{"id":"57dcfc7e-f1bd-4b42-80a6-e56cfb4bcfd8","arxiv_id":"2607.21628","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Wavelet-domain phase injection with low-frequency randomization improves realism and semantic consistency of sim-to-real translation, improving VLM planner ADE and FDE by about 5% on CARLA videos.","lead":"Sim-to-real translation that preserves structure by injecting the source image's phase in the wavelet domain, instead of the Fourier domain, produces more realistic and semantically consistent synthetic-to-real images and videos. The method needs no paired data, adds almost no inference cost, and on CARLA driving videos it reduced a vision-language planner's trajectory error by about 5%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core mechanism—latent-domain DT-CWPT phase injection—is untested: no experiment separates phase preservation from wavelet locality, LFR, or fine-tuning, so the reported gains may not be attributable to phase.","rationale":"The reader's weakest assumption correctly identifies the latent-domain transfer of Oppenheim–Lim as a structural premise that is never isolated. I agree because the paper's entire novelty—'Wavelet Phase Diffusion'—depends on phase being the carrier of structure in the DT-CWPT coefficients of a VAE latent. The reported ablations are properly targeted at wavelet-vs-Fourier and LFR on/off, but they do not test whether phase injection itself is responsible for the gains; a no-phase control is missing. This does not invalidate the empirical comparison, but it makes the central scientific claim conditional. The reader's conditional verdict is therefore appropriate. My proposed check is a single, concrete experiment that would settle the concern: replace source phase with noise phase and measure the impact on all headline metrics. If the control is comparable, the paper's mechanism attribution fails; if it degrades, the assumption is validated. I do not see a reason to move beyond CONDITIONAL: the concern is about mechanism attribution and missing isolation, not a demonstrated mathematical error.","tokens_in":11271,"tokens_out":10846,"duration_ms":116965,"concrete_test":"Train a phase-scrambled control identical to ψ-PD in every respect except that the source phase in all high-frequency packets is replaced by the corresponding noise phase (i.e., set ∠H(x)_l to ∠H(ϵ)_l in Eq. 6 before reconstruction), while keeping LFR, cutoff sweeps, and training data unchanged. Evaluate this control on vKITTI→KITTI (KID, mIoU, CLIP-IQA) and CARLA (ADE/FDE). If the control matches ψ-PD within noise, phase preservation is not the operative mechanism; if it degrades substantially, the latent-domain phase assumption is empirically supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ψ-PD achieves simultaneous realism, semantic consistency, and planning utility through wavelet-domain phase injection. The load-bearing assumption is introduced in Section 3: Oppenheim–Lim phase-magnitude separation is applied directly to a VAE latent x via a complex transform T. This transfer is never isolated experimentally. The ablations compare ψ-PD w/o LFR against NeuralRemaster, which tests wavelet-vs-Fourier bases but not phase-vs-no-phase; both methods inject source phase. No control randomizes or removes source phase while keeping LFR and training data fixed. Therefore the reported gains could in principle come from DT-CWPT locality, the low-frequency randomization, or the fine-tuning protocol rather than from phase-carrying geometric structure.\n\nA further tension makes this concern sharper: Eq. (5) replaces the low-frequency packet L(x) with noise, so the only source-phase information retained is in high-frequency wavelet packets. If Oppenheim–Lim is to justify structural preservation, that structure must be encodable in these high-frequency phases alone. The paper provides no evidence that latent-space DT-CWPT phase has this property. If the phase channel is not actually structure-bearing, the mechanism named in the title and abstract is not grounded, and the method's success would need a different explanation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ψ-PD, a diffusion-based sim-to-real translation method that performs phase injection in the dual-tree complex wavelet packet transform (DT-CWPT) domain rather than the Fourier domain, and replaces the low-frequency wavelet packet with the corresponding packet of random noise (Low-Frequency Randomization, LFR). The method is applied in the latent space of FLUX for images and Wan 2.2 for videos, after fine-tuning on unpaired open-domain data. Experiments on vKITTI→KITTI report state-of-the-art KID, CLIP-IQA, and mIoU among non-trivial methods, and