{"id":"1d316fea-bccd-4aad-91f2-46a404db507e","arxiv_id":"2506.06023","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A diffusion model fine-tuned on synthetically degraded stereo videos simultaneously generates a consistent stereo pair and restores low-resolution or compressed input, outperforming prior stereo generators on low-quality video.","lead":"This paper describes a video model that turns a single low-quality video into a pair of high-quality left-eye/right-eye views, combining stereo generation with video enhancement in one system. It matters because existing stereo generators assume clean input video, so they fail on typical user footage, and this system is trained on synthetic scenes to handle real noisy, blurry, compressed videos.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative 'outperforms' claim rests on in-domain synthetic metrics; real-world evidence is qualitative and lacks geometric stereo validation, leaving the central real-world generalization claim unsupported.","rationale":"The reader's weakest assumption—that the synthetic fine-tuning distribution is representative of real-world low-quality monocular videos—is indeed the central load-bearing premise. My analysis agrees on this point but sharpens it: the quantitative evidence (Table 3) is in-domain synthetic, so it cannot support the real-world generalization claim. Additionally, the real-world consistency metrics (LPIPS/CLIP view) are not diagnostic of stereo correctness; they measure perceptual similarity, not parallax accuracy. The paper's own acknowledged failure on specular/complex materials is direct evidence that the synthetic distribution does not cover important real-world appearance classes. This does not mean the method is invalid; it means the evidence presented is insufficient to support the fullest version of the central claim. The reader's CONDITIONAL verdict appropriately reflects that the method is plausible but the real-world claims are not fully demonstrated. Since my concern does not shift the verdict but adds a more precise technical justification, UNCHANGED is the appropriate recommendation. A concrete out-of-distribution test with geometric evaluation would settle whether the generalization concern actually lands.","tokens_in":13502,"tokens_out":2839,"duration_ms":30728,"concrete_test":"Evaluate the fine-tuned model on an out-of-distribution benchmark with known ground-truth depth/disparity, e.g., Sintel (synthetic but featuring complex materials, humans, and diverse lighting) or real stereo videos such as KITTI/DrivingStereo, applying the same degradation protocol (Eq. 1). Report (a) LPIPS against ground-truth right views, and (b) a geometric stereo error: estimate disparity from the generated left/right pair with an off-the-shelf stereo matcher (e.g., RAFT-Stereo) and compare against the known disparity map. If LPIPS degrades markedly relative to Table 3, or if geometric error is not better than StereoCrafter with FMA-Net restoration, the claims of real-world generalization and quantitative superiority would be substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims the method can be fine-tuned on a small synthetic stereo dataset and applied to low-quality real-world videos for simultaneous stereo generation and restoration, outperforming existing approaches. The quantitative support (Table 3) is LPIPS computed on 97 Kubric/ShapeNet test videos rendered with the same pipeline, same 14 object classes, same 458 environment maps, and the same degradation family used in training. This is fundamentally an in-distribution test of fitting, not generalization to real-world footage. For real-world inputs, the evidence is restricted to qualitative figures (Fig. 2, 4, 5), Table 1 (LPIPS/CLIP view and temporal consistency), and a small user study (Table 2, 15 viewers each scoring 5 of 15 videos). These metrics do not establish correct stereo geometry: LPIPS view consistency measures perceptual similarity between left and right views—a degenerate solution that outputs nearly identical views, or hallucinates plausible but geometrically wrong content, could score well. The paper itself admits a failure on specular and highly reflected materials (supplemental Fig. 6) because the training data lacks complex material appearance. Thus the load-bearing premise—that the small synthetic fine-tuning distribution is representative enough of low-quality real-world videos to transfer—is not actually tested by the paper's quantitative evaluations. The user study captures subjective quality on a handful of curated scenes, not the breadth of real-world degradation types, human faces, text, complex lighting, or materials. Without an out-of-distribution quantitative evaluation, the central claim of real-world applicability and superior performance on low-resolution inputs remains unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Restereo, a pipeline that fine-tunes a video diffusion model (initialized from StereoCrafter) on a small synthetic stereo dataset generated with Kubric/ShapeNet, adding degradation augmentation so that a single model can simultaneously restore a low-quality monocular input and generate a consistent stereo left/right video. The method trains both a left-to-right branch (conditioned on a warped occlusion mask) and a left-to-left restoration branch, and at inference runs both branches with histogram matching. Experiments compare against StereoCrafter and StereoDiffusion on synthetic and real videos, reporting