{"id":"831bcf85-e6c0-4704-8440-832c24ee4c3c","arxiv_id":"2506.08004","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Dynamic view synthesis from a monocular video is achieved by redesigning the noise initialization of a pretrained video diffusion model using a recursive interpolation and a stochastic latent modulation.","lead":"This paper proposes a training-free method for dynamic view synthesis from a single video by altering the noise fed into a pretrained video diffusion model. It claims state-of-the-art results by combining a recursive noise interpolation scheme with a stochastic latent inpainting step, without changing model weights.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"K-RNR's recursive latent, even after AdaIN normalization, is not shown to lie on the frozen denoiser's input manifold: Eq. (2) amplifies x0 beyond any valid forward-process state, so the reported gains could be source-signal leakage rather than inverse-problem solving.","rationale":"The central claim is that state-of-the-art dynamic view synthesis can be achieved purely by manipulating the initial noise of a frozen video diffusion model. The load-bearing requirement is that the constructed input is a valid diffusion state for the model being used. The paper shows in Fig. 4 and Fig. 5 that epsilon^(k) becomes increasingly aligned with x0 and that its mean and variance explode, then applies AdaIN to re-center and re-scale it. That operation fixes only the first two moments of the marginal distribution; it does not fix the joint distribution, the correlation structure, or the effective SNR relative to the timestep embedding. Since Eq. (2) gives a signal coefficient greater than one in the large-k limit, the normalized input can be far from the forward-process latents on which CogVideoX was trained. This is not merely a formal objection: if the input is essentially a scaled copy of x0, the high reconstruction scores may be measuring the model reproducing the source render, not synthesizing novel views. The proposed test would distinguish source leakage from genuine recursive-noise benefit by matching the cosine similarity to x0 with a valid convex combination and comparing outputs. I do not dispute the correctness of Propositions 4.1 and 4.2, but they are elementary; the empirical claim is what carries the paper, and it rests on this unverified distributional premise. The reader's rejection is therefore appropriate, and no verdict change is needed.","tokens_in":19215,"tokens_out":10340,"duration_ms":141517,"concrete_test":"On 100 OpenVid videos, compute xinit = AdaIN(epsilon^(10), epsilon^(3)) at t = 0.95T and measure (i) the sliced-Wasserstein distance to N(0, I) and (ii) the cosine similarity between xinit and x0. Then run the full pipeline with a control initial noise n = AdaIN(s * x0 + sqrt(1 - s^2) * epsilon, epsilon^(3)), choosing s so that cos(n, x0) equals the measured cosine of xinit. If the control matches K-RNR on PSNR and camera-pose metrics, the gains are explained by source-signal injection; if it does not, the K-RNR recursion contributes beyond simple leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 defines epsilon^(k) by recursively adding sqrt(alpha_bar_t) x0, with closed form Eq. (2). For fixed 0 < alpha_bar_t < 1 and k -> infinity, the x0 coefficient is sqrt(alpha_bar_t)/(1 - sqrt(1 - alpha_bar_t)) > 1, so the 'noise' is dominated by the source latent, not by Gaussian noise. AdaIN (Fig. 6) only matches the mean and variance of epsilon^(delta); it does not restore the joint distribution or the effective SNR expected by CogVideoX at t = 0.95T. The denoiser was trained on x_t = sqrt(alpha_bar_t) x0 + sqrt(1 - alpha_bar_t) epsilon, where the signal coefficient is < 1 and decays with t. Feeding a latent whose signal component is amplified creates a mismatch between the timestep conditioning and the actual input, so the reverse trajectory is not an inverse of the forward process. Fig. 4(a) actually confirms that cos(epsilon^(k), x0) grows with k, so the method is better described as injecting x0 into the noise channel than as solving an inverse problem. If the reported PSNR/SSIM gains in Table 2 mainly reflect reconstruction of the already-rendered source frames, the central claim that 'solely manipulating initial noise' achieves state-of-the-art novel-view synthesis is unsubstantiated. No manifold-level or distributional test is provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free