{"id":"0cf567bb-3c37-4eee-8dd1-3efed8846065","arxiv_id":"2507.01012","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DAM-VSR improves video super-resolution by first enhancing a key frame with an image super-resolution model, then using Stable Video Diffusion with a video ControlNet to propagate details while keeping motion aligned.","lead":"This paper presents DAM-VSR, a video super-resolution method that enhances one reference frame with an image super-resolution model and then animates it with a video diffusion model controlled by the low-quality input. It reports top perceptual-quality scores on real-world and AI-generated videos and introduces a bidirectional sampling scheme for longer clips.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The propagation of ISR-enhanced reference details to later frames (Sec 3.2) is unquantified; if it fails under large motion, DAM-VSR degenerates to single-frame ISR plus ControlNet and the SOTA claim is unsupported.","rationale":"After reading the paper in good faith, the framework is plausible and the ablations suggest each component contributes. The reader's weakest_assumption identifies the appearance-propagation step, and I agree this is the load-bearing condition. The entire SOTA claim on real-world and AIGC data depends on the ISR-enhanced reference frame improving all frames through SVD's temporal mechanism. The paper offers only a qualitative assertion in Sec 3.2 and a limitation note in Sec 5; it never measures per-frame quality or detail transfer. The quantitative comparisons in Table 2 aggregate over clips, and on real-world and AIGC data only non-reference metrics are reported. Because the ISR models (SupIR, InvSR) are strong, the first frame alone may score much higher than any baseline frame; averaging can make the whole clip appear superior even if the video diffusion adds no detail to later frames. The ablation in Table 1 is computed on UDM10 with ResShift and shows that training with the disentanglement improves LPIPS, but this is weak evidence of propagation under more complex real-world degradations and larger motions. A supporting gap is the missing degradation-pipeline reference in supplementary Sec 2, which prevents reproduction and leaves the training distribution unclear. The primary concern, however, is the unquantified propagation. The proposed concrete test, per-frame metrics conditioned on motion magnitude plus warped-residual correlation, would settle whether details really propagate. Given these gaps, the reader's CONDITIONAL verdict is appropriate; my read does not change it.","tokens_in":18091,"tokens_out":7409,"duration_ms":189471,"concrete_test":"On VideoLQ and AIGC29 clips, compute per-frame non-reference metrics (e.g., MUSIQ, CLIP-IQA) for DAM-VSR and for the 'w/o Disentanglement During Training and Inference' variant, and estimate per-frame optical flow magnitude. If the metric gain of DAM-VSR over the variant is concentrated in frame 1 and does not extend to frames with large flow, the propagation assumption fails. Additionally, warp frame 1's high-frequency residual to each later frame using estimated flow; if the correlation with that frame's residual is near zero under large motion, detail is not being propagated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, state-of-the-art performance on real-world and AIGC data, depends on the propagation assumption in Sec 3.2: after ISR on the reference frame (Eq. 2), 'these generated realistic details in the reference frame can be propagated to the remaining frames leveraging the powerful temporal generation capabilities of SVD.' The paper does not demonstrate this. Table 2 reports metrics averaged over entire clips, so a single very high-quality first frame, enhanced by SupIR/InvSR, can raise the average even if frames 2 through N are no better than the no-ISR ControlNet variant. The ablation in Table 1 reports only whole-clip numbers on UDM10, whose motions are limited, and the real-world and AIGC evaluations use only non-reference metrics. Sec 5 admits failure when 'ISR does not perform well,' but the relevant failure mode is broader: even a perfect ISR frame may fail to propagate under large inter-frame motion, occlusion, or scene changes, because SVD's temporal attention aligns latent content, not necessarily high-frequency detail. If propagation fails, DAM-VSR degenerates to a motion-only ControlNet with an ISR-boosted first frame, undermining both the temporal-consistency and detail-generation claims. A secondary reproducibility gap is the training degradation pipeline in supplementary Sec 2, whose citation is missing, leaving the training distribution unspecified. The paper should quantify per-frame fidelity and detail transfer across motion levels before the SOTA claim can be accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":18351,"tokens_out":6695,"duration_ms":75117,"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":[{"comment":"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.","section":"Section 3.2, Eq. (2), Table 2, Section 5"},{"comment":"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.","section":"Table 1 and Section 4.2.1"},{"comment":"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.","section":"Table 2 and Section 4.1.1"},{"comment":"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.","section":"Supplementary Section 2 (training strategy)"}],"minor_comments":[{"comment":"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":"Section 3.2"},{"comment":"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.","section":"Section 3.3.1, Eqs. (3)-(4)"},{"comment":"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":"Figures 3 and 4"},{"comment":"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.","section":"Section 3.4"},{"comment":"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.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the disentanglement idea is worth publishing once the propagation assumption is quantified and the evaluation is strengthened. I would encourage the editor to require the release of AIGC29 and the exact training-degradation specification as part of the revision; without these, the SOTA claims cannot be independently checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nCredit 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.\n\nSoft 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.\n\nOn 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.\n\nVerdict: 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.","headline":"Modular ISR+video-diffusion VSR with real promise; the central detail-propagation claim needs direct per-frame evidence before the SOTA talk.","tokens_in":18959,"tokens_out":1797,"would_cite":true,"duration_ms":24923,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["video super-resolution","appearance and motion disentanglement","Stable Video Diffusion","ControlNet","image super-resolution prior","motion-aligned bidirectional sampling","temporal consistency","AIGC video restoration"],"falsifier":"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.","tokens_in":17842,"feed_emoji":"🎬","tokens_out":6488,"duration_ms":63947,"temperature":0.7,"pith_summary":"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.","feed_headline":"One sharp frame can upscale a whole video","feed_subtitle":"A modular framework pairs image super-resolution with a video diffusion model and tops real-world benchmarks.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the pre-trained image-to-video diffusion model (SVD) whose temporal layers propagate reference-frame details and provide temporal consistency.","marker":"[Blattmann et al. 2023]"},{"why":"Provides the ControlNet design that the paper duplicates into a video ControlNet for motion and structure control.","marker":"[Zhang et al. 2023]"},{"why":"SupIR is the ISR method used for appearance enhancement on real-world and AIGC data.","marker":"[Yu et al. 2024]"},{"why":"ResShift is the fidelity-oriented ISR method used for synthetic benchmark experiments.","marker":"[Yue et al. 2024b]"},{"why":"InvSR is the generation-oriented ISR method used for real-world and AIGC data.","marker":"[Yue et al. 2024a]"},{"why":"Establishes that temporal self-attention maps reflect motion dynamics, justifying the 180-degree rotation used in backward sampling.","marker":"[Wang et al. 2024c]"},{"why":"Provides the YouHQ40 benchmark and the evaluation protocol, and is the main diffusion-based VSR baseline.","marker":"[Zhou et al. 2024]"},{"why":"SeedVR is the large-scale end-to-end video diffusion restorer whose data and GPU cost the paper contrasts with its own lightweight approach.","marker":"[Wang et al. 2025]"}],"fun_headline_variants":["Video super-resolution split into appearance and motion","Sharp reference frame drives temporal video upscaling","Image SR meets video diffusion for consistent details","Motion-aligned sampling extends video SR to long clips","One enhanced frame anchors video diffusion super-resolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Video super-resolution split into appearance and motion","Sharp reference frame drives temporal video upscaling","Image SR meets video diffusion for consistent details","Motion-aligned sampling extends video SR to long clips","One enhanced frame anchors video diffusion super-resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1296,"prompt_tokens":934,"completion_tokens":362,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":293}},"tokens_in":550,"tokens_out":362,"duration_ms":4813,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:00:18.184782+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}