DAM-VSR: Disentanglement of Appearance and Motion for Video Super-Resolution

Feng Gao; Guanying Chen; Kaihao Zhang; Le Li; Shaoshu Yang; Tao Wang; Wenhan Luo; Xiaoming Wei; Yong Zhang; Zhe Kong

REVIEW 4 major objections 5 minor 6 references

DAM-VSR: Disentanglement of Appearance and Motion for Video Super-Resolution

T0 review · 4 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 DAM-VSR splits video super-resolution into appearance from an enhanced reference frame and motion from a video ControlNet, reporting state-of-the-art real-world and AIGC results.

desk verdict Modular ISR+video-diffusion VSR with real promise; the central detail-propagation claim needs direct per-frame evidence before the SOTA talk. read the letter →

arxiv 2507.01012 v1 pith:622AAK4V submitted 2025-07-01 cs.CV

Zhe Kong , Le Li , Yong Zhang , Feng Gao , Shaoshu Yang , Tao Wang , Kaihao Zhang , Zhuoliang Kang

show 3 more authors

Xiaoming Wei Guanying Chen Wenhan Luo

This is my paper

classification cs.CV

keywords videosuper-resolutionappearanceandmotiondisentanglementStableDiffusionControlNetimagepriormotion-alignedbidirectionalsamplingtemporalconsistencyAIGCrestoration

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

This paper argues that real-world video super-resolution does not need a dedicated end-to-end video restoration model trained on enormous datasets. Instead, DAM-VSR splits the task into two parts: an image super-resolution model enhances a single reference frame, and the pre-trained video diffusion model Stable Video Diffusion (SVD) animates that enhanced frame while a video ControlNet forces the motion to match the low-quality input. The claim is that the realistic details introduced in the reference frame propagate through SVD's temporal layers to every other frame, yielding temporally consistent high-resolution video. The authors report state-of-the-art perceptual quality on real-world and AIGC video benchmarks, and a bidirectional sampling strategy that extends the method to long clips. If true, this makes VSR modular and cheap, since improvements in image super-resolution transfer immediately to video.

What carries the argument

The load-bearing object is the appearance–motion disentanglement itself: a high-quality reference image $\mathbf{B}_1 = \mathrm{ISR}(\mathbf{I}_1)$ supplies all appearance, while a video ControlNet conditioned on the low-quality clip $\mathbf{I}$ supplies all motion. The mechanism works as follows: the ISR output is encoded, repeated $k$ times, concatenated with noise in the channel dimension, and injected through cross-attention; the video ControlNet duplicates the UNet encoder, adds video embedding layers for the low-quality video, and feeds control features into the decoder skip connections. The second load-bearing component is motion-aligned bidirectional sampling: at each timestep, forward and backward denoising share the same latent noise, the backward pass uses the reversed clip and the high-quality end frame, and temporal self-attention maps are rotated 180 degrees, $A'_{k-i,k-j}=A_{i,j}$, so the backward pass follows the same motion path; predicted noises are averaged. Tile sampling and a fine-tuned VAE decoder handle resolution and fidelity.

What would settle it

Take a low-quality video with a fast camera pan or a large moving object, run DAM-VSR, and compare how much detail from the enhanced first frame appears in later frames: if a later frame loses the textures present in the first frame while a static or slow-motion clip retains them, then the claimed temporal propagation of appearance fails. A cleaner synthetic version: degrade a sharp video, enhance only the first frame with ISR, generate the rest with SVD plus ControlNet, and check whether non-reference frames recover the ground-truth texture rather than merely following the ControlNet-conditioned low-quality input.

Watch

Extended reading notes

Core claim

The central discovery is that appearance and motion can be separated in diffusion-based VSR: appearance comes from an ISR-enhanced reference image, motion from the low-quality video itself. The reference frame is super-resolved, repeated and concatenated with noise in latent space, and injected into the SVD UNet and the video ControlNet via cross-attention; the low-quality video is encoded by a video ControlNet that follows the image-ControlNet design. Because SVD is an image-to-video model, the enhanced appearance of the reference frame is carried to later frames by temporal attention, so the framework inherits SVD's temporal consistency. A motion-aligned bidirectional sampling stage generates long videos by running forward and backward generation with shared latent noise and rotated temporal attention maps, then averaging the predicted noise; adjacent clips share high-quality endpoint frames, avoiding flicker and error accumulation. The paper reports that this outperforms existing real-world and AIGC VSR methods on no-reference perceptual metrics.

