Pith. sign in

REVIEW 4 major objections 4 minor 37 references

A diffusion model that restores damaged film frame-by-frame at high resolution while preserving scene structure, trained on real restorations and requiring only one 24GB-VRAM GPU.

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 13:12 UTC pith:ZZGNNDZQ

load-bearing objection A coherent architecture write-up and a plausible dataset idea, but the paper's central empirical claim is entirely unrepresented — empty experiment sections and no numbers — so the contribution cannot currently be evaluated. the 4 major comments →

arxiv 2512.24946 v3 pith:ZZGNNDZQ submitted 2025-12-31 cs.CV cs.AIcs.MM

HaineiFRDM: Structure-Preserving Diffusion for Film Restoration under Fast Motion and Diverse Defects

Rongji Xun , Junjie Yuan , Zhongjie Wang This is my paper
classification cs.CV cs.AIcs.MM
keywords film restorationvideo diffusion modelpatch-based generationdefect removalfast motionhigh-resolution restorationfilm degradation datasetdiffusion priors
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.

This paper proposes HaineiFRDM, a video-diffusion model for restoring old film. It claims that a diffusion model's learned understanding of visual content lets it remove dust, scratches, and other defects even in fast-motion shots, without the optical-flow warping that makes previous methods erase limbs or distort structure. To make high-resolution work feasible on a single consumer 24GB-VRAM GPU, the model is trained and run on overlapping patches, with position-aware global prompt and frame fusion to keep patches coherent and a frequency-based texture module to preserve original texture. The authors also build a training dataset that pairs real before-after restoration videos with synthetic defect templates, and they claim their method outperforms existing open-source methods on both quantitative metrics and qualitative inspection. A sympathetic reader would care because the method targets two bottlenecks that block practical deployment: fast-motion structural errors and hardware requirements.

Core claim

Central claim: scene-structure damage in restored film comes not from the defects themselves but from faulty motion modeling, so a restoration model should not rely on optical flow. Instead, HaineiFRDM lets a pretrained video diffusion model use its intrinsic content-understanding to recognize and remove degradations while leaving clean content untouched. To scale to 2K/4K, the authors introduce patch-wise training and inference: each patch is restored with the help of whole-frame visual and textual information—via a position-aware global-frame fusion module and a global-prompt fusion module—so local patches stay consistent with the global scene. A patch-consistent inference framework first

What carries the argument

The load-bearing object is the Restoration-Guidance Network, a trainable side branch that injects learned residuals into a frozen pretrained video-diffusion U-Net, steering denoising toward the restored frame. Four mechanisms carry the argument. Global-frame fusion resizes the whole degraded frame, encodes it with a learned image encoder, adds the patch's bounding-box coordinates through a Fourier embedder, and injects this position-aware global feature via cross-attention, so a local patch can distinguish defect from content. Global-prompt fusion repeats this for a whole-frame text caption, letting the model exploit scene understanding. A texture-reconstruction module works in the 3D-FFT fr

Load-bearing premise

The real-degraded half of the training set is trustworthy: the before-after restoration videos are pixel-aligned and the defect masks, computed by frame differencing and filtered frame-by-frame by hand, mark actual defects rather than valid film content.

What would settle it

On a held-out real restoration video with fast motion, run the model and have human annotators flag any restored area where scene content (limbs, faces, text) changed shape or vanished; if the erasure rate matches or exceeds that of optical-flow baselines, the central content-understanding claim fails. A cheaper check: sample the training masks, compare them to human-labeled defects, and see whether frame-differencing masks have high precision—low precision would mean the model was trained to delete valid content.

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

If this is right

  • Fast-motion shots should no longer show limb disappearance or structural distortion, because no optical-flow warping is used at any stage.
  • A single 24GB-VRAM consumer GPU can restore high-resolution (2K/4K) frames, since the diffusion model runs on patches with features reused across patches.
  • Defect restoration is decoupled from global color and deblur adjustments, so the restored film preserves the director's intended look rather than being aggressively 'enhanced'.
  • The combination of real restored films and synthetic defect templates provides a training set that should transfer to unseen films better than purely synthetic data.
  • The global-residual inference framework should suppress blocking artifacts when full frames are assembled from patches.

