{"id":"2a46a369-c742-4f9c-b9ee-9725bc06e949","arxiv_id":"2501.00603","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DiC, a U-shaped diffusion model using only 3x3 convolutions, sparse skips, and stage-wise conditioning, beats transformer-based DiT on ImageNet FID with higher throughput.","lead":"The paper introduces DiC, a diffusion model built only from fast 3x3 convolutions instead of attention layers. It reports that DiC beats transformer-based diffusion models on ImageNet FID while running several times faster.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim depends on baseline FID/IS numbers sourced from the authors' own U-DiTs paper [38] rather than independently reproduced; if those baselines are stale or misaligned, the reported 'considerable margins' over diffusion transformers are not established.","rationale":"The reader's weakest assumption targets exactly the right spot: the external comparison in Table 6 is the load-bearing evidence for the abstract's plural claim about 'existing diffusion transformers,' and those numbers come from the authors' own prior work. My reading supports this as the single most important concern because the claimed 'considerable margins' would be unsupported if the baselines are not truly comparable. I did not find an independent technical flaw in the DiC architecture itself; the ablations are plausible and internally consistent. However, the absence of released code and the self-sourced baseline values raise the correctness risk from low to medium, which justifies keeping the verdict conditional. The concrete test I propose would settle the concern by checking whether a faithful DiT-XL/2 reproduction matches the paper's 20.05 or the DiT paper's 19.5, and whether the other transformer baselines actually take the aligned setting as favourably as DiC.","tokens_in":14827,"tokens_out":13196,"duration_ms":120523,"concrete_test":"Independently reproduce DiT-XL/2 under the exact setting described for Table 6: 400K iterations, ImageNet 256×256, batch size 256, learning rate 1e-4, seed 0, the DiT codebase, and the same sd-vae-ft-ema VAE. Compute FID-50K with the same sampler. If the reproduced FID is close to 20.05, the Table 6 baseline is fair; if it is materially lower (e.g., ≤19.5, matching the DiT paper), the reported margin is inflated. Additionally, obtain or request the U-DiTs training logs and checkpoints for all Table 6 baselines, or have an independent group rerun PixArt-α-XL/2 and DiffiT-XL/2 under their own recommended settings to confirm the comparison is not an artifact of aligned hyperparameters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result — that DiC surpasses existing diffusion transformers by considerable margins — is carried by Table 6, where all baseline FID/IS values (U-ViT-XL, DiT-XL/2, PixArt-α-XL/2, DiffiT-XL/2, DiT-LLaMA) are taken from U-DiTs [38], a paper co-authored by the present first author. Even the DiT-XL/2 400K FID of 20.05 used in Tables 4 and 6 is reported as 19.5 in the original DiT paper [30], suggesting the authors' rerun, not the official number, underpins the primary comparison. If these baselines were trained under hyperparameters, samplers, or evaluation details that differ from those used for DiC, the margins (e.g., 13.11 vs 20.05) could shrink or disappear. The manuscript provides no code, checkpoints, training logs, or seed variance, so an independent reader cannot verify whether the baseline runs are directly comparable. The internal ablations (Tables 1, 2) are consistent, but they do not validate the external comparison that the abstract's central claim relies on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DiC, a class-conditional latent diffusion model built entirely from stride-1 3x3 convolutions. The architecture is an encoder-decoder U-Net with sparse skip connections, stage-specific condition embeddings, mid-block condition injection, and conditional gating. The authors report that DiC outperforms DiT and other transformer baselines on ImageNet 256x256 and 512x512 in FID/IS at 400K iterations, while maintaining higher throughput, and they present ablations supporting each design choice.","tokens_in":15128,"tokens_out":7347,"duration_ms":60007,"significance":"If the results are reproducible, DiC would be a significant counterpoint to the field's shift toward isotropic transformers, demonstrating that a well-designed purely convolutional architecture