{"id":"9779f6c2-2187-46af-9b55-07299a91d10d","arxiv_id":"2509.05441","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A wavelet-based VAE that trains low- and high-frequency branches separately improves image reconstruction and diffusion generation.","lead":"Researchers propose a frequency-aware image autoencoder that splits images into low and high frequency parts with wavelets and optimizes each separately. On ImageNet, this yields sharper reconstructions and better diffusion-based image generation than prior tokenizers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Capacity not controlled: FA-VAE's two-branch architecture confounds frequency decoupling with added parameters in Tables 1 and 3; no matched-capacity coupled baseline is provided.","rationale":"After reading the paper in good faith, the strongest claim is that frequency decoupling is the cause of the improved fidelity. The most direct evidence is Table 3's comparison between FA-VAE and coupled-frequency retrained baselines. However, the appendix states FA-VAE has separate encoders and decoders for low and high frequency, doubling the parameter count relative to a single-branch coupled model. The paper only says 'similar latent dimensionality,' which is not parameter count or compute. This is a textbook confounding: the coupled baseline could underperform simply because it is smaller. The same issue affects Table 1's headline comparison with VA-VAE. The generation experiment adds another confound: LightningDiT with FA-VAE operates on a concatenated latent (z_L, z_H), increasing the latent dimensionality relative to VA-VAE's single latent; this could improve FID independently of frequency decoupling. No generation ablation is run with a coupled baseline. These gaps mean the paper has not yet established that the decoupling mechanism, rather than capacity or latent dimensionality, is responsible for the gains. The frequency-bias analysis (raw band losses) is also under-normalized, but it is secondary because even if the bias characterization is imperfect, the method's improvement is the core claim; the capacity confound threatens that core claim directly. A capacity-matched ablation would settle the issue. Therefore, the appropriate verdict remains conditional: the method shows promise and empirical gains, but the causal interpretation needs a controlled ablation.","tokens_in":16081,"tokens_out":6763,"duration_ms":67512,"concrete_test":"Retrain a coupled-frequency baseline with the same total parameter count as FA-VAE (e.g., a single encoder-decoder with the combined channel width of both branches, or a two-branch architecture with tied weights but joint optimization), using the same training budget, loss terms (including VF, GAN, LPIPS where applicable), and data. Then evaluate Table 3 metrics and LightningDiT generation FID with this baseline's latent under the same fusion/diffusion setup. If this capacity-matched coupled model matches FA-VAE's reconstruction loss and generation FID, decoupling per se is not responsible for the gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that explicitly decoupling low- and high-frequency optimization drives the reported reconstruction and generation gains. The ablation in Table 3 does not control for model capacity: FA-VAE uses two independent encoder-decoder pairs (Appendix, 'Low- and High-Frequency Encoding'), so its parameter count and per-step compute are roughly double that of the coupled-frequency variants, which reuse a single encoder-decoder architecture. The paper only matches 'similar latent dimensionality,' not parameter count or training budget. Thus the large drops in reconstruction loss (e.g., Rec. from 0.0125 to 0.0044 vs VA-VAE) and the generation FID improvement (5.14 to 3.24) could reflect additional capacity or the concatenated latent space used in generation (z_L and z_H fused) rather than the frequency-decoupling mechanism itself. Without a capacity-matched coupled baseline, the causal attribution is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the frequency behavior of latent tokenizers used in latent generative models. It reports that existing VAE variants reconstruct low-frequency content well but underfit high-frequency details, and attributes this to a low-frequency bias in the joint optimization objective. To address this, the paper proposes FA-VAE, which applies a Haar discrete wavelet transform, encodes low- and high-frequency subbands with two independent encoder-decoder pairs, and fuses the resulting latents for downstream