dynamic view synthesis method that operates entirely in the noise initialization phase of a frozen CogVideoX video diffusion model. It introduces K-order Recursive Noise Representation (K-RNR), which recursively combines the VAE-encoded latent with a DDIM-inverted latent and then applies adaptive instance normalization (AdaIN), and Stochastic Latent Modulation (SLM), which copies latent values from visible background regions into occluded regions via a random permutation. Experiments on OpenVid-1M, DAVIS, and AI-generated videos report improvements over GCD, TrajectoryAttention, DaS, TrajectoryCrafter, and ReCamMaster across visual quality, camera pose accuracy, and synchronization metrics. The paper frames the task as an inverse problem and claims that state-of-the-art performance can be achieved without weight updates or auxiliary modules.","tokens_in":19496,"tokens_out":5458,"duration_ms":66833,"significance":"If the central claim held, the paper would demonstrate that novel-view synthesis can be repurposed from a frozen video diffusion model by only manipulating initial noise, which is practically attractive and conceptually interesting. The paper includes formal propositions with proofs in the appendix, a broad set of baselines, and both full-reference and no-reference metrics. However, the core mechanism reduces to injecting the encoded latent into the noise channel, and the reported gains are not tied to a demonstrated inverse-problem solution. The evaluation lacks error bars, code, and an isolated assessment of newly visible regions, so the significance is currently below what the paper claims.","major_comments":[{"comment":"The closed-form expression in Eq. (2) shows that the coefficient of x0 is sqrt(ᾱ_t) * (1 - (sqrt(1-ᾱ_t))^k) / (1 - sqrt(1-ᾱ_t)), which grows with k and tends to a limit strictly larger than 1 as k→∞. Therefore the monotone increase in cosine similarity reported in Fig. 4(a) is a mathematical consequence of the recursion, not an empirical demonstration that the initialization is 'aligned' in a way that supports inverse problem solving. More importantly, at t=0.95T the frozen denoiser was trained on inputs with signal coefficient sqrt(ᾱ_t) < 1; substituting a latent whose signal component is amplified beyond this coefficient induces a train/test distribution shift. The paper provides no distributional or manifold-level evidence that AdaIN(ε^(k), ε^(δ)) lies in the support of inputs the denoiser expects at that timestep. The observed reconstruction gains may therefore reflect source-signal leakage rather than a valid solution to an inverse problem.","section":"§4.2, Eq. (2)"},{"comment":"SLM fills occluded latent regions by randomly copying values from visible, depth-near background positions. This is an ad-hoc stochastic permutation without any guarantee of spatial or temporal coherence; indeed, the limitations paragraph in §6 acknowledges that SLM 'can produce unstable or incoherent results when large regions become newly visible.' The quantitative evaluation does not isolate the quality of the completed regions: PSNR/SSIM/LPIPS in Table 2 are computed against ground-truth frames that do not contain ground truth for out-of-view content, so those numbers mostly reflect reconstruction of the already-visible content. A per-region or no-reference evaluation of newly synthesized areas is required before the claim of 'physically plausible synthesis in occluded areas' is supported.","section":"§4.4, Algorithm 1"},{"comment":"The central state-of-the-art claim is supported only by point estimates without error bars or significance tests. The hyperparameters k=10 and δ=3 are selected using the same evaluation set on which the ablations in Appendix E.2 and E.3 are performed, so the reported comparisons are at risk of selection bias. The baselines also rest on different base models (SVD, Wan, and CogVideoX), and no code or detailed evaluation protocol is released, making it difficult to verify that the comparisons are controlled. Reporting variance across seeds and trajectories and releasing code would be necessary to support the headline claim.","section":"Tables 1-2 and §5 (Implementation)"}],"minor_comments":[{"comment":"The title 'δblations' is a typo; it should be 'δ ablations'.","section":"Appendix E.3"},{"comment":"The notation '¯at' in the SNR formula 'SNR(t) = ¯at/(1−¯at)' should be 'ᾱ_t' with a subscript t for consistency with the rest of the paper.","section":"§3.2"},{"comment":"The