Pith. sign in

REVIEW 4 major objections 6 minor 68 references

LiftVSR: Lifting Image Diffusion to Video Super-Resolution via Hybrid Temporal Modeling with Only 4$\times$RTX 4090s

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LiftVSR claims the image diffusion prior of PixArt-α can be lifted to video super-resolution, beating video-diffusion methods on perceptual quality and temporal coherence using only 4×RTX 4090 GPUs.

desk verdict A promising efficiency-first VSR approach whose long-term consistency claim needs more evidence before it earns 'state-of-the-art'. read the letter →

arxiv 2506.08529 v1 pith:BQ5YR2LG submitted 2025-06-10 cs.CV

Xijun Wang , Xin Li , Bingchen Li , Zhibo Chen This is my paper · ORCID
classification cs.CV
keywords videosuper-resolutiondiffusionmodeltemporalmodelingdynamicattentionmemorycacheasymmetricsamplinglatenttransformerlong-videoconsistency
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

The paper claims that a pre-trained image diffusion model can be turned into a working video super-resolution (VSR) system without the multi-A100 training budgets of video diffusion models. Its proposed framework, LiftVSR, takes PixArt-α and adds two lightweight temporal components: Dynamic Temporal Attention, which models fine-grained motion within short 8-frame segments by predicting a flow path per attention head, and an Attention Memory Cache, which keeps a compressed memory of past segments so long videos stay coherent. An asymmetric sampling strategy trains different segments at different noise levels so that the cache from a nearly clean denoising step can condition the next, still-noisy segment without accumulating errors. On the tested benchmarks the model reports the best perceptual-quality scores among compared diffusion-based VSR methods and competitive fidelity, while training and inference fit on 4×RTX 4090 GPUs. If correct, this makes diffusion-grade video enhancement practical for long videos at consumer-GPU cost.

What carries the argument

The load-bearing object is the hybrid temporal modeling mechanism, which pairs two components inside a DiT-based image diffusion backbone (PixArt-α). Dynamic Temporal Attention (DTA) is an attention layer in which each head estimates a flow map from a reference frame and warps the key/value tokens of the whole short segment toward it, so that attention operates on aligned tokens; this keeps within-segment modeling cheap. Attention Memory Cache (AMC) is a cache unit of length $l=2$ that stores a compressed summary of past segments, read through a cross-attention query that shares weights with DTA's temporal attention and written through a gated update $g=\sigma(W_{gate}([F_{pool},F_{cache}]))$ that blends pooled new features with cached ones. The asymmetric sampling strategy is the third piece: during training different segments are diffused at distinct timesteps, so that at inference the cache harvested at the last denoising step of one segment can be injected as conditioning for the next segment's noisy latent without the model breaking on the feature mismatch. Together the three mechanisms carry the paper's claim that long-term consistency can be bought with a small, bounded memory cache instead of full-sequence attention.

What would settle it

Run the trained LiftVSR on a 300-frame video with a hard scene cut at a segment boundary, using the default 15-step DPM-Solver schedule, and inspect temporal profiles and DOVER scores around that boundary. If flicker, color drift, or remnants of the previous scene appear exactly at segment boundaries and grow with video length — or if the same video sampled at 5 and 50 steps shows inconsistent cache behavior — the cache exchangeability assumption is broken.

Watch

Extended reading notes

Core claim

LiftVSR claims to establish that an image-wise diffusion prior can be elevated to video super-resolution without the prohibitive cost of video diffusion models, by decomposing temporal modeling into two complementary mechanisms. Dynamic Temporal Attention (DTA) replaces conventional temporal attention: for each attention head it predicts a single token-flow path from a reference frame to other frames and warps the key and value tokens along it, so each attention operation touches far fewer tokens than deformable attention while still capturing correspondences. Attention Memory Cache (AMC) maintains a cache unit queried by a cross-attention module that shares parameters with DTA's temporal attention, and refreshed through a gated, average-pooled update, so historical segment information propagates forward with memory bounded by the cache size rather than video length. The asymmetric sampling strategy, adapted from diffusion-forcing ideas, trains different segments with distinct diffusion timesteps, and at inference uses the cache from the last denoising step as the condition for the next segment, which the authors argue avoids noise accumulation and cuts GPU memory. Across UDM10, REDS30, YouHQ40, SPMCS, and VideoLQ, the paper reports the highest MUSIQ, CLIPIQA, and DOVER scores among the compared diffusion VSR methods, with PSNR slightly below non-generative methods such as RealViformer.