Load-bearing premise

The method depends on the assumption that realistic details produced by the image super-resolution model on the reference frame are propagated by the video diffusion model's temporal layers to the remaining frames; if that propagation fails under strong motion or heavy degradation, the appearance path collapses and the method becomes a motion-only passthrough.

Editorial extensions

If this is right

Real-world video super-resolution can be built from off-the-shelf image super-resolution models and a pre-trained image-to-video diffusion model, avoiding the massive training data and GPU budgets of end-to-end video diffusion restorers.
Improvements in image super-resolution translate directly into video improvements without retraining the video pipeline, since the ISR model is a plug-in module.
Choosing a fidelity-oriented ISR method such as ResShift versus a generation-oriented one such as SupIR or InvSR gives a predictable fidelity-perceptual-quality trade-off for the output video.
Long videos can be processed by clip concatenation with shared high-quality endpoint frames, reducing flicker compared to independent clip processing and avoiding the error accumulation of auto-regressive generation.
The same disentanglement, with a depth-conditioned ControlNet, extends to video editing and video style transfer by modifying only the reference image.

Reading between the lines

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

Going beyond the paper, the framework predicts that any future improvement in blind image super-resolution will immediately lift video super-resolution quality on real-world footage, which is a cheap way to keep pace with the fast-moving ISR literature.
The bidirectional sampling idea suggests a testable scaling law: flicker between clips should drop as the overlap length and the number of shared high-quality endpoint frames increase; the paper does not quantify this relationship.
Because the method is a wrapper around SVD, it inherits SVD's failure modes; one testable consequence is that non-reference frame quality should degrade gracefully with increasing motion magnitude until it suddenly collapses when temporal attention can no longer track correspondences.
The same appearance-motion split could be applied to other conditional video generation tasks, such as keyframe interpolation or frame prediction, by choosing which reference frames to super-resolve; the paper only hints at editing and stylization.

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.

Desk Editor's Note

read the letter

Bottom line: this is a genuinely modular VSR recipe, and the paper's core intuition is sound: ISR is ahead of VSR, so feeding an ISR-enhanced reference frame into SVD plus a video ControlNet buys you detail and temporal consistency at a fraction of SeedVR's training cost. The combination is new, and the Fig. 2 analysis backs it up directly. The bidirectional sampling with rotated temporal attention maps is borrowed from Generative Inbetweening, but the adaptation to long-video VSR is clean and the ablation shows it helps.

Credit where it's due: the ablations are properly isolated and independent of the final design. Each component earns its place. The numbers on VideoLQ and AIGC29 are strong, and the modularity claim—swap the ISR model for fidelity or perceptual quality without retraining—is a real practical advantage.

Soft spots, in order. First, the central mechanism, ISR details propagating to the remaining frames, is asserted rather than demonstrated. Whole-clip averages can hide a strong first frame and mediocre rest. The stress-test concern is legitimate: the paper should report per-frame fidelity and ideally detail transfer under increasing motion. Until that is shown, "SOTA on real-world VSR" is overreach. Second, Table 1's baseline definitions are confusing; the text says "Baseline" is video ControlNet only, but the row labels and the w/o-disentanglement variants need a clearer mapping. Minor, but it makes the ablation hard to parse. Third, the supplementary materials have a missing citation for the degradation pipeline, leaving the training distribution underspecified. Sloppy. Fourth, no SeedVR comparison, despite SeedVR being the obvious scaling baseline. Fifth, no error bars anywhere. None of these are fatal, but they are exactly what a referee should request.

On the stress-test worry itself: the paper does not establish that propagation holds under strong motion or occlusion, and its limitation section only mentions ISR failure, not motion magnitude. That is a real gap, not a manufactured one. It can be closed with per-frame evaluation.