diffusion training. Reconstruction experiments on ImageNet report that FA-VAE roughly halves the reconstruction loss of VA-VAE (Table 1) and improves LightningDiT generation FID from 5.14 to 3.24 without CFG (Table 2). The paper also reports an ablation against coupled-frequency variants of several tokenizers (Table 3) and a per-class robustness analysis (Figure 5).","tokens_in":16336,"tokens_out":4258,"duration_ms":47627,"significance":"If the causal claim is established, the contribution is practically valuable: FA-VAE is a simple, modular tokenizer change that yields large reconstruction and generation gains in a standard latent diffusion pipeline, with no generation-time architectural modification. The paper is empirical and uses external metrics (rFID, gFID, LPIPS), so there is no circularity in the evaluation. However, the central attribution of these gains to frequency decoupling is not yet supported because the ablation does not control for model capacity, the frequency-bias analysis uses unnormalized band-wise losses, and Table 1 pools tokenizers trained on different datasets. With matched-capacity controls and normalized frequency metrics, the contribution would be considerably stronger.","major_comments":[{"comment":"The main ablation does not control for capacity. FA-VAE uses two independent encoder-decoder pairs (one for low frequency, one for the concatenated high-frequency tensor), while the coupled-frequency baselines reuse a single encoder-decoder architecture. The paper only states that 'all models are trained with similar latent dimensionality as used in the FA-VAE configuration' (Table 3), not that parameter counts, FLOPs, or training budgets are matched. The large reconstruction gains (e.g., Rec. 0.0125 to 0.0044 vs VA-VAE) could therefore come from the additional parameters/compute or from the different per-branch objectives (e.g., L1 high-frequency loss, no VF/LPIPS on the high branch) rather than from the decoupling mechanism itself. A capacity-matched coupled baseline is required to attribute the gains to frequency decoupling.","section":"Table 3 / Appendix 'Low- and High-Frequency Encoding'"},{"comment":"The claimed 'low-frequency bias' is supported only by raw per-band MSE values L_L and L_H. These bands have very different energies and coefficient counts: after a Haar DWT, natural-image energy is concentrated in the low-pass band, and the high-pass bands contain three orientations. A smaller raw high-frequency loss can simply reflect there being less high-frequency energy to reconstruct, not better optimization. To substantiate the bias claim, the authors should report normalized per-band errors (e.g., L_L/||x_L||^2 and L_H/||x_H||^2) or per-coefficient MSE. This affects the motivation of the paper, not just the presentation.","section":"Section 'Frequency Evaluation of Latent Embeddings' / Table 1"},{"comment":"Table 1 pools tokenizers with different training datasets and budgets. The appendix reveals that DC-AE* models were trained on ImageNet, SAM, FFHQ, and Mapillary Vistas, while KL-VAE* and VQ-VAE* models were trained on OpenImages, and FA-VAE is trained on ImageNet. The headline comparison against VA-VAE is more defensible, but the caption does not disclose these differences, and the statement that FA-VAE 'nearly halves the reconstruction loss of the strongest baseline' conflates architectural contribution with training-data and training-budget differences. Please restrict the main comparison to matched training setups or clearly annotate each row.","section":"Table 1 / Appendix 'Additional Evaluations'"},{"comment":"The generation comparison is between LightningDiT trained on the single VA-VAE latent and LightningDiT trained on the concatenated FA-VAE latent (z_L and z_H fused). This changes the latent dimensionality and the generative modeling problem, not only the frequency optimization. The reported gFID improvement (5.14 to 3.24) therefore conflates the effect of frequency decoupling with the effect of the richer, higher-dimensional latent input. A generation experiment using a coupled-frequency tokenizer with the same fused latent structure would clarify the contribution.","section":"Section 'Latent Fusion for Generative Modeling' / Table 2"}],"minor_comments":[{"comment":"The axis labels '2 × 100', '3 × 100', '4 × 100' are cryptic. Please clarify the frequency axis units or explain the indexing.","section":"Figure 2"},{"comment":"There is a typo: 'Similarly, To assess' should be 'Similarly, to assess'. Also, the notation V AEs appears with