LPIPS/CLIP consistency metrics, a small user study, and ablations on augmentation and training-data size.","tokens_in":13711,"tokens_out":5385,"duration_ms":49376,"significance":"If the central claims hold, this is a valuable contribution: a single model performing both stereo video generation and restoration, trained without large-scale real stereo video data. The idea of joint restoration and stereo generation with degradation augmentation is novel and practically useful, and the proposed use of synthetic data with ground-truth depth avoids the expensive preprocessing pipeline of prior work. The ablations on data size and augmentation are informative, and the commitment to release code and data is a strength. The main weaknesses are in evaluation: the primary quantitative generation-quality result is measured on a test set drawn from the same synthetic distribution as training, and the real-world evaluation does not verify stereo geometric correctness. These issues are addressable with additional experiments or with more carefully scoped claims.","major_comments":[{"comment":"The quantitative claim of outperforming existing approaches on low-resolution inputs is based solely on LPIPS computed on 97 Kubric test videos generated with the same renderer, the same 14 ShapeNet classes, the same 458 environment maps, and the same degradation family used in training. This is an in-distribution test and does not support the abstract's assertion that the method 'can be ... applied to low-quality real-world videos.' The real-world evidence consists of qualitative figures (Figs. 4–5), consistency metrics (Table 1), and a small user study (Table 2), none of which compare generated right views to ground-truth stereo. To support the central claim, the paper should either add a quantitative evaluation on real stereo videos with ground truth (or a held-out dataset with a different distribution) or explicitly scope the quantitative claim to the synthetic distribution.","section":"§4.3, Table 3"},{"comment":"The view consistency metric LPIPS (and CLIP cosine similarity) measures perceptual similarity between left and right views, so a degenerate solution that outputs nearly identical views would score well even with zero disparity. The paper does not report any geometric validation (e.g., disparity consistency, epipolar error, or warping error against a known stereo pair), so the low LPIPS view score does not establish correct stereo geometry. Given the paper's own admission of failure on specular and highly reflective materials (Sec. 4.2, supplemental Fig. 6), a geometric check is needed to rule out plausible-but-incorrect stereo content.","section":"§4.3, Table 1"},{"comment":"The fine-tuning data is limited to 14 ShapeNet object classes, simple materials, and degradation types from a Real-ESRGAN-style augmentation family; the paper acknowledges that this leads to failure on complex material appearance. This admitted limitation directly constrains the central claim of applicability to 'low-quality real-world videos,' which often contain diverse materials, textures, and scene types. The paper should either present evidence that the method handles a broader range of real-world content (e.g., a more diverse real-world test set with multiple scene categories) or relax the claim accordingly.","section":"§3.2 and §4.2"}],"minor_comments":[{"comment":"The sentence 'All test videos are not unseen during training' is ambiguous; presumably it should read 'All test videos are unseen during training.' Please clarify, because the validity of the quantitative results depends on this.","section":"§4.1"},{"comment":"The user study uses 15 participants, each scoring only 5 of 15 videos, and the standard deviations overlap between methods (e.g., 3D Stereo Effect 3.81±0.11 vs 4.07±0.10). A significance test or a larger participant pool would strengthen the claim of 'consistent improvements.'","section":"§4.3, Table 2"},{"comment":"The paper compares only with StereoCrafter and StereoDiffusion, while the abstract claims to outperform 'existing approaches.' Since other training-based methods (SpatialMe, StereoConversion, ImmersePro) are discussed in the related work, adding at least one such comparison or qualifying the claim would be appropriate.","section":"§4.1 / §4.2"},{"comment":"The disparity scaling S and the morphological dilation kernel size for the warped mask are set without ablations; reporting sensitivity of these choices would help reproducibility and robustness assessment.","section":"§3.3"},{"comment":"The notation in Eq. (1) is a bit inconsistent, since x is used for both left and right views after saying superscripts are omitted; please clarify the definitions and the exact role of the 'Up' operation in the training versus testing settings.","section":"§4.1, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the paper is generally well written, but the evaluation does not yet support the abstract's broad claim of applicability to low-quality real-world videos. I would encourage the editor to require either a quantitative real-world stereo evaluation or a careful rephrasing of the central claim. Also, the 'not unseen' typo in Sec. 4.1 should be fixed, as it directly affects the perceived validity of the test set."