text refers to 'Table 9' for the ablation results, but the ablation table is embedded in a figure environment ('Figure 9: Ablation on K-RNR...'), making the cross-reference inconsistent.","section":"Main text, Ablation Studies"},{"comment":"The use of AdaIN is described only as matching mean and variance; a citation to the original adaptive instance normalization work (e.g., Huang and Belongie, 2017) and a short explanation of why these statistics are the relevant ones would improve clarity.","section":"§4.2, Adaptive K-RNR"},{"comment":"The paper fixes the CFG scale at 6.0 but does not report sensitivity to this choice; given that CFG is known to interact strongly with noise initialization, a brief sensitivity analysis would aid reproducibility.","section":"§5, Implementation"}],"recommendation":"reject","confidential_remarks":"The paper has a clear empirical scope and the proofs in the appendix are correct, but the central conceptual claim is not supported: the recursive noise representation is mathematically equivalent to amplifying the source latent in the initialization, and no evidence is provided that this stays on the denoiser's input manifold. The SLM component is an unvalidated heuristic, and the SOTA claim rests on point estimates with selection of hyperparameters on the evaluation set. I do not see how these load-bearing issues can be fixed without substantially reworking the method and evaluation, so I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'd read this one for the recipe, not for the claims. The combination of K-RNR and Stochastic Latent Modulation for training-free DVS via pure noise initialization is genuinely new, and the qualitative results are plausible. The ablations show that the recursive mixing does improve reconstruction and that SLM helps fill occluded regions. That is real, if incremental, evidence that the heuristic does something.\n\nThe theory is thin, though. Proposition 4.1 is a direct corollary of setting alpha_bar_T=0, and Proposition 4.2 is just the closed form of a geometric series. The paper presents the growing cosine similarity in Fig. 4a as an empirical discovery, but Eq. (2) makes clear that the x0 coefficient grows geometrically with k, so the alignment gain is a mathematical necessity, not a new mechanism. The stress-test point about the input distribution is fair: for large k, the x0 coefficient exceeds 1, so epsilon^(k) is no longer a plausible forward-process latent. AdaIN only matches mean and variance; it does not restore the joint distribution or the effective SNR that CogVideoX expects at timestep 0.95T. Calling this \"inverse problem solving\" overstates what is happening—the method is better described as injecting the source latent into the noise channel.\n\nThe empirical evaluation is the weakest part. The custom benchmark of 1100 videos with five baselines is not enough to support a state-of-the-art claim. There are no error bars, no significance tests, and no code. The two most relevant training-free baselines are missing. The hyperparameters k=10 and delta=3 are tuned on the evaluation set, and the appendix shows that k=8 degrades; that suggests the method is sensitive to tuning. The PSNR gain over ReCamMaster is 14.50 vs 13.77, which is not decisive.\n\nTo be fair, the paper is honest about limitations and the appendix contains substantial ablations. The authors seem to have a real feel for what works empirically. The central idea—redesigning the noise initialization instead of adding modules or fine-tuning—is worth exploring. But as written, the load-bearing claims are unsupported.\n\nI would send this to a serious referee, not desk-reject it, because the direction is novel and the empirical results are non-trivially positive. But I'd expect major revision: proper comparisons, error bars, a distributional sanity check of the modulated latent, and a more measured tone. If those come through, it could be a useful paper for practitioners in video editing and camera control.","headline":"A new training-free noise-modulation recipe for dynamic view synthesis, but the state-of-the-art claim outruns the evidence.","tokens_in":20098,"tokens_out":2059,"would_cite":false,"duration_ms":26407,"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":"This paper shows that dynamic view synthesis reduces to editing the starting noise of a frozen video diffusion model.","keywords":["dynamic view synthesis","monocular video","noise