can match or beat transformer backbones at comparable FLOPs while being faster in practice. The paper's strengths include a clean set of ablations (Tables 1 and 2) showing monotonic improvements, measured throughput numbers, and a clear progression from standard architectures to the final design. However, the external validity of the headline numbers is currently limited by the reliance on inherited baseline statistics and the absence of error bars.","major_comments":[{"comment":"The central claim that DiC surpasses existing diffusion transformers by considerable margins rests entirely on baseline FID/IS values taken from ref. [38], a paper co-authored by the first author. In particular, the DiT-XL/2 value of 20.05 used in Tables 4 and 6 differs from the 19.5 reported in the original DiT paper [30], so the comparison is not against the official numbers. The paper should either reproduce all baselines in the same codebase with the same 400K setting, or provide the exact checkpoint and sampling details for the inherited numbers. Without this, the margins such as 13.11 vs 20.05 are not yet established.","section":"Sec. 4.2, Tables 4 and 6"},{"comment":"The FLOPs comparisons are inconsistent: DiC FLOPs are reported both raw and Winograd-adjusted, while DiT FLOPs are given only in raw form. Since Winograd applies only to 3x3 convolutions, the entries '116.1 (57.2)' next to '118.6' overstate the efficiency advantage. The measured throughput (TP) is a fairer metric and should be emphasized; the FLOPs tables should either use unadjusted FLOPs for all models or adjust all applicable operations. Additionally, the text in Sec. 4.3 saying 'DiT-XL/2 requires 524.7G FLOPs (with Winograd optimization)' appears to be a typo, as Winograd does not apply to DiT.","section":"Sec. 4.1, Tables 3, 6, and 7"},{"comment":"All FID/IS values are single-run estimates without error bars or multiple seeds. Diffusion training is known to have run-to-run variance, and the close margins in Table 9 (DiC-H 2.25 vs DiT-XL/2 2.27) and the ordering of the ablations in Tables 1-2 could change with repeated runs. The authors should report at least two or three seeds for the key comparisons, or state the variance if fewer runs are available.","section":"Sec. 4.2, Tables 4-7"}],"minor_comments":[{"comment":"The table header and row alignment are confusing: for DiT-XL/2, the '16.2' appears to be throughput, not Winograd FLOPs, but it is placed in the Wino. column. Please reformat the table and correct the accompanying text.","section":"Sec. 4.3, Table 7"},{"comment":"The caption reads 'cf g= 4'; this should be 'cfg=4'.","section":"Figure 5 caption"},{"comment":"The claim that PixArt-α-XL/2 is compared under the same class-conditional setting should be clarified, since PixArt-α is originally a text-to-image model; state explicitly how it was adapted per [38].","section":"Sec. 4.2"},{"comment":"The phrase 'prevents any leakage of label between stages' is imprecise; synchronized label-drop during training ensures consistent conditioning dropout, not information leakage. Rephrase for accuracy.","section":"Sec. 3.3"},{"comment":"The acknowledgment that baseline statistics in Table 6 are from [38] is good, but it should also appear prominently in the main text near the table.","section":"Supplementary, 'Credit'"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is well-founded. The paper's headline comparative claims depend on baseline numbers from a paper co-authored by the first author, and the manuscript would be considerably strengthened by providing independent baseline reproductions or detailed training/evaluation logs. I would encourage the editor to request this before final acceptance, and to consider whether the lack of code release is acceptable for the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"DiC is a serious empirical architecture study: a pure 3x3-conv U-Net with sparse skips and stage-specific conditioning that beats DiT-XL/2 decisively at the 400K-iteration ImageNet budget and roughly matches it at longer training, while running several times faster. I think the core claim is likely true, and the paper deserves a real referee.\n\nWhat is genuinely new is the design combination and the empirical finding that, for pure 3x3 convs, hourglass beats isotropic, sparse skips beat dense skips, and stage-specific conditioning helps. The ablations in Tables 1 and 2 are clean and consistent. The authors also did something I appreciate: they tried to align RIN, EDM2, and HDiT to the DiT training setting and honestly report that those methods converge slowly or fail outright. That is useful negative evidence.