Verdict: this deserves a serious referee. I would send it out and ask for the per-frame analysis, SeedVR comparison, clearer ablation definitions, and the missing citation. It's also a good reading-group paper; the framework is easy to explain and the failure-mode discussion is productive.

Referee Report

4 major / 5 minor

Summary. The paper proposes DAM-VSR, a video super-resolution framework built on Stable Video Diffusion (SVD) together with a video ControlNet. The method disentangles the task into appearance enhancement and motion control: an image super-resolution (ISR) model enhances the first (reference) frame of each clip, and a video ControlNet conditions the SVD generation on the low-quality input video to control motion. The authors further introduce a motion-aligned bidirectional sampling strategy for long videos, alternating forward and backward generation with shared rotated temporal attention maps and blending the predicted noises. Experiments are reported on synthetic benchmarks (UDM10, YouHQ40, REDS30), on the real-world VideoLQ dataset, and on a self-collected AIGC29 dataset, with the claimed outcome that DAM-VSR achieves state-of-the-art performance on real-world and AIGC data.

Significance. If the central claim is substantiated, the paper would demonstrate a practical and lightweight alternative to end-to-end video restoration models: a frozen image-to-video diffusion model combined with a strong ISR model and a video ControlNet can generate realistic, temporally consistent details without training at the scale of SeedVR. The modular disentanglement idea is attractive, and the ablation study in Table 1 broadly supports the contribution of the three proposed components. The bidirectional sampling strategy is a clever mechanism for long-video consistency. However, the current evidence does not yet establish the central mechanism: the propagation of ISR-enhanced reference details to all frames is asserted rather than quantified, and the headline SOTA claims on real-world and AIGC data rest on non-reference metrics over a small, unshared dataset without statistical or human evaluation. These gaps are load-bearing for the main claim, so the paper needs additional analysis and experiments before it can be accepted.

major comments (4)

[Section 3.2, Eq. (2), Table 2, Section 5] The central mechanism of the framework is asserted but not directly demonstrated. After Eq. (2), the paper states that 'these generated realistic details in the reference frame can be propagated to the remaining frames leveraging the powerful temporal generation capabilities of SVD,' but all reported metrics in Table 2 are averaged over whole clips. An enhanced first frame can therefore raise the clip-level averages even if frames 2 through 14 receive no benefit from the ISR reference. The limitation section (Section 5) only addresses the case 'when ISR does not perform well,' which is a different failure mode from the one in which a perfect ISR frame fails to propagate under large motion, occlusion, or scene changes. Since the SOTA claim for real-world and AIGC data depends on this propagation, the paper should quantify it with per-frame fidelity curves (e.g., PSNR/LPIPS as a function of frame index on UDM10 and YouHQ40), a comparison of first-frame versus later-frame quality, and a breakdown by motion magnitude using optical-flow-based grouping.
[Table 1 and Section 4.2.1] The ablation configurations are not sufficiently specified. The text defines 'Baseline' as training only a video ControlNet conditioned on low-quality videos, which is the same description given for 'w/o Disentanglement During Training and Inference'; yet the two rows report different PSNR values (24.775 versus 25.379). The reader cannot determine which of the three components (disentanglement, bidirectional sampling, VAE-decoder fine-tuning) are active in each row. In addition, 'w/o Disentanglement only During Training' is described as using the same ControlNet while 'conducting reference image enhancement only during inference,' but a network trained without a reference-image condition would not know how to consume the enhanced reference frame, and the injection mechanism is not explained. Please define each row as a toggle vector over the three components and specify the exact training and inference inputs for every variant.
[Table 2 and Section 4.1.1] The headline claim of state-of-the-art performance on real-world and AIGC data is supported only by non-reference metrics on VideoLQ and on a self-collected set of 29 videos (AIGC29), with no error bars, no statistical significance tests, and no human perceptual study. Some of these metrics, especially BRISQUE, are known to respond to texture smoothing or generated artifacts and may not reflect perceived quality, so the large gaps over baselines are not yet convincing. The authors should report standard deviations over multiple runs or per-video score distributions, and should add a forced-choice or MOS user study on the real-world and AIGC clips. Releasing AIGC29 publicly would also allow independent verification of the AIGC claim.
[Supplementary Section 2 (training strategy)] The training degradation pipeline is not reproducible as written. The sentence 'Throughout the training process, we adhere to the degradation pipeline of to generate low-quality and high-quality video pairs' contains a missing citation, so the degradation model used to create training pairs is unspecified. This matters because the method targets real-world blind VSR and the degradation distribution is what the ControlNet learns to invert. In addition, the training data description ('a large-scale high-resolution dataset from the website') is too vague to reproduce. Please specify the degradation model with a reference, describe the data source and license, and provide details of the train/test split.