Where Pith is reading between the lines

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

  • If the content-understanding claim holds, the method should extend to non-scratch degradations (dye fading, frame loss, flicker) by retraining or fine-tuning on paired data, because the architecture contains no explicit defect detector that would need redesign.
  • The dataset construction bottleneck suggests that real-world performance is gated by the quality and scale of before-after restoration pairs; collecting more pairs with verifiable masks is the most direct test of generalization.
  • The patch-consistent inference scheme, with global-residual fusion and cross-patch feature reuse, is a general recipe that could transfer to other high-resolution video restoration tasks, such as VHS or archival footage.
  • A testable extension: measure restoration consistency on scenes where content repeats across shots; the position-aware fusion should keep repeated objects identical, a failure point for optical-flow baselines.

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 / 4 minor

Summary. The paper proposes HaineiFRDM, a patch-based latent diffusion framework for high-resolution film restoration. The method combines a ControlNet-style restoration-guidance network with global prompt/frame fusion modules, a frequency-domain texture-reconstruction module, and a patch-consistent inference procedure that fuses global residuals according to a cosine weighting schedule (Eqs. 1-2). The authors also introduce a dataset containing real-degraded film clips paired with restored frames and synthetic degradation generated from human-labeled defect templates. The abstract and introduction make strong empirical claims: superior restoration quality over existing open-source methods, strong structural consistency under fast motion, and training/inference on a single 24GB-VRAM consumer GPU.

Significance. If the empirical claims were supported, the paper would offer a meaningful step toward practical diffusion-based film restoration on consumer hardware. The architecture description is coherent and several design choices—position-aware global fusion, frequency-based texture conditioning, and global-residual inference weighting—are technically plausible. The dataset proposal, including categorized defect templates and real restoration-laboratory footage, is potentially valuable to the community. However, because the experimental section is almost entirely empty, no conclusion about the method's actual performance can be drawn. The contribution is therefore currently an unvalidated system description rather than a demonstrated advance.

major comments (4)
  1. [§6.2, §6.3, abstract] The central claim of 'superior restoration quality' and 'comprehensive experimental results conclusively demonstrate superiority' is unsupported: both quantitative subsections (6.2, 6.3) are empty headings. No PSNR, SSIM, LPIPS, BRISQUE, or FVD values are reported anywhere, and no comparison against RTN, DeepRemaster, TAPE, or Mamba-based baselines appears. Without any numeric results, the reader cannot verify the abstract's claims, and the claim itself is load-bearing for the paper's contribution.
  2. [§6.4, §6.5] The qualitative evaluation (6.4) and ablation study (6.5) are placeholder text, not results. Section 6.4 merely states 'show pathe newsreel results here'; no figures or qualitative comparisons are included. Section 6.5 lists ablation topics (6.5.1-6.5.5) but provides no metrics, no visual comparisons, and no discussion of the effect of removing components. The only quantitative statement, 'to reach PSNR=20dB... global prompt uses 12w iterations... enlarged fixed grid prompt costing 20w iterations,' is a training-convergence observation, not a restoration-quality metric, and it cannot support the claim that the proposed modules improve output quality.
  3. [§5 and §3 (real-degraded data)] The real-degraded training set is constructed by frame-differencing 'before-after' comparison videos and computing binary defect masks from paired frames. The text concedes a 'pixel-mismatch problem' due to deflickering/color toning and relies on manual frame-by-frame filtering to obtain a 5-minute training set. No independent validation of mask quality is provided. If misaligned pairs or incorrect masks remain, the model would be trained to erase valid content—the exact limb-disappearance failure attributed to optical-flow methods. Given that the paper's real-world generalization claims rest partly on this data, the absence of any mask-quality analysis is a substantive gap.
  4. [§7] The limitations listed in the conclusion directly qualify the paper's headline claims. The model 'still has some trouble recognizing and repainting for some constant lines scratches,' the synthetic pipeline 'may still suffer from generating LQ and GT pair with color distortion,' generation is 'slow,' and there is 'loss in exquisite details.' These are not minor caveats; together they undermine the advertised 'strong structural consistency' and 'superior restoration quality.' The conclusions should be calibrated to what is actually demonstrated, which—given the empty experimental sections—is nothing quantitative.
minor comments (4)
  1. [Throughout] There are numerous typos and formatting issues: 'dateset' (Sec. 3), 'LIPIPS' (Sec. 6.1.2), 'radom-sampled' (Eq. 3), 'freuqency' (Fig. 1 caption), and 'emplyed' (Sec. 1). Figures 4 and 5 are labeled only 'Example of a short caption' and do not show actual restored outputs.
  2. [§3 and §5] Sections 3 and 5 are nearly identical in content; this duplication should be removed or one section should be renamed and made the definitive dataset description.
  3. [§4.3] The loss weights α_p=1 and α_d=81 are introduced without any sensitivity analysis; a brief justification or ablation would be helpful (though not decisive given the larger experimental gap).
  4. [References] Reference [7] is truncated ('with no $'), and reference [19] mixes the RVRT paper (Liang et al.) with a shifted-window citation; the numbering is inconsistent with the citation context.