on CARLA report that ψ-PD is the only method to improve both realism and VLM-planner ADE/FDE. The paper also demonstrates zero-shot instance-level translation via spatially varying cutoff maps.","tokens_in":11549,"tokens_out":4811,"duration_ms":47647,"significance":"If the reported results are robust, the contribution is useful: ψ-PD requires no paired data or inference-time conditioning, introduces negligible overhead, supports instance-level control, and is evaluated with a downstream planning metric, which is a valuable complement to standard realism metrics. The ablation design comparing DT-CWPT vs. Fourier phase injection and LFR on/off is conceptually appropriate, and the authors provide code and models. The central mechanistic claim, however, is not isolated: no experiment separates phase preservation from wavelet locality, LFR, or fine-tuning, and the empirical evidence consists of single-run point estimates without error bars.","major_comments":[{"comment":"The causal role of phase is not established by the ablations. Comparing ψ-PD w/o LFR with NeuralRemaster varies the transform (DT-CWPT vs. Fourier) and packet-wise thresholding while both methods inject source phase; comparing full ψ-PD with ψ-PD w/o LFR varies only the low-frequency packet. No arm randomizes or removes the source phase in the high-frequency packets while keeping DT-CWPT locality, LFR, and fine-tuning fixed. The reported gains could therefore be due to wavelet locality, LFR, or the training protocol rather than to phase-carrying structure. Add a no-phase control (e.g., use noise phase in all packets under the same training setup) and report realism, consistency, and planning metrics for it.","section":"§4.2, Eq. (5); §5.6"},{"comment":"The Oppenheim–Lim premise is imported from pixel-domain signals to VAE latents and DT-CWPT coefficients without validation. Because Eq. (5) discards the low-frequency packet, the only source-phase information that survives is in high-frequency packets; the paper supplies no evidence that latent-space high-frequency wavelet phase encodes geometric or semantic structure. Provide a diagnostic (e.g., comparing phase-only vs. magnitude-only latent reconstructions for depth/semantic metrics, or a control using noise phase) to ground the transfer.","section":"§3 and §4.2"},{"comment":"All reported metrics are single-run point estimates with no standard errors, confidence intervals, or significance tests. The headline margins are small (e.g., KID 4.41 vs. 4.52; ADE reduction 5.4%), so without repeated runs or paired tests the \"only method\" and improvement claims are not statistically supported. Report multiple runs/seeds or, at minimum, bootstrap confidence intervals.","section":"Tables 1 and 2"},{"comment":"The final inference cutoff r and depth J appear to be selected from the evaluation benchmark: the text states inference uses r = half Nyquist and J=4, while §5.6 sweeps r to construct Pareto frontiers. This is tuning on the test set unless a validation protocol is defined. Specify how r and J are chosen (held-out scenes/data) and report sensitivity of Tables 1–2 to r and J.","section":"§5.1 and §5.6"}],"minor_comments":[{"comment":"The symbol L is used both for the low-frequency packet and for the number of high-frequency packets L=2^J−1. This collision makes equations such as Eq. (6) confusing; rename one of them.","section":"§4.1"},{"comment":"f_Nyq is not defined for the VAE latent space; specify its value/units. Also clarify whether LFR is applied always at inference or with probability p=0.8 as during training.","section":"§4.2, Eq. (4)"},{"comment":"The axis label 'Downsample from   to   ' is incomplete; the figure and caption should be completed.","section":"Figure 3"},{"comment":"It would help to state the number of test frames/images per method and whether the same random seed is used. Also specify whether CLIP-IQA is reference-free or reference-based, since that affects interpretation.","section":"Tables 1–2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, well-scoped engineering paper that gets the right baseline comparisons in, and the instance-level translation trick is genuinely new and useful. But the headline mechanism — that wavelet phase is what preserves structure — is not actually tested. The ablations compare wavelet vs Fourier and LFR on/off, but every method injects source phase. There is no control where you randomize or remove phase while keeping wavelet locality, LFR, and training fixed. So the reported gains could come from wavelet support, the low-frequency randomization, or the fine-tuning protocol, not from phase carrying structure. Also note Eq. (5) replaces the low-frequency packet with noise, so any source structure has to be carried by high-frequency wavelet phases alone. The paper gives no evidence that latent-space DT-CWPT phase has that property. That is a real gap, not a nitpick.