minor comments (5)

[Section 3.2] The description of the reference-image latent ('after K times repeating and then the image latent concatenates with a random noise in the channel dimension') would benefit from an explicit tensor-shape statement, because it is currently unclear whether the reference latent is concatenated with the noisy latent along the channel axis or used as an additional condition.
[Section 3.3.1, Eqs. (3)-(4)] The networks F and G are used in Eqs. (3) and (4) before being introduced formally; the text refers to them as the forward and backward generation processes but does not define the symbols. Please introduce F and G explicitly before their first use.
[Figures 3 and 4] The pipeline diagrams are dense, and labels such as 'Predicted reverse noise' and 'Frame Reverse' are not explained in the captions; a short caption-level walkthrough of the bidirectional sampling would make the figures self-contained.
[Section 3.4] The video editing and style transfer applications are demonstrated only with qualitative examples, and a depth-conditioned ControlNet was trained for these tasks. Adding a sentence that these results are illustrative rather than quantitatively evaluated would set appropriate expectations.
[Table 2] The table caption does not state which ISR method was used for each dataset; the text in Section 4.3.1 says ResShift is used for synthetic data and SupIR/InvSR for real-world and AIGC data, but this information should appear in the table caption or in a footnote.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the central claim is empirically tested against external baselines, with only minor non-load-bearing self-citations.

full rationale

The paper's derivation does not reduce to its inputs. Eq. (2) defines the enhanced reference frame A_hat_1 = ISR(I_1), and the rest of the pipeline feeds that frame and the low-quality clip to SVD plus a video ControlNet. The contribution of the reference enhancement is tested by ablations (Table 1 rows c/d/g) rather than assumed: a ControlNet trained without the reference condition performs worse, and injecting ISR at inference into a model not trained for it does not match the full method. Evaluations in Table 2 are against external baselines on standard and collected benchmarks, and the real-world/AIGC metrics are not produced by fitting a parameter to the test set. The only self-citations (e.g., [Kong et al. 2025], [Zhao et al. 2024]) are contextual related-work citations and do not carry the argument. The propagation claim in Sec. 3.2 ('these generated realistic details in the reference frame can be propagated to the remaining frames') is an empirical assumption rather than a circular step; the paper itself lists the dependence on ISR quality as a limitation (Sec. 5). A separate reproducibility gap is the appendix's missing citation for the training degradation pipeline ('we adhere to the degradation pipeline of to generate low-quality and high-quality video pairs'), which should be fixed but is not circularity. Overall no prediction is equivalent to a fitted input, so the paper is not circular; at most there are non-load-bearing self-citations and an unquantified propagation assumption.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

All assumptions are domain assumptions about pre-trained models and data distributions; there are no new postulated physical entities. The two most fragile assumptions are the temporal-attention-map reversal and the propagation of ISR details through SVD's temporal layers.

free parameters (6)

learning rate = 8e-5
Constant learning rate for AdamW, chosen by hand for training (Sec 4.1).
perceptual loss weight alpha = 1
In VAE-decoder fine-tuning loss L = L2 + alpha * Lperceptual + beta * LGAN (Appendix Sec 2).
GAN loss weight beta = 0.025
Same loss equation as above.
SDEdit initial noise ratio = 60% of steps
Noise added to low-quality input to reduce sampling steps to 18 (Appendix Sec 3).
clip length K = 14
SVD's native frame length, used for all clips (Sec 3.2).
training iterations per stage = 60K / 45K / 10K
Three-stage training schedule (Appendix Sec 2).