Circularity Check

0 steps flagged

No circular derivation found; central claims are unverified but not circular.

full rationale

I inspected the claimed derivation chain: the method is a ControlNet-style diffusion model with global fusion, frequency modules, and a patch-consistent inference scheme. Nothing in the paper defines a quantity in terms of the target prediction, fits a parameter and then reports it as a prediction, or imports a load-bearing result from the authors' own prior work (there are no self-citations). The training losses (Eq. 3-6) are standard supervised losses; the defect masks are generated from paired degraded/restored frames or from synthetic templates, so using them in the defect loss is supervised training rather than circular prediction. The inference-time global residual in Sec. 4.2 is produced by the same model from downsampled frames and then used as a reference; this is a coarse-to-fine bootstrapping design, not a reduction of the final output to its own input. The real circularity-adjacent concerns are evidentiary, not definitional: the quantitative sections are empty headings, the evaluation would run on the authors' own synthetic and self-collected data, and the Section 7 limitations undercut the empirical claims. These are serious credibility and validity problems, but they do not satisfy the required standard of exhibiting a specific equation or fitted parameter that is equivalent to its input by construction. Accordingly, the honest circularity finding is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No new physical or representational entities are introduced; the Global Prompt / Global Frame Fusion / Texture-Reconstruction modules are architectural components, and the defect templates are data artifacts, so independent-evidence assessment is not applicable. The central claim rests on 3 hand-chosen parameters and 5 domain assumptions listed above.

free parameters (3)
  • alpha_p (preprocess loss weight) = 1
    Eq. (6): 'we empirically use alpha_p = 1 and alpha_d = 81' - hand-chosen, no sensitivity analysis (the corresponding ablation is absent).
  • alpha_d (defect loss weight) = 81
    Eq. (6): empirical choice, no justification or ablation.
  • global-residual weight coefficient in GRFM = 0.25
    Eq. (2): c_t = 0.25*(1 + cos(pi*(T-t)/T)); the 0.25 scale and cosine schedule are chosen without derivation or ablation.
axioms (5)
  • domain assumption A pretrained video diffusion model's content prior transfers to film-defect restoration through ControlNet-style residual injection.
    Section 4.1: Restoration-Guidance Network injects learned residuals into a pretrained UNet; the paper offers no diagnostic that the prior 'understands' scratch/dust degradation.
  • domain assumption CLIP image/text embeddings of the full degraded frame, fused with Fourier-encoded patch coordinates, provide sufficient global context to resolve patch-level ambiguity.
    Section 4.1.1 Global Fusion Module; rests on CLIP's patch-token quality for dense restoration cues, not established for film defects.
  • domain assumption Binary defect masks computed by paired-frame differencing on real restoration videos are accurate after manual filtering.
    Sections 3/5: the paper admits pixel-mismatch and remaining defects in source videos; mask quality is never independently measured.
  • domain assumption Hand-labeled defect templates sampled by appearance probability approximate the distribution of real film defects.
    Section 3 synthetic workflow; no distributional comparison between synthetic and real defects is provided.
  • standard math 3D FFT/iFFT of latent features is a valid invertible domain for texture-consistency constraints.
    Section 4.1.2 uses 3D-FFT, MLP on real/imag parts, cross-attention, 3D-iFFT; invertibility is standard, but the claim that cross-attention in the frequency domain preserves texture is empirical, not guaranteed.