\n\nThe main soft spot is baseline provenance. Table 6 inherits all baseline FID/IS numbers from U-DiTs [38], co-authored by the present first author. The paper discloses this, and the caption claims a universally aligned 400K setting, but the DiT-XL/2 number they use is 20.05, not the 19.5 in the original DiT paper. That discrepancy is not load-bearing given the huge gap (13.11 vs 20.05), but an independent reader cannot verify the inherited numbers without code, logs, or checkpoints. I would ask for release of code and checkpoints, plus a short explanation of why the rerun differs from the published DiT number.\n\nSecond, all FIDs are single-run. One seed at a few model sizes would help, especially because the paper makes strong comparative claims. Third, the abstract says DiC \"surpasses existing diffusion transformers by considerable margins.\" That is accurate for the 400K-budget comparisons in Tables 4 and 6, but Table 9 shows DiC-H at 2M iterations essentially matching DiT-XL/2 at 7M (2.25 vs 2.27). The real story is faster convergence and much higher throughput, not a quality margin at scale. The wording should be tightened.\n\nThe Winograd-aware FLOPs are a reasonable secondary metric; they are usually labeled, though Fig. 4 mixes adjusted and unadjusted points and could mislead.\n\nNet: this is a solid counterpoint to the transformer consensus, with enough evidence and honest internal comparisons to justify peer review. I would accept it with revision, mainly asking for code/checkpoints, seed info or error bars, and a more precise abstract. I would bring it to our reading group and would likely cite it in work on efficient generative backbones.","headline":"DiC is a real empirical architecture study with a likely-true core claim, but its headline margins lean on baseline numbers from the authors' own prior work and the abstract overstates the long-training comparison.","tokens_in":15638,"tokens_out":2768,"would_cite":true,"duration_ms":27974,"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":"DiC, a diffusion model made entirely of stride-1 3x3 convolutions, claims to surpass transformer-based diffusion models on ImageNet generation while running faster.","keywords":["diffusion models","3x3 convolution","U-Net","diffusion transformer","ImageNet generation","classifier-free guidance","efficient generative models"],"falsifier":"Independently re-run DiT-XL/2 and the other baselines under the same 400K-iteration ImageNet 256x256 setting with the same VAE and optimizer, and compute FID/IS on the same reference statistics; if DiT-XL/2's FID drops from 20.05 toward DiC's 13.11, or if DiC's numbers move when re-measured, the claimed margin collapses. A second check is to measure DiC and DiT throughput on the same GPU and kernel library, since DiC's reported Winograd FLOPs assume hardware and software support for that algorithm.","tokens_in":14662,"feed_emoji":"🖼️","tokens_out":6863,"duration_ms":58077,"temperature":0.7,"pith_summary":"The paper argues that the recent turn to transformer backbones in diffusion models is not necessary: a carefully designed CNN made entirely of stride-1 3x3 convolutions can beat diffusion transformers on class-conditional ImageNet generation. On the standard DiT 400K-iteration setting, the largest DiC variant reaches FID 13.11 and Inception Score 100.15 on 256x256 ImageNet, against FID 20.05 and IS 66.74 for DiT-XL/2, while processing images at roughly 313 per second versus about 67. The gains come from an hourglass encoder-decoder shape that widens the receptive field, sparse skip connections that cut redundant channel concatenation, and a set of conditioning changes matched to the U-shaped structure: stage-specific embeddings, mid-block injection, and conditional gating. If the results hold, speed-critical and high-resolution generation can be built on simple convolutions rather than on self-attention with quadratic complexity.","feed_headline":"Pure 3x3-conv diffusion outruns transformer-based rivals","feed_subtitle":"DiC hits FID 13.11 on ImageNet 256, vs DiT-XL/2's 20.05, while generating faster.","key_machinery":"The load-bearing object is the Conv3x3 Basic Block: two sequential stride-1 3x3 convolutions with