Load-bearing premise

The whole scheme rests on the assumption that a memory cache taken from a nearly fully denoised video segment can be safely injected as conditioning into the next, still-noisy segment; if training at distinct timesteps per segment does not teach the model to tolerate that feature mismatch, the long-term consistency mechanism fails on longer videos or different sampling step counts.

Editorial extensions

If this is right

  • Memory cost for long videos stops growing with video length: only the current 8-frame segment plus the small cache need to be live during inference.
  • Long-term coherence is obtained without a video diffusion model, so labs without 8-16 A100 clusters can train and run diffusion-based VSR.
  • Because the cache comes from the last, nearly clean denoising step, both GPU memory and accumulated noise across segments are reduced relative to caching every step.
  • On the reported benchmarks, the approach beats both image-diffusion VSR methods (Upscale-A-Video, MGLD-VSR) and a video-diffusion method (STAR) on perceptual metrics while using far less compute.

Reading between the lines

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

  • The shared weights between AMC's query attention and DTA's temporal attention suggest the cache acts as a set of summary tokens read in the same attention space; a cheap test would be shrinking the cache length from 2 to 1 or growing it to 4 to see how long-range coherence trades against memory.
  • Because the asymmetric sampling ties segments to distinct timesteps, the model may be sensitive to the number of inference sampling steps; testing at 5 and 50 steps would reveal whether the cache-conditioning strategy generalizes beyond the trained 15-step schedule.
  • The cache naturally carries one scene's appearance into the next segment, so a hard scene cut is a stress test the paper does not report; a gated reset triggered by content change would be a natural extension.
  • If the community shifts consistency reporting from Ewarp, which rewards blur, to temporal profiles and DOVER as argued here, comparisons among diffusion VSR methods could change materially.
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

4 major / 6 minor

Summary. The paper proposes LiftVSR, a video super-resolution framework built on the image diffusion model PixArt-α. It introduces two temporal modeling components: Dynamic Temporal Attention (DTA) for short-range intra-segment alignment and Attention Memory Cache (AMC) for cross-segment long-range consistency, plus an asymmetric sampling strategy (ASS) that uses a cache from the last denoising step of a segment to condition the next segment. The authors report state-of-the-art no-reference perceptual metrics (MUSIQ, CLIPIQA, DOVER) on four synthetic benchmarks and one real-world benchmark with training and inference on only four RTX 4090 GPUs, alongside ablations showing monotonic DOVER gains as DTA, AMC, and ASS are added.

Significance. If the perceptual-quality claims hold, LiftVSR would be a useful efficiency-oriented alternative to video-diffusion-based VSR: it needs far less GPU than STAR or VEnhancer, and its DTA/AMC design is a plausible way to lift an image prior to long videos. The ablations are clearly structured, and the paper provides a substantial set of comparisons and a qualitative temporal-profile analysis. However, the 'state-of-the-art' claim is currently too broad: the method is consistently below non-generative baselines in PSNR, SSIM, and LPIPS and is also below STAR in Ewarp on all synthetic benchmarks, so the central claim rests entirely on no-reference perceptual metrics and needs either a higher bar of evidence or a carefully qualified claim.