pith-pipeline@v1.3.0-alltime-deepseek · 10411 in / 16921 out tokens · 163757 ms · 2026-08-03T13:12:20.115974+00:00 · methodology

0 comments
read the original abstract

Existing film-restoration methods frequently fail under fast motion, producing limb disappearance and structural distortion due to inaccurate motion modeling. Moreover, high-resolution restoration under spatially-persistent and mixed defects remains insufficiently studied. We propose HaineiFRDM, a Film Restoration Diffusion Model that leverages the content modeling capability of diffusion models for content-aware restoration, removing defects while preserving scene structure.To enable scalable high-resolution restoration, we adopt a patch-wise strategy with position-aware global fusion modules to maintain cross-patch coherence. We further introduce a frequency-based module to enhance texture consistency and a patch-consistent inference framework to alleviate blocking artifacts introduced by patch-based processing.We also construct a film restoration dataset comprising categorized defect templates, professionally restored films, and realistic synthetic degradations.Extensive experiments demonstrate our superior restoration quality with strong structural consistency. Our design also reduces memory requirements, enabling high-resolution restoration on a single 24GB-VRAM GPU.Code and the dataset will be released at https://anonymous.4open.science/r/HaineiFRDM.

Figures

Figures reproduced from arXiv: 2512.24946 by Junjie Yuan, Rongji Xun, Zhongjie Wang.

Figure 1
Figure 1. Figure 1: Training pipeline of our proposed HaineiFRDM. The model is input degraded patched frames and extracts each frame features with Preprocess Module. Then the frame features input into ControlNet, in which we use Global-Prompt-Fusion Module and Global-Frame-Fusion Module to help the model have a global-frame awareness and use freuqency mdoule to maintain origin frame textures. Lastly, the learned ControlNet re… view at source ↗
Figure 3
Figure 3. Figure 3: Example of a short caption, which should be centered. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Synthesize data workflow. intensive dust, cigarette-burn holes, flickering and constant scratches, etc. Furthermore, we randomly apply quality degradation(rescale,jpeg compression) and film texture sim￾ulation(add grain) to sampled frames to increase diversity . After that, we add color to defects in templates refer to defects on colored films and then fuse the frames and the defects to get the degraded fr… view at source ↗
Figure 4
Figure 4. Figure 4: Example of a short caption, which should be centered. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example of a short caption, which should be centered. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Synthesize data workflow. different camera angles and shot sizes. And we collected different types film defect templates including sparse and intensive dust, cigarette-burn holes, flickering and constant scratches, etc. Furthermore, we randomly apply quality degradation(rescale,jpeg compression) and film texture sim￾ulation(add grain) to sampled frames to increase diversity . After that, we add color to de… view at source ↗

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

37 extracted references · 3 linked inside Pith

  1. [1]

    https : / / www

    Diamant film restoration software. https : / / www . hs - art . com / index . php / research - main / deeprestore-menu. 2, 3, 6

  2. [2]

    https://mtifilm.com/ restoration/

    Mti film restoration software. https://mtifilm.com/ restoration/. 3, 6

  3. [3]

    Reference-based restoration of digitized analog videotapes

    Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Reference-based restoration of digitized analog videotapes. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1659– 1668, 2024. 2

    2024

  4. [4]

    Multidiffusion: Fusing diffusion paths for controlled image generation

    Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 3, 6

    2023

  5. [5]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 2

    Pith/arXiv arXiv 2023

  6. [6]

    Basicvsr: The search for essential compo- nents in video super-resolution and beyond

    Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential compo- nents in video super-resolution and beyond. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947–4956, 2021. 2

    2021

  7. [7]

    Demofusion: Democratising high- resolution image generation with no $

    Ruoyi Du, Dongliang Chang, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Demofusion: Democratising high- resolution image generation with no $. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6159–6168, 2024. 3, 6