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's core idea is a single diffusion model that does both stereo video generation and restoration, using dual left-to-right and left-to-left training branches and Real-ESRGAN-style degradation augmentation on Kubric synthetic data. That combination is genuinely new relative to StereoCrafter and StereoDiffusion, and the ablations (Table 4) show the augmentation and data scale matter. Fine-tuning from StereoCrafter on 958 synthetic videos and getting qualitative wins on real low-res footage is a plausible and useful direction.\n\nThe soft spot is exactly where the stress-test lands. The quantitative 'outperforms' claim (Table 3) is LPIPS on 97 Kubric test videos drawn from the same renderer, object classes, environment maps, and degradation family used for training. That is in-distribution fitting evidence, not real-world generalization. The real-world support is qualitative figures, a 15-participant user study with 5 videos each, and LPIPS/CLIP view-consistency scores that measure perceptual similarity, not geometric stereo correctness. A degenerate solution that makes left and right views nearly identical would score well on LPIPS view consistency. The paper itself admits failure on specular and highly reflective materials because the training data lacks complex material appearance. So the abstract's claim of applying to low-quality real-world videos and outperforming is only partially supported.\n\nNone of this is fatal. The internal logic is coherent, the pipeline makes sense, and the synthetic-data-plus-degradation recipe is a reasonable path. The paper is honest about the specular failure. Missing code/data and no error bars on main tables are minor at this stage. The deeper fix is an out-of-distribution evaluation: real low-res videos with a quantitative proxy, or at least a larger and more diverse user study, plus a check on stereo geometry beyond perceptual similarity.\n\nWho gets value: researchers working on stereo video generation and restoration, especially those who want to avoid large real stereo datasets. It deserves a serious referee; I would send it to review rather than desk reject, with a request that the authors address the generalization gap.","headline":"A genuinely new dual-branch recipe for joint stereo generation and restoration, with solid ablations but real-world generalization claims outrunning the in-domain synthetic metrics.","tokens_in":14380,"tokens_out":1526,"would_cite":true,"duration_ms":15607,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single video diffusion model can turn a low-resolution monocular video into a restored, consistent stereo pair.","keywords":["stereo video generation","video restoration","diffusion models","synthetic data","warped mask conditioning","low-resolution video","latent video diffusion","view consistency"],"falsifier":"Run the released model on real low-quality videos with content outside the 14 ShapeNet classes, such as human faces, text, and highly reflective materials, and compare the restored left view and generated right view against StereoCrafter with Real-ESRGAN preprocessing; visible inconsistency or no sharpness gain on these inputs would weaken the claim that the method applies generally to low-quality real-world videos, and the paper already reports such a failure for specular highlights.","tokens_in":13252,"feed_emoji":"🎥","tokens_out":5931,"duration_ms":53181,"temperature":0.7,"pith_summary":"The paper claims that one video diffusion model can take a low-resolution monocular video and produce both a restored left view and a consistent right view, simultaneously performing stereo generation and restoration. Existing stereo video methods assume high-quality input and only inpaint occluded regions, whereas this method fine-tunes the model on degraded synthetic stereo clips and conditions it on warped masks, so one model learns both restoration and stereo generation. The practical appeal is that training uses a relatively small synthetic dataset with ground-truth depth, avoiding expensive real stereo capture, and the method is reported to outperform StereoCrafter and StereoDiffusion on low-resolution inputs in LPIPS, CLIP consistency, and user ratings. The central claim is that this transfer from synthetic degraded data to real low-quality video works.","feed_headline":"One model restores low-res video and makes 3D stereo","feed_subtitle":"Trained only on synthetic clips, it sharpens the left view and generates a consistent right view in one pass.","key_machinery":"The machinery is a fine-tuned latent video diffusion U-Net, built on the Stable Video Diffusion architecture and initialized from StereoCrafter, trained with two randomly sampled branches: a left-to-right branch whose conditioning input is a disparity-warped, degraded left-view latent concatenated with a resized binary warped mask, and a left-to-left branch whose conditioning is a degraded left-view latent with a zero mask. Both branches supervise with mean-square error against clean target latents, and only the temporal layers are trainable. Training data comes from Kubric-rendered ShapeNet objects with 458 environment maps, degraded with blurring, downsampling, Gaussian noise, and JPEG compression using fixed random seeds per video, following the Real-ESRGAN degradation recipe. During inference the same two branches denoise with the warped mask and zero mask, and histogram matching aligns brightness between the two output views.","core_discovery":"On the paper's own terms, the discovery is that restoration and stereo generation do not need separate models: one diffusion U-Net, fine-tuned on degraded synthetic stereo videos and conditioned on a binary warped