major comments (4)
  1. [Section 4.2, Tables 1 and 5] The abstract and conclusion claim 'state-of-the-art results' without qualification, but the method is below RealViformer on PSNR, SSIM, and LPIPS on every synthetic benchmark (e.g., UDM10 PSNR 27.654 vs 29.557, SSIM 0.7885 vs 0.8515, LPIPS 0.2923 vs 0.2338). The SOTA claim is thus entirely delegated to no-reference metrics (MUSIQ, CLIPIQA, DOVER). At minimum, the paper should explicitly frame the contribution as perceptual-quality SOTA and report the objective metric gap in the abstract. The current wording is misleading.
  2. [Section 3.4] The asymmetric sampling strategy is under-specified. The paper states that training 'diffuses different segments with distinct timesteps' but does not specify the distribution over per-segment timesteps, the number of segments, or whether the cache is exposed to all denoising steps. The ablation in Table 2 validates ASS only at the default 15-step DPM-Solver setting on YouHQ40; no longer videos, more segments, or other step counts are tested. Therefore the claim that the strategy 'avoids noise accumulation' is not established beyond the exact evaluated configuration. This is a load-bearing point because it is the mechanism claimed to preserve long-term consistency.
  3. [Section 3.3] The AMC query operation is not sufficiently described for reproducibility. The paper says the cache query uses 'a cross-attention module, which shares parameters with temporal attention in the DTA module,' but DTA is a flow-warping attention (Eqs. 2-3) over a reference-frame query and multi-head keys/values. It is unclear how a cache tensor of length l=2 is integrated into this attention, what the key/value projections are, and whether positional encoding is applied to the cache. Please provide the exact forward pass of the cache query and state the tensor shapes.
  4. [Section 4.2, Table 1] The temporal-consistency evidence is one-sided. Ours has larger Ewarp than both STAR and Upscale-A-Video on all four synthetic datasets (e.g., REDS30: 0.613 vs 0.179 for STAR; YouHQ40: 0.241 vs 0.093 for STAR). The paper attributes this to detail generation, but then it should provide a quantitative temporal coherence metric (e.g., a learned video quality metric that captures temporal artifacts, or a user study) to support the 'state-of-the-art temporal coherence' claim. A single subjective temporal profile (Fig. 6) is not strong enough to carry the claim.
minor comments (6)
  1. [Section 4.2] The reference to 'Table 5' for the main quantitative results is incorrect; Table 1 is the main table. The same error appears in the 'Quantitative Comparison' paragraph.
  2. [Appendix A.2] The phrase 'reduce to warping error' should read 'reduces the warping error'.
  3. [Table 4] The row labels (a), (b), (v) are confusing; the table does not clearly indicate which configuration corresponds to Video VAE and Color Correction. Please use explicit checkmarks or rename the rows.
  4. [Section 3.1] The sentence 'decoded back to the image space eith a V AE decoder' contains a typo; it should be 'with a VAE decoder'.
  5. [Abstract and Figure 1] The abstract uses '4×RTX 4090s' while Figure 1 uses '4xRTX 4090s'; please standardize the notation.
  6. [Table 2] Adding AMC (row c) causes MUSIQ to drop from 67.328 to 66.282, which is a non-monotonic change. This is not discussed; please comment on whether this is within expected noise or indicates a trade-off.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LiftVSR's temporal modules are newly proposed and externally evaluated, with no fitted parameter renamed as a prediction.

full rationale

LiftVSR's derivation chain is self-contained against external components. The diffusion backbone is the pre-trained, externally published PixArt-α model [8]; the temporal contributions, Dynamic Temporal Attention and Attention Memory Cache, are new modules whose computations are specified in Eqs. 2–5 and whose hyperparameters are listed in Table 3. The asymmetric sampling strategy is credited to the external Diffusion Forcing work [6] as an inspiration, not to the authors' own prior results, and it is then tested through the ablation in Table 2 rather than assumed. No parameter is fitted to a benchmark and subsequently reported as a prediction: all quantitative results are measured on held-out external datasets (REDS30, SPMCS, UDM10, YouHQ40, VideoLQ) against published baselines, and the perceptual and temporal metrics used are standard external evaluation tools. No equation reduces to its own inputs: Eq. 1 is the standard diffusion denoising objective, Eqs. 2–3 are flow-warped attention, and Eqs. 4–5 are a gated cache update, none of which presuppose the claimed DOVER scores, MUSIQ gains, or temporal-coherence conclusions. Self-citation is not load-bearing anywhere in the argument; the paper's own contributions are validated by ablations rather than by appeal to prior work. Appendix C acknowledges genuine limitations such as difficulty with small objects, faces, and text, the overhead of overlapping sampling for arbitrary resolutions, and the lack of real-time inference, but these limitations do not indicate that any claimed result is equivalent to an input by definition. The paper therefore does not exhibit significant circularity.