inconsistent spacing throughout the manuscript.","section":"Section 'Frequency Evaluation of Latent Embeddings'"},{"comment":"The phrase 'with input wavelet representation instead of input pixel representation' is unclear. The table compares coupled-frequency training on wavelet inputs against decoupled FA-VAE; please state this explicitly.","section":"Table 3 caption"},{"comment":"The term 'fairness' is unusual for per-class reconstruction error. The analysis shows lower top-100 class-wise errors, which is better described as per-class robustness or worst-class fidelity, not fairness in the usual sense.","section":"Section 'Fairness in Latent Representations'"},{"comment":"The sentence 'We experiment with two different f16 tokenizers: one trained without alignment for high frequencies and one with VF loss using DINOv2 for low frequencies' is confusing. Presumably these are two branches of FA-VAE, but the phrasing suggests two separate tokenizers. Please clarify.","section":"Appendix 'Implementation Details'"},{"comment":"Reference [Li et al. 2024b] is cited as 'ICLR 2025 (withdrawn)'. Citing a withdrawn paper is problematic; please replace with the published/available version or remove.","section":"References"},{"comment":"The statement 'We believe our method, if trained on larger datasets, would exhibit even more beneficial results' is speculative and not supported by the reported experiments. Either provide supporting evidence or remove it.","section":"Appendix 'Additional Evaluations'"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clear and useful empirical contribution, but the central causal claim is currently under-supported. The most important fix is a capacity-matched coupled-frequency baseline for both reconstruction and generation, plus normalized frequency-error reporting. If those are added, I would be supportive. I also suggest the authors consider whether the 'fairness' framing and the cross-dataset Table 1 comparison invite unnecessary criticism; both can be tightened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on arXiv:2509.05441. The paper reports a genuine, sizable improvement by splitting the VAE into two branches that handle Haar wavelet low and high subbands separately. Reconstruction loss drops from 0.0105 (VA-VAE) to 0.0044, and LightningDiT trained on the fused latent gets gFID 3.24 versus 5.14 with the same diffusion training budget. Those are solid numbers, and the residual power spectrum in Figure 2 shows the improvement is not just a scalar artifact.\n\nThe new thing is the specific design: two independent encoder-decoder pairs, an asymmetric objective (low branch gets DINOv2 VF alignment; high branch uses plain L1 plus GAN), and a concatenative fusion into a single latent that plugs into LightningDiT. That combination is not in the prior wavelet-based generation literature, which mostly injects frequency processing into the diffusion backbone rather than into the tokenizer.\n\nThe soft spots are real. The biggest is the capacity confound. FA-VAE has two encoder-decoder pairs, so roughly twice the parameters of the coupled variants in Table 3. The paper says these variants were trained with 'similar latent dimensionality,' but that doesn't control parameter count or training budget. Without a matched-capacity coupled baseline, you can't attribute the gains to decoupling; extra capacity alone might buy most of it. That directly weakens the paper's central claim.\n\nSecond, the frequency-bias analysis reports raw per-band MSE. High-frequency wavelet coefficients have intrinsically smaller energy, so a lower high-frequency loss is partly a scale effect. A proper analysis needs normalization by band energy or energy ratios. The residual spectrum figure is more convincing, but the stated bias is overstated.\n\nThird, Table 1 mixes tokenizers trained on different datasets (ImageNet, OpenImages, multi-domain), so the ranking isn't apples-to-apples. The retrained coupled ablation in Table 3 is better than most papers in this area, which partially mitigates this.\n\nMinor: the generation setup says 'no architectural changes' while concatenating two latents, which would change the first-layer input dimension of the diffusion model. Likely just phrasing, but worth fixing.