GroupNorm, GELU activation, a residual shortcut, and conditioning injected into the second convolution. Around it, the hourglass encoder-decoder uses downsampling and upsampling so each 3x3 kernel sees a region equivalent to 6x6 or 12x12 pixels in the original image, compensating for the kernel's narrow receptive field. Sparse (strided) skip connections, applied every few blocks instead of every block, cut the channel-concatenation cost that dense skips impose at scale. Stage-specific embedding tables give each encoder/decoder stage its own condition vector aligned to that stage's feature dimension, and mid-block injection plus conditional gating (borrowed from DiT's AdaLN) let the condition modulate features more precisely. The paper counts Winograd acceleration as halving the effective FLOPs of the 3x3 stride-1 convolutions, which underpins the reported throughput advantage.","core_discovery":"The central claim is that a purely convolutional diffusion model can surpass diffusion transformers in both generation quality and throughput when its architecture is matched to the 3x3 kernel's receptive-field limits. DiC replaces the transformer block with two stride-1 3x3 convolutions, builds the network as an encoder-decoder hourglass, keeps only sparse skip connections between stages, and conditions each stage with its own embedding injected mid-block under an AdaLN-style gating scheme. On ImageNet 256x256 at 400K iterations with no guidance, DiC-XL reduces FID from 20.05 (DiT-XL/2) to 13.11 and raises IS from 66.74 to 100.15; with classifier-free guidance DiC-XL reaches FID 3.89 versus 6.24. Scaled further, DiC-H reaches FID 2.25 with guidance after 2M iterations, slightly better than DiT-XL/2's 2.27 after 7M iterations, at roughly 2.4 times the throughput. The paper also reports the same pattern at 512x512 resolution.","pith_inferences":["The conditioning recipe (stage-specific embeddings, mid-block injection) addresses a general mismatch in U-shaped networks where a single condition embedding table is shared across stages with different channel widths; this design lesson likely transfers to other encoder-decoder generative models, including U-shaped vision transformers and hybrid diffusion backbones.","If the reported margins hold under independent replication, the current preference for isotropic transformer backbones in diffusion models may be a convenience rather than a necessity; simplified convolutional architectures could define a better efficiency frontier for real-time and high-resolution synthesis, where attention's quadratic cost is most punishing.","The Winograd-based FLOP accounting suggests that reported GFLOPs for convolution-based models can be misleadingly high; comparing actual latency on target hardware, including which kernel libraries are used, matters more than theoretical FLOPs for deciding between CNN and transformer backbones.","A testable prediction follows from the receptive-field argument: on very high-resolution images, where a few downsampling stages no longer give the deepest 3x3 kernels a view of the whole image, DiC's advantage should shrink unless the hourglass depth is increased, which would indicate that the receptive-field expansion is indeed the mechanism."],"forward_implications":["On the standard DiT 400K schedule, DiC-S, DiC-B, and DiC-XL all beat their DiT counterparts by large margins (for example FID 13.11 versus 20.05 at XL size), so the advantage holds across model scales.","Because DiC's cost grows roughly linearly with image resolution while DiT's self-attention grows quadratically, the quality and speed gap widens at higher resolutions such as 512x512.","DiC-H converges fast: with no guidance it reaches FID 9.73 at 600K steps, close to DiT-XL/2's 9.62 after 7M steps, suggesting CNN backbones need far less compute to reach a given quality.","With classifier-free guidance and longer training, DiC-H reaches FID 2.25 on ImageNet 256x256, matching or slightly beating DiT-XL/2's 2.27 while training for 2M instead of 7M iterations and sampling at 2.4 times the throughput.","Combining DiC with representation alignment (REPA/U-REPA) yields FID 1.74 after 1M iterations, exceeding the plain-training results and showing the architecture benefits from the same convergence accelerations as transformers."],"supporting_citations":[{"why":"Defines the DiT baseline, the 400K/2M/7M training settings, and the AdaLN conditioning that DiC adopts