Assumptions & free parameters 7 free parameters · 8 assumptions · 3 invented entities

The central claim rests on pretrained external priors (PixArt-alpha, SVD Video VAE), on hand-chosen hyperparameters that define the efficiency-quality trade-off, and on newly introduced modules DTA and AMC whose behavior is only validated internally. The asymmetric sampling schedule is the least specified load-bearing component. No new physical entities are proposed.

free parameters (7)
  • DTA segment length = 8 frames
    Chosen in Section 4.1. It controls the short-term temporal window and the compute of the attention operation, and it directly shapes the hybrid efficiency claim.
  • AMC cache length = 2
    Set in Appendix Table 3. It controls how much history is stored and therefore the long-term consistency versus memory trade-off.
  • Number of sampling steps = 15
    Set in Appendix A.2. It balances perceptual quality and speed and directly influences the reported metrics and runtime.
  • Temporal overlap length = 1 frame
    Set in Appendix A.2. Used to stitch segments and enhance temporal coherence across segment boundaries.
  • DTA insertion block interval = 3
    Set in Appendix Table 3. Determines which DiT blocks receive temporal attention and affects the modeling capacity of DTA.
  • Training batch sizes = image: 64, video: 8
    Chosen in Section 4.1. These settings affect training stability and are part of the low-compute training claim.
  • Learning rate = 1e-5
    Chosen in Section 4.1. A standard but hand-selected value that affects convergence and final quality.
assumptions (8)
  • domain assumption Pretrained PixArt-alpha and SVD Video VAE priors transfer to video super-resolution
    The method's quality depends on these external models; the paper provides no independent evidence of transferability beyond its own outputs.
  • domain assumption DTA flow estimation network can learn reliable dense correspondences in latent space from concatenated query and key tokens
    Load-bearing for fine-grained temporal alignment; no theoretical or independent experimental support is given.
  • domain assumption No-reference metrics MUSIQ, CLIPIQA, and DOVER are valid proxies for subjective quality
    The SOTA claim rests on these metrics, while PSNR, SSIM, and LPIPS are worse for LiftVSR than for several baselines.
  • domain assumption E_warp is not a suitable temporal consistency measure for textured outputs
    Used to explain away higher E_warp values, but no quantitative alternative consistency metric is provided.
  • domain assumption Diffusion Forcing style asymmetric sampling transfers to segment-wise cache interaction in VSR
    Borrowed from reference [6], but the exact timestep schedule and training procedure are underspecified in Section 3.4.
  • ad hoc to paper AMC query cross-attention can share parameters with DTA temporal attention
    Architectural choice stated in Section 3.3 without an ablation or theoretical justification.
  • domain assumption Synthetic degradation pipeline and training data cover real-world degradations
    Generalization to VideoLQ is tested, but the coverage of all real-world distortions is not guaranteed.
  • standard math Standard LDM forward and reverse process is a valid training objective
    Equation (1) relies on standard diffusion theory rather than a proof in this paper.
invented entities (3)
  • Dynamic Temporal Attention (DTA) module
    purpose: Adds a learned flow-estimation and warping path to temporal attention so the image DiT can align features within an 8-frame segment.
    Its benefit is shown only in the paper's ablation Table 2; there is no external falsifiable handle outside this paper.
  • Attention Memory Cache (AMC) unit
    purpose: Stores gated historical segment features and conditions the next segment to maintain long-term consistency.
    Supported only by internal ablations and sample videos; no artifact outside the paper validates it.
  • Token flow map F_flow
    purpose: A dense per-head flow field predicted to warp key and value tokens across frames in DTA.
    An intermediate learned quantity whose correctness is not independently testable outside the trained model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LiftVSR: Lifting Image Diffusion to Video Super-Resolution via Hybrid Temporal Modeling with Only 4$\times$RTX 4090s." pith.science (2026). https://pith.science/paper/BQ5YR2LG