\n\nNet: competent, well-written paper with a plausible method and strong empirical results. The causal attribution is not established, but the method itself is worth taking seriously. I'd send it to peer review and ask for a capacity-matched ablation and normalized frequency analysis. It will be a stronger paper after that. Recommendation: engage with it; it deserves referee time and is likely citable as a practical tokenizer improvement.","headline":"The wavelet two-branch VAE shows real gains on ImageNet reconstruction and generation, but the claim that decoupling is the cause is confounded by unmatched capacity in the ablations.","tokens_in":16790,"tokens_out":3727,"would_cite":true,"duration_ms":40426,"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":"A wavelet-based autoencoder that trains low- and high-frequency latents separately recovers fine image detail that conventional latent tokenizers lose.","keywords":["frequency-aware VAE","Haar wavelet","latent tokenizer","high-frequency fidelity","latent diffusion","image reconstruction","ImageNet","texture detail"],"falsifier":"Train a coupled-frequency variant of the FA-VAE architecture with identical encoder-decoder parameter counts, same loss weights, and same training epochs (matching the decoupled model in everything except the separation of frequency bands), then compare reconstruction loss, LPIPS, and generation FID. If the coupled variant matches FA-VAE's numbers, the decoupling is not the cause of the improvement.","tokens_in":16005,"feed_emoji":"🌊","tokens_out":2706,"duration_ms":26552,"temperature":0.7,"pith_summary":"The paper claims that existing latent image tokenizers systematically over-prioritize low-frequency content, smoothing away the textures, edges, and fine structures that make generated images look real. To fix this, the authors propose FA-VAE, a variational autoencoder that splits an image into low- and high-frequency bands with a Haar wavelet transform, learns a separate latent code for each band, and recombines them with an inverse transform. On ImageNet-1K, FA-VAE nearly halves the reconstruction loss of the strongest prior tokenizer (VA-VAE) and cuts the generation FID of a LightningDiT diffusion pipeline from 5.14 to 3.24. The paper argues that explicit frequency decoupling during tokenizer training is what preserves fine detail, and that this improves both reconstruction and downstream image generation.","feed_headline":"Wavelet split nearly halves image reconstruction loss","feed_subtitle":"Training low- and high-frequency latents separately sharpens textures and cuts generation FID from 5.14 to 3.24.","key_machinery":"The Haar discrete wavelet transform is the load-bearing mechanism: it separates an image into one low-frequency approximation band and three high-frequency detail bands, allowing the training objective to be decoupled into L_low and L_high with separate latents z_L and z_H. The two encoder-decoder pairs are then specialized, with the low branch supervised by frequency and perceptual losses, and the high branch supervised by an L1 loss plus adversarial term, and a simple concatenation fuses the bands into one latent for diffusion modeling.","core_discovery":"The paper's central claim is a diagnosis plus a remedy: standard variational and vector-quantized tokenizers jointly optimize reconstruction across the whole image, and this joint objective is dominated by low-frequency energy, leaving high-frequency details under-fit. FA-VAE addresses the root cause by decomposing each image with a level-1 Haar discrete wavelet transform into low-frequency (LL) and high-frequency (LH, HL, HH) components, encoding and decoding each band with its own encoder-decoder pair, and fusing the sampled latents by concatenation before applying the inverse wavelet transform. The low-frequency branch keeps a VA-VAE-style objective with a vision foundation alignment loss","pith_inferences":["The reconstruction gain may partly come from having two encoder-decoder pairs instead of one, since the ablation only matches 'similar latent dimensionality' and the paper does not state equal parameter counts or training budgets; a capacity-matched coupling experiment would isolate the decoupling effect.","The high-frequency branch's avoidance of pretrained feature losses is itself a claim worth testing: an LPIPS or DINOv2 loss applied directly to the high-frequency band might further improve detail fidelity, or might conflict with the L1 objective.","Because the residual power spectra show lower energy across both bands, one testable consequence is that FA-VAE latents should also improve downstream tasks that depend on sharp edges, such as super-resolution finetuning or image editing with diffusion inversion.","The paper leaves open whether a learned, adaptive decomposition (rather than a