\n\nWhat is good: the method is simple and the evaluation is thoughtful. The DT-CWPT phase injection is a genuine extension over Fourier-based NeuralRemaster, and LFR is a plausible mechanism for decoupling from synthetic illumination. The zero-shot instance-level translation via spatial cutoff maps is a nice capability that falls out of the design, and the downstream VLM planning metric is a welcome addition to sim-to-real evaluation. The ablation comparing wavelet vs Fourier and LFR vs no-LFR is the right experimental design.\n\nThe soft spots beyond the missing phase control: all metric tables are single-run point estimates with no error bars or significance tests. Some of the claimed margins are small (e.g., KID 4.41 vs 4.52), and without variance we cannot tell if they are real. The final operating point (r and J) appears to be selected by a sweep on the evaluation benchmark, with no stated validation split. The paper says code is available but nothing is directly verifiable from the manuscript. These are all addressable.\n\nIf the method works, it works because injecting source phase in a localized transform plus randomizing low-frequency content gives a better realism–consistency trade-off. That is plausible. But the title and abstract claim phase is the reason, and the paper does not earn that specific claim. A control that randomizes phase in the wavelet domain, or one that uses a different structure-bearing transform, would settle it.\n\nWho should read this: people working on sim-to-real translation, diffusion-based editing, and closed-loop evaluation with VLM planners. It deserves a serious referee, but the review should ask for a phase-randomization control, multi-seed stats, and a clearly separated validation split. I would not desk-reject it.","headline":"Solid engineering extension of NeuralRemaster with a genuine new capability, but the phase mechanism is never isolated; treat the strong claims as conditional.","tokens_in":12055,"tokens_out":2004,"would_cite":true,"duration_ms":21957,"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":"Wavelet-domain phase injection with low-frequency randomization achieves photorealistic, semantically consistent sim-to-real translation without paired data or conditioning, and measurably improves a downstream vision-language planner.","keywords":["sim-to-real translation","wavelet phase diffusion","dual-tree complex wavelet packet transform","phase preservation","low-frequency randomization","diffusion models","semantic consistency","autonomous driving"],"falsifier":"Run two matched vKITTI→KITTI translations with ψ-PD: one with genuine source phase injection and one with source and noise phases swapped. If swapped-phase outputs keep mIoU and DepSSIM at the same level, phase is not the carrier of the reported structural consistency. As a second check, profile pixel values across a high-contrast edge: Fourier phase injection should show overshoot and ringing while DT-CWPT should not.","tokens_in":11132,"feed_emoji":"🚗","tokens_out":5639,"duration_ms":50126,"temperature":0.7,"pith_summary":"The paper's central claim is that the phase-magnitude decomposition known from signal processing also works when applied to the latent space of a diffusion model, provided the transform is spatially localized. By replacing Fourier phase preservation with a dual-tree complex wavelet packet transform, the method avoids the global ringing and boundary leakage that made Fourier phase injection hallucinate false geometry. A second component randomizes the low-frequency wavelet packet, letting the model ignore synthetic illumination and generate in-distribution real appearance. On vKITTI→KITTI, the method is the only one to lead simultaneously in realism (KID 4.41), perceptual quality (CLIP-IQA 0.561), and semantic consistency (mIoU 43.50) among translation methods; on CARLA video it reduces a vision-language planner's average trajectory error by 5.4% and final error by 5.1%. The same spatially local cutoff map enables zero-shot instance-level translation, where only a selected object becomes photorealistic while the rest of the frame remains untouched.","feed_headline":"Cut planner error 5.4% with wavelet-phase sim-to-real translation","feed_subtitle":"No paired data or conditioning: wavelet phase keeps lane and traffic semantics while adding real-world appearance.","key_machinery":"The central mechanism is phase injection in the dual-tree complex wavelet packet transform (DT-CWPT) domain. The transform decomposes a latent into