@misc{pith2026250608529,
  author       = {Pith},
  title        = {Pith review of: LiftVSR: Lifting Image Diffusion to Video Super-Resolution via Hybrid Temporal Modeling with Only 4$\times$RTX 4090s},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BQ5YR2LG}},
  note         = {Machine review of arXiv:2506.08529}
}
abstract

Diffusion models have significantly advanced video super-resolution (VSR) by enhancing perceptual quality, largely through elaborately designed temporal modeling to ensure inter-frame consistency. However, existing methods usually suffer from limited temporal coherence and prohibitively high computational costs (e.g., typically requiring over 8 NVIDIA A100-80G GPUs), especially for long videos. In this work, we propose LiftVSR, an efficient VSR framework that leverages and elevates the image-wise diffusion prior from PixArt-$\alpha$, achieving state-of-the-art results using only 4$\times$RTX 4090 GPUs. To balance long-term consistency and efficiency, we introduce a hybrid temporal modeling mechanism that decomposes temporal learning into two complementary components: (i) Dynamic Temporal Attention (DTA) for fine-grained temporal modeling within short frame segment ($\textit{i.e.}$, low complexity), and (ii) Attention Memory Cache (AMC) for long-term temporal modeling across segments ($\textit{i.e.}$, consistency). Specifically, DTA identifies multiple token flows across frames within multi-head query and key tokens to warp inter-frame contexts in the value tokens. AMC adaptively aggregates historical segment information via a cache unit, ensuring long-term coherence with minimal overhead. To further stabilize the cache interaction during inference, we introduce an asymmetric sampling strategy that mitigates feature mismatches arising from different diffusion sampling steps. Extensive experiments on several typical VSR benchmarks have demonstrated that LiftVSR achieves impressive performance with significantly lower computational costs.

Figures

Figures reproduced from arXiv: 2506.08529 by the authors.

Figure 1
Figure 1. Visual and speed comparisons on both real-world and AI-generated videos. Our proposed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of our LiftVSR. Our LiftVSR is built upon a pre-trained DiT model. It [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the inference process. The input video is divided into multiple segments, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Visual comparison of our method with state-of-the-art methods on synthetic low-quality [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of our method with state-of-the-art methods on real-world videos from [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Temporal profiles comparision. We select a row of pixels from the generated video and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Comparison of results with different sampling steps. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Comparison of results across different segments with and without AMC module. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparisons on real-world videos. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Qualitative comparisons on synthetic low-quality videos. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 26 canonical work pages

  1. [1]

    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, Do- minik 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. 9

    arXiv 2023

  2. [2]

    Video super-resolution transformer

    Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021

    arXiv 2021

  3. [3]

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

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

    work page 2021

  4. [4]

    Basicvsr++: Improv- ing video super-resolution with enhanced propagation and alignment

    Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improv- ing video super-resolution with enhanced propagation and alignment. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972–5981, 2022

    work page 2022

  5. [5]

    Investigating tradeoffs in real-world video super-resolution

    Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962–5971, 2022

    work page 2022

  6. [6]

    Diffusion forcing: Next-token prediction meets full-sequence diffusion.Advances in Neural Information Processing Systems, 37:24081–24125, 2024

    Boyuan Chen, Diego Martí Monsó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion.Advances in Neural Information Processing Systems, 37:24081–24125, 2024

    2024

  7. [7]

    Pixart- δ: Fast and controllable image generation with latent consistency models.arXiv preprint arXiv:2401.05252, 2024

    Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- δ: Fast and controllable image generation with latent consistency models.arXiv preprint arXiv:2401.05252, 2024

    arXiv 2024

  8. [8]

    Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    arXiv

Show all 68 references
  1. [9]

    Flatten: optical flow-guided attention for consistent text-to-video editing.arXiv preprint arXiv:2310.05922, 2023

    Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan- Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. Flatten: optical flow-guided attention for consistent text-to-video editing.arXiv preprint arXiv:2310.05922, 2023

    arXiv

  2. [10]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. InProceedings of the IEEE international conference on computer vision, pages 764–773, 2017

    2017

  3. [11]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023

    2023 arXiv

  4. [12]

    Venhancer: Generative space-time enhancement for video generation

    Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024

    2024 arXiv

  5. [13]

    Prompt-to-prompt image editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    2022 arXiv