    2024

  8. [8]

    Mamba: Linear-time sequence mod- eling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces. InFirst conference on lan- guage modeling, 2024. 2

    2024

  9. [9]

    Scalecrafter: Tuning-free higher- resolution visual generation with diffusion models

    Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher- resolution visual generation with diffusion models. InThe Twelfth International Conference on Learning Representa- tions, 2023. 3

    2023

  10. [11]

    Cogvlm2: Visual language models for image and video understanding.arXiv preprint arXiv:2408.16500,

    Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qing- song Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding.arXiv preprint arXiv:2408.16500,

    Pith/arXiv arXiv

  11. [12]

    Fouriscale: A frequency perspective on training-free high-resolution image synthesis

    Linjiang Huang, Rongyao Fang, Aiping Zhang, Guanglu Song, Si Liu, Yu Liu, and Hongsheng Li. Fouriscale: A frequency perspective on training-free high-resolution image synthesis. InEuropean conference on computer vision, pages 196–212. Springer, 2024. 3

    2024

  12. [13]

    Deepremaster: tem- poral source-reference attention networks for comprehensive video enhancement.ACM Transactions on Graphics (TOG), 38(6):1–13, 2019

    Satoshi Iizuka and Edgar Simo-Serra. Deepremaster: tem- poral source-reference attention networks for comprehensive video enhancement.ACM Transactions on Graphics (TOG), 38(6):1–13, 2019. 2

    2019

  13. [14]

    Latent space super-resolution for higher-resolution image generation with diffusion models

    Jinho Jeong, Sangmin Han, Jinwoo Kim, and Seon Joo Kim. Latent space super-resolution for higher-resolution image generation with diffusion models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2355–2365, 2025. 3

    2025

  14. [15]

    Deepenhancer: Temporally consistent focal transformer for comprehensive video enhancement

    Qin Jiang, Qinglin Wang, Lihua Chi, Wentao Ma, Feng Li, and Jie Liu. Deepenhancer: Temporally consistent focal transformer for comprehensive video enhancement. InPro- ceedings of the 2024 International Conference on Multimedia Retrieval, pages 969–977, 2024. 2

    2024

  15. [16]

    When fast fourier transform meets transformer for image restoration

    Xingyu Jiang, Xiuhui Zhang, Ning Gao, and Yue Deng. When fast fourier transform meets transformer for image restoration. InEuropean Conference on Computer Vision, pages 381–402. Springer, 2024. 5

    2024

  16. [17]

    Efficient optimization of inpainting scheme and line scratch detection for old film restoration

    Seong-Whan Kim and Ki-Hong Ko. Efficient optimization of inpainting scheme and line scratch detection for old film restoration. InPacific Rim International Conference on Artifi- cial Intelligence, pages 623–631. Springer, 2006. 2

    2006

  17. [18]

    On missing data treatment for degraded video and film archives: a survey and a new bayesian ap- proach.IEEE Transactions on Image Processing, 13(3):397– 415, 2004

    Anil C Kokaram. On missing data treatment for degraded video and film archives: a survey and a new bayesian ap- proach.IEEE Transactions on Image Processing, 13(3):397– 415, 2004. 2

    2004

  18. [19]

    Recurrent video restoration trans- former with guided deformable attention.Advances in Neural Information Processing Systems, 35:378–393, 2022

    Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration trans- former with guided deformable attention.Advances in Neural Information Processing Systems, 35:378–393, 2022. 4

    2022

  19. [20]

    Restoring degraded old films with recursive recurrent transformer networks

    Shan Lin and Edgar Simo-Serra. Restoring degraded old films with recursive recurrent transformer networks. InProceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6718–6728, 2024. 2

    2024

  20. [21]

    Accdiffusion: An accurate method for higher-resolution im- age generation

    Zhihang Lin, Mingbao Lin, Meng Zhao, and Rongrong Ji. Accdiffusion: An accurate method for higher-resolution im- age generation. InEuropean Conference on Computer Vision, pages 38–53. Springer, 2025. 3

    2025

  21. [22]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 2

    2021

  22. [23]

    Subject- diffusion: Open domain personalized text-to-image genera- tion without test-time fine-tuning

    Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject- diffusion: Open domain personalized text-to-image genera- tion without test-time fine-tuning. InACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 5

    2024

  23. [24]

    Making old film great again: Degradation-aware state space model for old film restoration

    Yudong Mao, Hao Luo, Zhiwei Zhong, Peilin Chen, Zhi- jiang Zhang, and Shiqi Wang. Making old film great again: Degradation-aware state space model for old film restoration. InProceedings of the Computer Vision and Pattern Recogni- tion Conference, pages 28039–28049, 2025. 2 8

    2025

  24. [25]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 5

    2021

  25. [26]

    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, pages 10684–10695, 2022. 2

    2022

  26. [27]

    Image processing for restoration of old film se- quences

    Takahiro Saito, Takashi Komatsu, Tomohisa Hoshi, and Toshi- aki Ohuchi. Image processing for restoration of old film se- quences. InProceedings 10th International Conference on Image Analysis and Processing, pages 709–714. IEEE, 1999. 2

    1999

  27. [28]

    Hierarchical patch diffusion mod- els for high-resolution video generation

    Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, and Sergey Tulyakov. Hierarchical patch diffusion mod- els for high-resolution video generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7569–7579, 2024. 3

    2024

  28. [29]

    Is one gpu enough? pushing image generation at higher-resolutions with founda- tion models.arXiv preprint arXiv:2406.07251, 2(3):5, 2024

    Athanasios Tragakis, Marco Aversa, Chaitanya Kaul, Roder- ick Murray-Smith, and Daniele Faccio. Is one gpu enough? pushing image generation at higher-resolutions with founda- tion models.arXiv preprint arXiv:2406.07251, 2(3):5, 2024. 3

    Pith/arXiv arXiv 2024

  29. [30]

    Bring- ing old films back to life

    Ziyu Wan, Bo Zhang, Dongdong Chen, and Jing Liao. Bring- ing old films back to life. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17694–17703, 2022. 2, 7

    2022

  30. [31]

    Svfr: A unified framework for gener- alized video face restoration

    Zhiyao Wang, Xu Chen, Chengming Xu, Junwei Zhu, Xi- aobin Hu, Jiangning Zhang, Chengjie Wang, Yuqi Liu, Yiyi Zhou, and Rongrong Ji. Svfr: A unified framework for gener- alized video face restoration. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 7406–7415,

  31. [32]

    Fastcomposer: Tuning-free multi- subject image generation with localized attention.Interna- tional Journal of Computer Vision, 133(3):1175–1194, 2025

    Guangxuan Xiao, Tianwei Yin, William T Freeman, Fr ´edo Durand, and Song Han. Fastcomposer: Tuning-free multi- subject image generation with localized attention.Interna- tional Journal of Computer Vision, 133(3):1175–1194, 2025. 5

    2025

  32. [33]

    Focal attention for long-range interactions in vision transformers.Advances in Neural Information Processing Systems, 34:30008–30022,

    Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers.Advances in Neural Information Processing Systems, 34:30008–30022,

  33. [34]

    Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization

    Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. InEuropean conference on computer vision, pages 74–91. Springer, 2024. 6

    2024

  34. [35]

    Motion- guided latent diffusion for temporally consistent real-world video super-resolution

    Xi Yang, Chenhang He, Jianqi Ma, and Lei Zhang. Motion- guided latent diffusion for temporally consistent real-world video super-resolution. InEuropean conference on computer vision, pages 224–242. Springer, 2024. 2

    2024

  35. [36]

    Inf- dit: Upsampling any-resolution image with memory-efficient diffusion transformer

    Zhuoyi Yang, Heyang Jiang, Wenyi Hong, Jiayan Teng, Wendi Zheng, Yuxiao Dong, Ming Ding, and Jie Tang. Inf- dit: Upsampling any-resolution image with memory-efficient diffusion transformer. InEuropean Conference on Computer Vision, pages 141–156. Springer, 2025. 3, 6

    2025

  36. [37]

    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, pages 3836–3847, 2023. 4

    2023

  37. [38]

    Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution

    Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535–2545, 2024. 2 9

    2024