one real low-frequency packet and complex high-frequency packets with localized spatial support; source phase is injected packet-wise according to a cutoff radius map, so structural information is preserved locally without global spectral coupling. The second component, Low-Frequency Randomization (LFR), replaces the source low-frequency packet with the noise packet, removing the synthetic illumination prior. Together they form the structured noise that drives the diffusion backbone, and the same cutoff map enables instance-level translation. The","core_discovery":"The paper proposes ψ-PD, a training recipe for sim-to-real translation that needs no paired real-synthetic data and no inference-time conditioning signals. The method constructs diffusion noise by taking a Gaussian sample and, in the dual-tree complex wavelet packet (DT-CWPT) domain, injecting the phase of the source latent while randomizing the low-frequency packet. Because wavelet packets have localized support, the phase constraint acts where it is needed instead of coupling all frequency bands globally as the Fourier transform does. The paper reports that this yields a strictly better realism–consistency operating point than prior phase-preserving diffusion: best KID 4.41, best CLIP-IQA","pith_inferences":["A natural stress test is to swap the roles: inject source magnitude with noise phase. If semantic consistency survives, the mechanism is not specifically phase; if it collapses, the phase hypothesis is confirmed.","The same DT-CWPT phase injection could likely be ported to other generative backbones without retraining them, since the construction lives in the transform and noise schedule rather than the model weights.","The spatial cutoff map suggests a cheap interactive editing tool: a user-provided mask or depth map could set per-region cutoff radii, making photorealism a continuous local dial rather than a global setting.","Low-Frequency Randomization might be imported into Fourier-domain methods as a partial fix for global illumination bias, although the ringing problem would remain; comparing the two would isolate how much of the gain comes from locality versus randomization."],"forward_implications":["Sim-to-real pipelines can gain realism and semantic consistency without paired datasets or control modules, removing a major data-engineering cost.","Because the overhead is under 0.15% of inference time, the method can be dropped into existing image and video diffusion backbones with negligible latency cost.","Downstream planners benefit directly: translated CARLA video reduces predicted-trajectory error by about 5%, suggesting closed-loop simulation evaluation becomes more reliable.","Instance-level translation is available zero-shot from a per-pixel cutoff map, enabling selective photorealism editing of single objects in a scene.","Localized phase injection avoids the geometric hallucinations (wrong vehicle orientation, dashboard artifacts) that Fourier-domain phase preservation produced in the qualitative comparisons."],"fun_headline_variants":["Wavelet phase injection cuts sim-to-real planner error by 5.4%","No paired data: wavelet phase preserves semantics in sim-to-real","Localized wavelet phase beats global Fourier for sim-to-real","Wavelet phase diffusion improves realism without paired data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that phase carries spatial structure in a VAE's latent codes and in DT-CWPT coefficients just as it does in pixel-domain images; the paper provides no isolated experiment testing that transfer, so if latent-space phase does not align with geometry the mechanism loses its grounding.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet phase injection cuts sim-to-real planner error by 5.4%","No paired data: wavelet phase preserves semantics in sim-to-real","Localized wavelet phase beats global Fourier for sim-to-real","Wavelet phase diffusion improves realism without paired data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1298,"prompt_tokens":808,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":552,"tokens_out":490,"duration_ms":5117,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:02:32.181809+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run two matched vKITTI→KITTI translations with ψ-PD: one with genuine source phase injection and one with source and noise phases swapped. If swapped-phase outputs keep mIoU and DepSSIM at the same level, phase is not the carrier of the reported structural consistency. As a second check, profile pixel values across a high-contrast edge: Fourier phase injection should show overshoot and ringing while DT-CWPT should not.","supporting_citations":[],"review_version":1}