  6. [14]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  7. [15]

    Video diffusion models.Advances in Neural Information Processing Systems, 35:8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models.Advances in Neural Information Processing Systems, 35:8633–8646, 2022

    2022

  8. [16]

    Video super-resolution via bidirectional recurrent convolutional networks.IEEE transactions on pattern analysis and machine intelligence, 40(4):1015–1028, 2017

    Yan Huang, Wei Wang, and Liang Wang. Video super-resolution via bidirectional recurrent convolutional networks.IEEE transactions on pattern analysis and machine intelligence, 40(4):1015–1028, 2017

    2017

  9. [17]

    Video super- resolution with recurrent structure-detail network

    Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super- resolution with recurrent structure-detail network. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XII 16, pages 645–660. Springer, 2020

    2020

  10. [18]

    Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation

    Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3224–3232, 2018

    2018

  11. [19]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019. 10

    2019

  12. [20]

    Imagic: Text-based real image editing with diffusion models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6007–6017, 2023

    2023

  13. [21]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 5148–5157, 2021

    2021

  14. [22]

    Mucan: Multi-correspondence aggregation network for video super-resolution

    Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. InComputer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X 16, pages 335–351. Springer, 2020

    2020

  15. [23]

    Diffvsr: Enhancing real-world video super-resolution with diffusion models for advanced visual quality and temporal consistency.arXiv preprint arXiv:2501.10110, 2025

    Xiaohui Li, Yihao Liu, Shuo Cao, Ziyan Chen, Shaobin Zhuang, Xiangyu Chen, Yinan He, Yi Wang, and Yu Qiao. Diffvsr: Enhancing real-world video super-resolution with diffusion models for advanced visual quality and temporal consistency.arXiv preprint arXiv:2501.10110, 2025

    2025 arXiv

  16. [24]

    Lsdir: A large scale dataset for image restoration

    Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775–1787, 2023

    2023

  17. [25]

    Vrt: A video restoration transformer.IEEE Transactions on Image Processing, 2024

    Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer.IEEE Transactions on Image Processing, 2024

    2024

  18. [26]

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

    2022

  19. [27]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. InProceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136–144, 2017

    2017

  20. [28]

    Diffbir: Toward blind image restoration with generative diffusion prior

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffbir: Toward blind image restoration with generative diffusion prior. InEuropean Conference on Computer Vision, pages 430–448. Springer, 2024

    2024

  21. [29]

    Learning trajectory-aware trans- former for video super-resolution

    Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware trans- former for video super-resolution. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687–5696, 2022

    2022

  22. [30]

    Video-p2p: Video editing with cross-attention control

    Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024

    2024

  23. [31]

    Evalcrafter: Benchmarking and evaluating large video generation models

    Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2024

  24. [32]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    2022

  25. [33]

    Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048, 2024

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048, 2024

    2024 arXiv

  26. [34]

    Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study

    Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition worksh...

    2019

  27. [35]

    Openvid-1m: A large-scale high-quality dataset for text-to-video generation

    Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 11

    2024 arXiv

  28. [36]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  29. [37]

    Fatezero: Fusing attentions for zero-shot text-based video editing

    Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932–15942, 2023

    2023

  30. [38]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn 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

    2022

  31. [39]

    Frame-recurrent video super- resolution

    Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super- resolution. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6626–6634, 2018

    2018

  32. [40]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  33. [41]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  34. [42]

    Detail-revealing deep video super-resolution

    Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. InProceedings of the IEEE international conference on computer vision, pages 4472–4480, 2017

    2017

  35. [43]

    Tdan: Temporally-deformable alignment network for video super-resolution

    Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360–3369, 2020

    2020

  36. [44]

    Ntire 2017 challenge on single image super-resolution: Methods and results

    Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. InProceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114–125, 2017

    2017

  37. [45]

    Deformable non-local network for video super-resolution.IEEE Access, 7:177734–177744, 2019

    Hua Wang, Dewei Su, Chuangchuang Liu, Longcun Jin, Xianfang Sun, and Xinyi Peng. Deformable non-local network for video super-resolution.IEEE Access, 7:177734–177744, 2019

    2019

  38. [46]

    Exploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 2555–2563, 2023

    2023

  39. [47]

    Seedvr: Seeding infinity in diffusion transformer towards generic video restoration

    Jianyi Wang, Zhijie Lin, Meng Wei, Yang Zhao, Ceyuan Yang, Fei Xiao, Chen Change Loy, and Lu Jiang. Seedvr: Seeding infinity in diffusion transformer towards generic video restoration. arXiv preprint arXiv:2501.01320, 2025

    2025 arXiv

  40. [48]

    Exploiting diffusion prior for real-world image super-resolution.International Journal of Computer Vision, 132(12):5929–5949, 2024

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution.International Journal of Computer Vision, 132(12):5929–5949, 2024

    2024

  41. [49]

    Edvr: Video restoration with enhanced deformable convolutional networks

    Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019

    2019

  42. [50]

    Real-esrgan: Training real-world blind super-resolution with pure synthetic data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. InProceedings of the IEEE/CVF international conference on computer vision, pages 1905–1914, 2021

    1905

  43. [51]

    Exploring video quality assessment on user generated contents from aesthetic and technical perspectives

    Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. InProceedings of the IEEE/CVF International Conference on ...

    2023

  44. [52]

    Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. InProceedings of the IEEE/CVF International Conference ...

    2023

  45. [53]

    One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Processing Systems, 37:92529–92553, 2024

    Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Processing Systems, 37:92529–92553, 2024. 12

    2024

  46. [54]

    Animesr: Learning real-world super- resolution models for animation videos.Advances in Neural Information Processing Systems, 35:11241–11252, 2022

    Yanze Wu, Xintao Wang, Gen Li, and Ying Shan. Animesr: Learning real-world super- resolution models for animation videos.Advances in Neural Information Processing Systems, 35:11241–11252, 2022

    2022

  47. [55]

    Diffir: Efficient diffusion model for image restoration

    Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095–13105, 2023

    2023

  48. [56]

    Star: Spatial-temporal augmentation with text-to-video models for real-world video super-resolution.arXiv preprint arXiv:2501.02976, 2025

    Rui Xie, Yinhong Liu, Penghao Zhou, Chen Zhao, Jun Zhou, Kai Zhang, Zhenyu Zhang, Jian Yang, Zhenheng Yang, and Ying Tai. Star: Spatial-temporal augmentation with text-to-video models for real-world video super-resolution.arXiv preprint arXiv:2501.02976, 2025

    2025 arXiv

  49. [57]

    Easyanimate: A high-performance long video generation method based on transformer architecture.arXiv preprint arXiv:2405.18991, 2024

    Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture.arXiv preprint arXiv:2405.18991, 2024

    2024

  50. [58]

    Video enhancement with task-oriented flow.International Journal of Computer Vision, 127:1106–1125, 2019

    Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow.International Journal of Computer Vision, 127:1106–1125, 2019

    2019

  51. [59]

    Rerender a video: Zero-shot text-guided video-to-video translation

    Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. InSIGGRAPH Asia 2023 Conference Papers, pages 1–11, 2023

    2023

  52. [60]

    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

    2024

  53. [61]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    2024 arXiv

  54. [62]

    Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations

    Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. InProceedings of the IEEE/CVF international conference on computer vision, pages 3106–3115, 2019

    2019

  55. [63]

    A feature-enriched completely blind image quality evaluator.IEEE Transactions on Image Processing, 24(8):2579–2591, 2015

    Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator.IEEE Transactions on Image Processing, 24(8):2579–2591, 2015

    2015

  56. [64]

    The unrea- sonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unrea- sonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  57. [65]

    I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models

    Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, Jingren Zhou, and Alibaba Group. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models

  58. [66]

    Realviformer: Investigating attention for real-world video super-resolution

    Yuehan Zhang and Angela Yao. Realviformer: Investigating attention for real-world video super-resolution. InEuropean Conference on Computer Vision, pages 412–428. Springer, 2024

    2024

  59. [67]

    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

    2024

  60. [68]

    Deformable detr: Deformable transformers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020. 13 Appendix A Architecture A.1 More detials on Training Our proposed LiftVSR is constructed b...

    2010 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.