initialization","zero-terminal SNR","DDIM inversion","video diffusion model","recursive noise representation","latent modulation"],"falsifier":"Initialize the frozen CogVideoX denoiser with the AdaIN-normalized K-RNR latent and with a standard Gaussian latent of identical mean and variance, then compare reconstruction PSNR and identity preservation: if the Gaussian latent performs as well, the recursive structure is not doing the work, and if the normalized latent is far from the training noise distribution the outputs should visibly degrade.","tokens_in":18941,"feed_emoji":"🎥","tokens_out":10108,"duration_ms":109970,"temperature":0.7,"pith_summary":"Dynamic view synthesis (DVS) is treated here as an inverse problem in the noise initialization phase: the paper claims that a frozen video diffusion model can synthesize novel views from a single video with no weight updates, auxiliary modules, or synthetic pre-training. The obstacle is the Zero-Terminal SNR Collapse Problem: when the training schedule forces the terminal signal coefficient $\\bar\\alpha_T=0$, the forward map $\\Phi_T(x_0,\\epsilon)=\\epsilon$ is not injective in $x_0$, so DDIM inversion cannot uniquely recover the source latent. The authors' fix, K-order Recursive Noise Representation, recursively mixes the VAE-encoded latent with the DDIM-inverted latent in closed form, and Adaptive K-RNR applies AdaIN to keep the scale stable. Newly visible regions are completed by Stochastic Latent Modulation, which samples visible background latents and permutes them into occlusion-masked positions. A sympathetic reader would care because, if true, camera control becomes a property of input design rather than architecture, making DVS accessible without retraining.","feed_headline":"Starting-noise edits alone outperform trained view synthesis","feed_subtitle":"A recursive noise initialization lets a frozen diffusion model synthesize novel views with no weight updates.","key_machinery":"The load-bearing object is the K-order Recursive Noise Representation, the closed-form identity $\\epsilon^{(k)} = \\left(\\sum_{i=1}^k \\sqrt{\\bar\\alpha_t}(\\sqrt{1-\\bar\\alpha_t})^{i-1}\\right)x_0 + (\\sqrt{1-\\bar\\alpha_t})^{k}\\epsilon_{\\mathrm{inv}}$, which converts the collapsed terminal latent into a tunable mixture of source structure and inverted noise. Two auxiliary mechanisms carry the rest: Adaptive K-RNR, which applies adaptive instance normalization (AdaIN), $\\mathrm{AdaIN}(\\epsilon^{(k)},\\epsilon^{(\\delta)})$, to suppress the recursion's exploding mean and variance, and Stochastic Latent Modulation, the visibility-aware permutation operator that fills occlusion-masked latent regions by resampling visible background values. Together these replace the weight updates and camera branches of prior systems.","core_discovery":"On its own terms, the paper's central discovery is that the non-injectivity induced by zero-terminal SNR is bypassable by constructing the initialization noise recursively. With the VAE-encoded latent $x_0$ as pivot, the recursion $\\epsilon^{(1)}=\\sqrt{\\bar\\alpha_t}x_0+\\sqrt{1-\\bar\\alpha_t}\\epsilon_{\\mathrm{inv}}$, $\\epsilon^{(k)}=\\sqrt{\\bar\\alpha_t}x_0+\\sqrt{1-\\bar\\alpha_t}\\epsilon^{(k-1)}$ has the closed form shown in Eq. (1) and its continuous-depth generalization Eq. (2). The paper shows cosine similarity to $x_0$ rises with $k$, that the norm approaches the expected noise scale, and that the exploding mean and variance can be tamed by $\\tilde{x}_{\\mathrm{init}} = \\mathrm{AdaIN}(\\epsilon^{(k)}, \\epsilon^{(\\delta)})$ with $\\delta=3$. It then claims that Stochastic Latent Modulation, a permutation that copies latent values from visible depth-near background locations into occluded positions, lets the frozen denoiser complete regions revealed by camera motion. The reported experiments on 1000 OpenVid videos plus DAVIS and AI-generated videos show the method ahead of GCD, TrajectoryAttention, DaS, TrajectoryCrafter, and ReCamMaster on visual quality, camera pose, and synchronization metrics.","pith_inferences":["Editorial inference: because the zero-terminal-SNR collapse is schedule-level rather than model-specific, K-RNR should transfer to other diffusion models trained with the same schedule; porting it to a second backbone would be a direct test.","Editorial inference: the paper's own stated limitation that SLM can be unstable when large regions become newly visible implies the method's guarantees are