mask, can simultaneously sharpen the input left view and synthesize a consistent right view. The left-to-right branch inpaints the disoccluded regions marked by the warped mask, while the left-to-left branch with a zero mask teaches the network to restore the input itself. Because both branches share weights and are trained with the same degradation augmentation, the model learns to enhance and generate under a single consistent inference pass; the paper reports that the generated pairs are sharper and more view-consistent than StereoCrafter with Real-ESRGAN or FMA-Net preprocessing.","pith_inferences":["If the transfer assumption holds, the same mask-conditioned two-branch recipe could extend to other pairwise or multi-view tasks, such as generating consistent views at larger baselines or simultaneous denoising and novel-view synthesis from degraded input.","The reported failure on specular highlights suggests the limiting factor is material and lighting diversity in the synthetic data, not the architecture, so adding procedurally varied materials and lighting would be a direct, testable improvement.","Because the model currently processes 16-frame clips, applying it to long videos requires temporal chunking; testing consistency across chunk boundaries is an obvious next evaluation.","The fixed-seed degradation augmentation is likely load-bearing for temporal restoration consistency; ablating it with independent per-frame degradations would reveal how much of that consistency depends on this choice."],"forward_implications":["A low-quality monocular video no longer needs a separate super-resolution pass before stereo conversion; the same model restores and generates both views.","Synthetic data with ground-truth depth is enough to fine-tune a stereo-video generator and restorer, avoiding large real-world stereo capture and stereo-matching pipelines.","The model produces sharper left and right views than StereoCrafter with Real-ESRGAN or FMA-Net preprocessing, with better LPIPS and CLIP view consistency.","Generation quality improves as the fine-tuning dataset grows from 10 to 958 videos, so scaling synthetic data diversity is a direct route to better results.","Because both branches share weights and run in one model, no additional restoration network or extra storage is needed at inference."],"supporting_citations":[{"why":"Supplies the pretrained weights the method initializes from and the main baseline for low-quality stereo generation comparisons.","marker":"[53]"},{"why":"Provides the degradation recipe of blurring, downsampling, Gaussian noise, and JPEG compression used to synthesize low-quality training data.","marker":"[45]"},{"why":"Renders the synthetic stereo training videos with ground-truth depth and left and right camera pairs.","marker":"[12]"},{"why":"Contributes the disparity-based forward warping and mask estimation used to build the warped video and occlusion mask.","marker":"[44]"},{"why":"Estimates video depth at inference time to produce the warped input and mask for real-world videos.","marker":"[16]"},{"why":"Defines the Stable Video Diffusion architecture whose temporal layers are fine-tuned.","marker":"[3]"}],"fun_headline_variants":["One diffusion model restores and generates stereo video","Single model sharpens left view and synthesizes right view","Restore and generate: single model for low-res stereo video","One U-Net restores and generates stereo video jointly","Low-res in, 3D out, no separate upscaler"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fine-tuning distribution, synthetic ShapeNet objects with simple materials plus Real-ESRGAN-style degradations, is representative enough of real-world low-quality monocular video that the model will restore and generate consistent stereo on real footage.","fun_headline_variants_meta":{"raw":{"variants":["One diffusion model restores and generates stereo video","Single model sharpens left view and synthesizes right view","Restore and generate: single model for low-res stereo video","One U-Net restores and generates stereo video jointly","Low-res in, 3D out, no separate upscaler"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0009,"raw_usage":{"total_tokens":3833,"prompt_tokens":862,"completion_tokens":2971,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":2889}},"tokens_in":478,"tokens_out":2971,"duration_ms":20248,"temperature":1.0,"reasoning_tokens":2889,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T06:01:53.384725+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released model on real low-quality videos with content outside the 14 ShapeNet classes, such as human faces, text, and highly reflective materials, and compare the restored left view and generated right view against StereoCrafter with Real-ESRGAN preprocessing; visible inconsistency or no sharpness gain on these inputs would weaken the claim that the method applies generally to low-quality real-world videos, and the paper already reports such a failure for specular highlights.","supporting_citations":[{"cited_title":"Kubric: A scalable dataset generator","cited_arxiv_id":null,"evidence_quote":"Renders the synthetic stereo training videos with ground-truth depth and left and right camera pairs."},{"cited_title":"Stereodiffusion: Training-free stereo image generation using latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Contributes the disparity-based forward warping and mask estimation used to build the warped video and occlusion mask."}],"review_version":1}