings of the 2nd International Conference on Learning Representations (ICLR), 2014

    2014

  23. [23]

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

    H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang, “Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models,”arXiv preprint arxiv:2308.06721, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Spatial transformer networks,

    M. Jaderberg, K. Simonyan, A. Zisserman, and K. Kavukcuoglu, “Spatial transformer networks,” inAd- vances in Neural Information Processing Systems, vol. 28, 2015, pp. 2017–2025

    2015

  25. [25]

    Smartcontrol: Enhancing controlnet for handling rough visual conditions,

    X. Liu, Y . Wei, M. Liu, X. Lin, P. Ren, X. Xie, and W. Zuo, “Smartcontrol: Enhancing controlnet for handling rough visual conditions,”arXiv preprint arXiv:2404.06451, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  27. [27]

    Adam: A method for stochastic opti- mization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic opti- mization,”arXiv preprint arXiv:1412.6980, 2017

    Pith/arXiv arXiv 2017

  28. [28]

    The unreasonable effectiveness of deep features as a percep- tual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a percep- tual metric,” inProceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2018

    2018

  29. [29]

    Dreamsim: Learning new dimensions of hu- man visual similarity using synthetic data,

    S. Fu, N. Tamir, S. Sundaram, L. Chai, R. Zhang, T. Dekel, and P. Isola, “Dreamsim: Learning new dimensions of hu- man visual similarity using synthetic data,”arXiv preprint arXiv:2306.09344, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  31. [31]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    2021

  32. [32]

    Learning transferable visual models from natural language su- pervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language su- pervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  33. [33]

    Image quality metrics: Psnr vs. ssim,

    A. Hore and D. Ziou, “Image quality metrics: Psnr vs. ssim,” in2010 20th International Conference on Pattern Recognition (ICPR). IEEE, 2010, pp. 2366–2369

    2010

  34. [34]

    Im- age quality assessment: from error visibility to structural simi- larity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Im- age quality assessment: from error visibility to structural simi- larity,”IEEE Transactions on Image Processing (TIP), vol. 13, no. 4, pp. 600–612, 2004

    2004