fixed Haar filter) would preserve even more detail; a learned filter bank could serve as a strict generalization of this method."],"forward_implications":["If the decoupling claim holds, any latent tokenizer can adopt frequency-split encoders and decoders without changing the downstream generation backbone, yielding sharper text, edges, and textured regions.","The frequency-aware latent space should transfer to other latent diffusion and autoregressive pipelines, improving perceptual quality wherever the tokenizer is replaced.","The two-band structure suggests a natural way to trade off detail versus compression: the high-frequency latent can be down-weighted or dropped when bandwidth is limited, while preserving global structure.","The class-wise reconstruction analysis implies frequency-aware tokenization reduces representation bias toward easy, low-texture categories, which could improve generative fairness across rare or detailed classes.","The same wavelet-decoupling recipe may extend to video, 3D shapes, and medical images, where fine structures are also concentrated in high-frequency subbands."],"supporting_citations":[{"why":"Supplies the VA-VAE baseline, the LightningDiT diffusion backbone, and the vision-foundation alignment loss design that FA-VAE extends.","marker":"(Yao, Yang, and Wang 2025)"},{"why":"Provides the normalization strategy applied to the Haar wavelet low- and high-frequency components.","marker":"(Mulcahy 1997)"},{"why":"DINOv2 is the pretrained foundation model used for the low-frequency vision foundation alignment loss.","marker":"(Oquab et al. 2023)"},{"why":"Provides the LDM architecture and training scheme that FA-VAE's tokenizer backbone follows.","marker":"(Rombach et al. 2022)"},{"why":"MAR's encoder-decoder architecture and training setup are inherited for the individual frequency branches.","marker":"(Li et al. 2024a)"},{"why":"LPIPS is used as the perceptual loss and evaluation metric for reconstruction quality.","marker":"(Zhang et al. 2018)"},{"why":"VQGAN-style adversarial loss is used for low-frequency regularization.","marker":"(Esser et al. 2021)"}],"fun_headline_variants":["Wavelet split sharpens textures, cuts FID by 37%","Frequency decoupling restores sharp detail in generation","FA-VAE: wavelet-aware latents fix low-frequency bias","Split latents by wavelet to recover high-frequency detail"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The central ablation compares FA-VAE against coupled-frequency baselines that are said to use 'similar latent dimensionality,' but the paper does not confirm equal parameter counts or training budgets, so the reconstruction gains could partly come from extra encoder-decoder capacity rather than from frequency decoupling itself.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet split sharpens textures, cuts FID by 37%","Frequency decoupling restores sharp detail in generation","FA-VAE: wavelet-aware latents fix low-frequency bias","Split latents by wavelet to recover high-frequency detail"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1422,"prompt_tokens":776,"completion_tokens":646,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":578}},"tokens_in":520,"tokens_out":646,"duration_ms":6756,"temperature":1.0,"reasoning_tokens":578,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:23:11.955483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a coupled-frequency variant of the FA-VAE architecture with identical encoder-decoder parameter counts, same loss weights, and same training epochs (matching the decoupled model in everything except the separation of frequency bands), then compare reconstruction loss, LPIPS, and generation FID. If the coupled variant matches FA-VAE's numbers, the decoupling is not the cause of the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the VA-VAE baseline, the LightningDiT diffusion backbone, and the vision-foundation alignment loss design that FA-VAE extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the normalization strategy applied to the Haar wavelet low- and high-frequency components."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the LDM architecture and training scheme that FA-VAE's tokenizer backbone follows."},{"cited_title":"A.; Shechtman, E.; and Wang, O","cited_arxiv_id":null,"evidence_quote":"LPIPS is used as the perceptual loss and evaluation metric for reconstruction quality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"VQGAN-style adversarial loss is used for low-frequency regularization."}],"review_version":1}