{"id":"de9be778-1982-4ea4-b66f-f573c6b20654","arxiv_id":"2508.06392","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FVGen uses GAN-based adversarial distillation and softened reverse KL divergence to compress a video diffusion teacher for novel-view synthesis into a four-step student with comparable quality.","lead":"FVGen distills a multi-step video diffusion model for novel-view synthesis into a student that generates novel views in four sampling steps, cutting sampling time by more than 90% with comparable visual quality. It may make sparse-view 3D reconstruction practical when pre-trained video diffusion models would otherwise need many sampling passes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mode collapse in the four-step student is not ruled out: the reverse-KL/GAN objective is mode-seeking, and the reported visual-quality metrics do not bound view diversity or downstream reconstruction consistency.","rationale":"The reader's weakest assumption—that the four-step GAN+softened-RKL student faithfully preserves the teacher's distribution and view diversity—is the same load-bearing condition I identify. I ground it further in the objective's mode-seeking character rather than in the abstract's silence. The full text is mostly illegible, so I cannot rule out that the paper already contains diversity or reconstruction ablations; the corrupted tables and equation fragments are consistent with an ablation study but do not reveal the actual numbers. I found no internal contradiction or obvious mathematical error, only a missing empirical comparison that is essential to the stated downstream-reconstruction motivation. Therefore I do not move the reader's UNVERDICTED verdict; I recommend an additional check that would either settle the concern or confirm the claim. No independent support such as code or formal verification is visible in the corrupted text, so the burden remains on the reported experiments.","tokens_in":13682,"tokens_out":5329,"duration_ms":66808,"concrete_test":"On the same sparse-view scenes used in the paper, sample 8 view sequences from FVGen and from the teacher with identical random seeds. Compute (i) mean pairwise LPIPS within each sampled set and (ii) per-frame FID of FVGen outputs relative to teacher outputs. Then feed each sampled set into the downstream reconstruction module and report per-scene Chamfer distance / PSNR for both. If FVGen's pairwise LPIPS is substantially below the teacher's while per-frame FID is similar, the visual-quality metric is masking view collapse, and the central claim would need to be conditioned on a diversity-preserving term or on explicit reconstruction metrics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the student's conditional distribution over view sequences to match the teacher closely enough that downstream sparse-view reconstruction from student samples is not systematically degraded. The paper's distillation objective minimizes a softened reverse KL term, which is mode-seeking for the student distribution, together with an adversarial GAN loss; both terms encourage sharp, high-likelihood outputs while providing no explicit diversity pressure. With only four sampling steps, such a student can plausibly attain the abstract's 'similar or even better visual quality' on per-frame fidelity metrics while dropping low-probability but reconstruction-critical variations (e.g., alternative plausible geometry or hallucinated content in unseen regions). The abstract reports visual quality and sampling time, but I see no direct comparison of teacher vs student sample diversity, nor a downstream reconstruction robustness metric over multiple student samples. This is a missing-support concern: the evaluation as described is consistent with a collapsed student that scores well on FID/PSNR-style metrics yet fails the paper's stated goal of supplying useful dense observations for sparse-view reconstruction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FVGen, a framework for accelerating novel-view synthesis from video diffusion models (VDMs). The proposed method distills a multi-step denoising teacher model into a few-step student model using a combination of generative adversarial networks and softened reverse KL-divergence minimization. The central empirical claims are: (1) the student generates the same number of novel views as prior VDM-based methods in as few as four sampling steps; (2) visual quality is similar or better; and (3) sampling time is reduced by more than 90%. The intended downstream benefit is faster sparse-view 3D reconstruction, where VDMs would otherwise need multiple runs.","tokens_in":14004,"tokens_out":1563,"duration_ms":19356,"significance":"If the empirical claims hold, FVGen addresses a real bottleneck in VDM-based novel-view synthesis and sparse-view reconstruction: slow autoregressive or multi-step sampling. A four-step student that preserves quality while reducing runtime by an order of magnitude would be practically significant. The contribution is a distillation methodology applied to video diffusion models, and the paper's framing around downstream reconstruction is a useful application context. However, the significance can be assessed only after the experiments and technical details are verified; the abstract alone provides no quantitative support, and the full text supplied to me is corrupted and unreadable.","major_comments":[{"comment":"The provided manuscript text is corrupted (mojibake) and unreadable. Sections, equations, tables, and experimental results cannot be accessed. As a result, the central empirical claims — four-step sampling, >90% time reduction, and similar/better