assumptions (5)

domain assumption SVD's image-to-video generative prior transfers to video super-resolution.
The method relies on SVD animating an enhanced reference frame into a high-quality clip; assumed in Sec 3.2 and tested only via ablations on the final pipeline, not in isolation.
domain assumption Temporal self-attention maps of SVD encode motion dynamics, and reversing their row/column order yields the motion of the reversed video.
This is the core of the bidirectional sampling (Sec 3.3.1); it is borrowed from Wang et al. 2024c and assumed to hold when a video ControlNet is added.
domain assumption A video ControlNet built by cloning the UNet encoder, adding video embedding layers, and adding its outputs to skip connections provides effective motion control.
Extended from image ControlNet design (Zhang et al. 2023), asserted in Sec 3.2 without a dedicated ablation of ControlNet strength.
domain assumption The collected 300K-clip training set and the (unidentified) degradation pipeline yield realistic low-quality/high-quality pairs for training.
Appendix Sec 2 says 'we adhere to the degradation pipeline of' with the reference left blank; dataset composition is not described.
standard math Diffusion v-prediction objective is a valid training target for the denoising U-Net.
Sec 3.1 restates the standard latent diffusion v-prediction loss from SVD.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DAM-VSR: Disentanglement of Appearance and Motion for Video Super-Resolution." pith.science (2026). https://pith.science/paper/622AAK4V

@misc{pith2026250701012,
  author       = {Pith},
  title        = {Pith review of: DAM-VSR: Disentanglement of Appearance and Motion for Video Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/622AAK4V}},
  note         = {Machine review of arXiv:2507.01012}
}

read the original abstract

Real-world video super-resolution (VSR) presents significant challenges due to complex and unpredictable degradations. Although some recent methods utilize image diffusion models for VSR and have shown improved detail generation capabilities, they still struggle to produce temporally consistent frames. We attempt to use Stable Video Diffusion (SVD) combined with ControlNet to address this issue. However, due to the intrinsic image-animation characteristics of SVD, it is challenging to generate fine details using only low-quality videos. To tackle this problem, we propose DAM-VSR, an appearance and motion disentanglement framework for VSR. This framework disentangles VSR into appearance enhancement and motion control problems. Specifically, appearance enhancement is achieved through reference image super-resolution, while motion control is achieved through video ControlNet. This disentanglement fully leverages the generative prior of video diffusion models and the detail generation capabilities of image super-resolution models. Furthermore, equipped with the proposed motion-aligned bidirectional sampling strategy, DAM-VSR can conduct VSR on longer input videos. DAM-VSR achieves state-of-the-art performance on real-world data and AIGC data, demonstrating its powerful detail generation capabilities.

Figures

Figures reproduced from arXiv: 2507.01012 by the authors.

**Figure 1.** Visualization comparisons with state-of-the-art methods on both real-world and AIGC videos. Our DAM-VSR demonstrates remarkable upscaling [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Analysis of the role of ISR enhancement. a) Input video. b) The [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The overall pipeline of the proposed DAM-VSR method. Our method introduces an appearance and motion disentanglement framework for VSR. To [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: The model architecture of the disentangled forward generation (the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Illustration of long video processing. Through clip concatenation, [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: alitative ablation study of DAM-VSR. 4.2 Ablation Study To validate the e ectiveness of various components within DAMVSR, we conduct an ablation study encompassing the following elements: Disentanglement Framework, Fine-tuning VAE Decoder and Motion-aligned Bidirectio…

**Figure 7.** Figure 7: The experimental results of the proposed framework for other applications, including video editing and video style transfer. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

**Figure 8.** Figure 8: alitative comparison on synthetic low-quality videos from YouHQ40. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: alitative comparison on real-world videos from VideoLQ. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

**Figure 10.** Figure 10: alitative comparison on AIGC videos [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

**Figure 4.** Figure 4: The architecture consists of a denoising UNet and a video [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 11.** Figure 11: Illustration of tile sampling, which can handle the generation of [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

6 extracted references · 4 canonical work pages

[4]

StyleMaster: Stylize Your Video with Artistic Generation and Translation

StyleMaster: Stylize Your Video with Artistic Generation and Translation. arXiv preprint arXiv:2412.07744 (2024). Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. 2024. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In CVPR. 25669–25680. Zon...