regime-dependent, so large rotation or translation sweeps are the cleanest stress test.","Editorial inference: a per-video choice of $k$ could replace the fixed $k=10$ by stopping recursion when the latent norm crosses the expected Gaussian noise scale.","Editorial inference: replacing SLM's random permutation with deterministic visibility-aware inpainting in latent space would isolate whether randomness is actually needed for plausible completions."],"forward_implications":["Any video diffusion model trained with a zero-terminal-SNR schedule becomes a candidate for training-free dynamic view synthesis, because K-RNR restores the alignment that DDIM inversion loses at $\\bar\\alpha_T=0$.","Since no weights change, the same frozen backbone can be reused across many camera trajectories and scenes, with only the depth estimation and noise construction recomputed per video.","The reported results imply that camera-conditioned attention branches and per-video LoRA fine-tuning are not required for competitive DVS, shifting the design space from architecture to initialization.","Adaptive K-RNR's hyperparameters matter: the ablations show quality peaking around $k=6$-$7$ and $\\delta=3$ and degrading beyond them, so the gain is tied to these operating points.","If SLM generalizes, occlusion completion in DVS can be treated as a latent-sampling problem rather than an inpainting or 3D-reconstruction problem."],"supporting_citations":[{"why":"This reference supplies the frozen CogVideoX video diffusion model whose latent space and denoiser carry the whole method.","marker":"[57]"},{"why":"This reference provides DDIM inversion, the deterministic inversion whose collapse under zero-terminal SNR motivates K-RNR.","marker":"[45]"},{"why":"This reference argues that noise schedules should enforce zero terminal SNR, the training choice that creates the non-injectivity problem.","marker":"[30]"},{"why":"This reference supplies the camera-conditioning and dynamic point-cloud rendering recipe that the method adapts for novel views.","marker":"[58]"},{"why":"This reference provides the monocular depth estimation used to lift videos into dynamic 3D point clouds.","marker":"[19]"},{"why":"This reference is the OpenVid-1M dataset on which the main quantitative comparison is performed.","marker":"[36]"}],"fun_headline_variants":["Recursive noise initialization for training-free view synthesis","Inverting video diffusion with recursive noise latents","Training-free dynamic views via recursive noise representation","Frozen diffusion model synthesizes novel views via noise edits","Zero-shot view synthesis by recursive latent noise modulation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's load-bearing premise is that the recursively mixed and re-centered noise is still a plausible input for the frozen denoiser; the paper shows scale and similarity improve with recursion depth but never shows the normalized latent lies in the distribution the model was trained on.","fun_headline_variants_meta":{"raw":{"variants":["Recursive noise initialization for training-free view synthesis","Inverting video diffusion with recursive noise latents","Training-free dynamic views via recursive noise representation","Frozen diffusion model synthesizes novel views via noise edits","Zero-shot view synthesis by recursive latent noise modulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00014,"raw_usage":{"total_tokens":1172,"prompt_tokens":967,"completion_tokens":205,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":133}},"tokens_in":583,"tokens_out":205,"duration_ms":3075,"temperature":1.0,"reasoning_tokens":133,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:21:00.248996+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Initialize the frozen CogVideoX denoiser with the AdaIN-normalized K-RNR latent and with a standard Gaussian latent of identical mean and variance, then compare reconstruction PSNR and identity preservation: if the Gaussian latent performs as well, the recursive structure is not doing the work, and if the normalized latent is far from the training noise distribution the outputs should visibly degrade.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF winter conference on applications of computer vision","cited_arxiv_id":null,"evidence_quote":"This reference argues that noise schedules should enforce zero terminal SNR, the training choice that creates the non-injectivity problem."}],"review_version":1}