visual quality — cannot be verified. This is a load-bearing issue for an empirical paper. The authors must provide a readable manuscript with full experimental details.","section":"Full text"},{"comment":"The abstract reports only qualitative claims ('similar (or even better) visual quality', 'more than 90% time reduction') without any metrics, baseline names, dataset identifiers, or error bars. Since the full text is unavailable, these claims are unsupported. The revision must include concrete evaluation numbers and a clear comparison protocol.","section":"Abstract"},{"comment":"The distillation objective combines a softened reverse KL term and an adversarial GAN loss, both of which are mode-seeking for the student distribution. The paper's downstream goal requires the student to supply diverse observations for sparse-view reconstruction. The abstract reports no teacher-student diversity comparison and no downstream reconstruction robustness metric over multiple student samples. This leaves open the risk that per-frame quality metrics are good while reconstruction-critical variation is lost. Please add an explicit diversity or downstream-consistency evaluation, e.g., measuring variance across student samples or comparing reconstruction metrics against teacher-generated samples.","section":"Objective (GAN + softened reverse KL)"}],"minor_comments":[{"comment":"The phrase 'particularly when working with sparse input views (more than 2)' is vague. Clarify the view-count regime and whether 'more than 2' means 3+, or includes 2.","section":"Abstract"},{"comment":"The term 'same number of novel views' is ambiguous: it should state the exact number of generated views and compare it explicitly to prior VDM-based methods.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The full text is corrupt in the version I received; I cannot judge the technical soundness or experimental validity. I recommend the editors obtain a clean version before further consideration. The mode-collapse concern raised by the stress-test is plausible and should be addressed explicitly in the experiments; it is not a circularity problem, but a missing-support issue for the claimed downstream benefit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is clear and sensible: take the established GAN-plus-reverse-KL distillation recipe, apply it to video diffusion models, and target novel-view synthesis with only four sampling steps. If the 90% time reduction holds without quality loss, that would make sparse-view reconstruction pipelines much more practical, especially when you need multiple VDM samples. I think that's a real contribution, not just a tweak. The authors are also honest about the motivation—VDMs are slow when run repeatedly for spatial coverage.\n\nWhat I can't do is verify the claims. The full text I received is badly corrupted, so I only have the abstract and some garbled fragments. The abstract gives no numbers, baselines, or dataset details. The stress-test concern about mode collapse is a reasonable question to put to the authors: a GAN plus a mode-seeking reverse KL term can produce sharp samples that look good per-frame while losing diversity that matters for downstream reconstruction. But I don't know whether the paper already addresses that with a diversity metric or downstream reconstruction consistency check. The abstract doesn't mention it, so it's at least a missing piece of evidence.\n\nThat said, this paper deserves a serious referee. The problem is important, the approach is plausible, and the claim is concrete and falsifiable. The right move is to send it to review and ask the authors for a clean, complete version. I'd want the reviewer to check: how the student is initialized, how the teacher is used during distillation, whether the student is evaluated on the same downstream reconstruction task as the teacher, and whether the reported visual quality metrics are accompanied by any diversity or consistency measure.\n\nFor me, I'd want to see the actual paper before citing it or putting it in front of a reading group. But the work is clearly in the realm of something a CV/3D vision venue should look at.","headline":"The four-step video diffusion distillation idea is worth taking seriously, but the corrupted full text means I can't check the actual experiments—so treat the strong claims as unverified.","tokens_in":14381,"tokens_out":1330,"would_cite":false,"duration_ms":18610,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FVGen distills a multi-step video diffusion teacher into a four-step student that generates the same number of novel views with equal or better quality and cuts sampling time by more than 90%.","keywords":["novel view synthesis","video diffusion models","diffusion distillation","adversarial training","reverse KL divergence","few-step sampling","sparse-view 3D reconstruction"],"falsifier":"Run a controlled comparison: for the same sparse-view input, sample many view sets from the four-step student and from the full multi-step teacher; if the student's generated views have measurably lower diversity, or if feeding the student's views into the same reconstruction pipeline yields systematically worse geometry than the teacher's views, the central claim fails. A second check is timing: the claimed reduction of more than 90% should hold end-to-end, not just for the denoising loop.","tokens_in":13644,"feed_emoji":"🎥","tokens_out":4365,"duration_ms":42338,"temperature":0.7,"pith_summary":"FVGen is a method for making novel-view synthesis from sparse input images practical by drastically cutting the sampling cost of video diffusion models. It tries to establish that