work page Pith review arXiv 2024
[5]

Upscale-A-Video: Temporal-Consistent Diﬀusion Model for Real-World Video Super-Resolution. In CVPR. 2535–2545. 10 • Zhe Kong, Le Li, Yong Zhang, Feng Gao, Shaoshu Yang, Tao Wang, Kaihao Zhang, Zhuoliang Kang, Xiaoming Wei, Guanying Chen, and Wenhan Luo Source Video Edited Video Source Video Style Transferred Video Fig. 7. The experimental results of the p...

work page
[30]

Using SDEdit, we add noise equivalent to 60% of the steps to the low-quality input as initial noise, eﬀectively reducing the number of sampling steps to 18, minimizing inference time

work page
[2022]

High-resolution image synthesis with latent diﬀusion models. In CVPR. 10684– 10695. Claudio Rota, Marco Buzzelli, and Joost van de Weijer. 2025. Enhancing Perceptual Quality in Video Super-Resolution through Temporally-Consistent Detail Synthesis using Diﬀusion Models. ECCV (2025). Jingfan Tan, Xiaoxu Chen, Tao Wang, Kaihao Zhang, Wenhan Luo, and Xiaocun Cao

work page 2025
[2023]

TCSVT (2023)

Blind face restoration for under-display camera via dictionary guided trans- former. TCSVT (2023). Jingfan Tan, Hyunhee Park, Ying Zhang, Tao Wang, Kaihao Zhang, Xiangyu Kong, Pengwen Dai, Zikun Liu, and Wenhan Luo. 2024. Blind face video restoration with temporal consistent generative prior and degradation-aware prompt. In ACMM. 1417–1426. Qi Tang, Yao Z...

arXiv 2023
[2024]

Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR. 8640–8650. Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. 2024. Vrt: A video restoration transformer. TIP (2024). Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zh...

arXiv 2024

Pith tools

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

[1] [4]

StyleMaster: Stylize Your Video with Artistic Generation and Translation

StyleMaster: Stylize Your Video with Artistic Generation and Translation. arXiv preprint arXiv:2412.07744 (2024). Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. 2024. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In CVPR. 25669–25680. Zon...

work page Pith review arXiv 2024

[2] [5]

Upscale-A-Video: Temporal-Consistent Diﬀusion Model for Real-World Video Super-Resolution. In CVPR. 2535–2545. 10 • Zhe Kong, Le Li, Yong Zhang, Feng Gao, Shaoshu Yang, Tao Wang, Kaihao Zhang, Zhuoliang Kang, Xiaoming Wei, Guanying Chen, and Wenhan Luo Source Video Edited Video Source Video Style Transferred Video Fig. 7. The experimental results of the p...

work page

[3] [30]

Using SDEdit, we add noise equivalent to 60% of the steps to the low-quality input as initial noise, eﬀectively reducing the number of sampling steps to 18, minimizing inference time

work page

[4] [2022]

High-resolution image synthesis with latent diﬀusion models. In CVPR. 10684– 10695. Claudio Rota, Marco Buzzelli, and Joost van de Weijer. 2025. Enhancing Perceptual Quality in Video Super-Resolution through Temporally-Consistent Detail Synthesis using Diﬀusion Models. ECCV (2025). Jingfan Tan, Xiaoxu Chen, Tao Wang, Kaihao Zhang, Wenhan Luo, and Xiaocun Cao

work page 2025

[5] [2023]

TCSVT (2023)

Blind face restoration for under-display camera via dictionary guided trans- former. TCSVT (2023). Jingfan Tan, Hyunhee Park, Ying Zhang, Tao Wang, Kaihao Zhang, Xiangyu Kong, Pengwen Dai, Zikun Liu, and Wenhan Luo. 2024. Blind face video restoration with temporal consistent generative prior and degradation-aware prompt. In ACMM. 1417–1426. Qi Tang, Yao Z...

arXiv 2023

[6] [2024]

Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR. 8640–8650. Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. 2024. Vrt: A video restoration transformer. TIP (2024). Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zh...

arXiv 2024