a multi-step video diffusion teacher can be distilled into a student that produces the same number of novel views in only four denoising steps, with visual quality similar to or better than the teacher, while removing more than 90% of the sampling time. If true, this removes the main practical bottleneck that keeps video-diffusion-based sparse-view reconstruction slow, while preserving enough view consistency and diversity for downstream 3D reconstruction. The paper argues that the right distillation recipe pairs an adversarial GAN loss with a softened reverse KL-divergence objective, rather than a standard one-step regression target.","feed_headline":"Novel views in 4 steps: video diffusion runs 90% faster","feed_subtitle":"Distilling a multi-step video diffusion teacher into a four-step student keeps image quality and cuts sampling time by more than 90 percent.","key_machinery":"The central object is a distilled few-step video diffusion student. The mechanism is distribution-matching distillation rather than one-step regression: the student is optimized so that, starting from random noise, its four-step denoising trajectory yields outputs that are judged real by a GAN discriminator trained against the teacher's samples, and that minimize a softened reverse KL divergence with the teacher's output distribution, which punishes the student where the teacher has probability mass and avoids over-smoothing. This combination carries the claim: the GAN term preserves high-frequency detail, and the softened reverse KL term preserves multi-view consistency and diversity across","core_discovery":"On its own terms, the paper's central discovery is that a video diffusion model used for novel-view synthesis does not need dozens of iterative denoising steps at inference time. FVGen first trains a multi-step video diffusion model as teacher, then distills it into a student that generates the full set of novel views in as few as four sampling steps. The student is trained with a GAN discriminator that matches the perceptual quality and detail of the teacher's output, together with a softened reverse KL-divergence term that keeps the student's distribution close to the teacher's without the mode-averaging that a plain forward KL objective can cause. On real-world datasets, the paper reports","pith_inferences":["Editorially: the GAN-plus-softened-reverse-KL combination may transfer to other conditional video generation tasks, such as inpainting or temporal interpolation, where teacher sampling is the main cost.","Editorially: the strongest untested risk is distribution collapse or reduced view diversity; a natural test is to measure the entropy or FID of generated view sets against the teacher, and to check that reconstruction quality degrades gracefully when student views are fed into a 3D reconstruction module.","Editorially: if the distillation preserves the teacher's view consistency, FVGen could serve as a fast proposal generator for multi-view stereo or as a data-augmentation engine for training reconstruction networks."],"forward_implications":["If correct, sparse-view 3D reconstruction pipelines can run FVGen instead of a full multi-step video diffusion model and keep similar geometry and visual quality.","The same distillation recipe could apply to any pre-trained video diffusion model, not only novel-view synthesis, to obtain few-step generation.","The 90% sampling-time reduction makes repeated video diffusion runs practical, which matters for spatial coverage when more than two sparse views need to be connected.","Four-step inference brings video-diffusion-based novel view synthesis closer to interactive or real-time use.","The student's speed makes it feasible to generate dense observations on demand, filling unseen areas in sparse-view reconstruction without long waits."],"supporting_citations":[],"fun_headline_variants":["Novel views in 4 steps: video diffusion 90% faster","Four steps to fresh views: video diffusion distilled","90% faster novel views via video diffusion distillation","Sparse views to dense views in 4 diffusion steps","Video diffusion view synthesis: 4 steps, 90% speedup"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that a four-step student trained with adversarial and softened reverse-KL losses can match the teacher's full output distribution, including view diversity and multi-view consistency, without systematic bias or mode collapse.","fun_headline_variants_meta":{"raw":{"variants":["Novel views in 4 steps: video diffusion 90% faster","Four steps to fresh views: video diffusion distilled","90% faster novel views via video diffusion distillation","Sparse views to dense views in 4 diffusion steps","Video diffusion view synthesis: 4 steps, 90% speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00137,"raw_usage":{"total_tokens":5398,"prompt_tokens":756,"completion_tokens":4642,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":4559}},"tokens_in":500,"tokens_out":4642,"duration_ms":32778,"temperature":1.0,"reasoning_tokens":4559,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:44:21.493117+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison: for the same sparse-view input, sample many view sets from the four-step student and from the full multi-step teacher; if the student's generated views have measurably lower diversity, or if feeding the student's views into the same reconstruction pipeline yields systematically worse geometry than the teacher's views, the central claim fails. A second check is timing: the claimed reduction of more than 90% should hold end-to-end, not just for the denoising loop.","supporting_citations":[],"review_version":1}