as gating; all main comparisons are measured against it.","marker":"[30]"},{"why":"Supplies the aligned 400K-iteration FID/IS numbers for DiT-XL/2, U-ViT-XL, PixArt-alpha, DiffiT, and DiT-LLaMA in Table 6, the core of the comparison claim.","marker":"[38]"},{"why":"U-ViT is the isotropic-with-skips baseline; its skip design is contrasted with DiC's sparse skips, and its reported numbers serve as a baseline in several comparison tables.","marker":"[1]"},{"why":"ADM provides the canonical convolutional U-Net block and the mid-block condition injection position that DiC revisits.","marker":"[11]"},{"why":"The Winograd algorithm is cited to justify the claim that stride-1 3x3 convolutions save most multiplications, supporting the speed and effective-FLOPs claims.","marker":"[26]"},{"why":"ConvNeXt motivates the switch from SiLU to GELU in DiC's basic block.","marker":"[28]"},{"why":"RepVGG demonstrates that simple 3x3 convolutions can be both fast and accurate, the starting premise for DiC's speed goal.","marker":"[12]"}],"fun_headline_variants":["3x3 convs only: DiC outruns diffusion transformers in speed and FID","DiC: pure 3x3 conv diffusion beats DiT with higher speed","Hourglass 3x3 convs: DiC tops DiT without attention","Rethinking 3x3 convs: DiC hits FID 13.11 vs DiT's 20.05","3x3 conv diffusion: faster generation, better FID than DiT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison against diffusion transformers depends on the baseline FID/IS numbers in the main table being obtained under exactly the same aligned 400K-iteration DiT training setting, but those numbers are inherited from a prior study co-authored by DiC's own first author rather than re-run here.","fun_headline_variants_meta":{"raw":{"variants":["3x3 convs only: DiC outruns diffusion transformers in speed and FID","DiC: pure 3x3 conv diffusion beats DiT with higher speed","Hourglass 3x3 convs: DiC tops DiT without attention","Rethinking 3x3 convs: DiC hits FID 13.11 vs DiT's 20.05","3x3 conv diffusion: faster generation, better FID than DiT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001163,"raw_usage":{"total_tokens":4838,"prompt_tokens":993,"completion_tokens":3845,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":3727}},"tokens_in":609,"tokens_out":3845,"duration_ms":24346,"temperature":1.0,"reasoning_tokens":3727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:46:40.545201+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Independently re-run DiT-XL/2 and the other baselines under the same 400K-iteration ImageNet 256x256 setting with the same VAE and optimizer, and compute FID/IS on the same reference statistics; if DiT-XL/2's FID drops from 20.05 toward DiC's 13.11, or if DiC's numbers move when re-measured, the claimed margin collapses. A second check is to measure DiC and DiT throughput on the same GPU and kernel library, since DiC's reported Winograd FLOPs assume hardware and software support for that algorithm.","supporting_citations":[{"cited_title":"Scalable diffusion models with transformers","cited_arxiv_id":null,"evidence_quote":"Defines the DiT baseline, the 400K/2M/7M training settings, and the AdaLN conditioning that DiC adopts as gating; all main comparisons are measured against it."},{"cited_title":"U-dits: Downsample tokens in u-shaped diffusion transformers, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the aligned 400K-iteration FID/IS numbers for DiT-XL/2, U-ViT-XL, PixArt-alpha, DiffiT, and DiT-LLaMA in Table 6, the core of the comparison claim."},{"cited_title":"All are worth words: A vit backbone for diffusion models","cited_arxiv_id":null,"evidence_quote":"U-ViT is the isotropic-with-skips baseline; its skip design is contrasted with DiC's sparse skips, and its reported numbers serve as a baseline in several comparison tables."},{"cited_title":"Diffusion models beat gans on image synthesis","cited_arxiv_id":null,"evidence_quote":"ADM provides the canonical convolutional U-Net block and the mid-block condition injection position that DiC revisits."},{"cited_title":"A convnet for the 2020s","cited_arxiv_id":null,"evidence_quote":"ConvNeXt motivates the switch from SiLU to GELU in DiC's basic block."},{"cited_title":"Repvgg: Making vgg-style convnets great again","cited_arxiv_id":null,"evidence_quote":"RepVGG demonstrates that simple 3x3 convolutions can be both fast and accurate, the starting premise for DiC's speed goal."}],"review_version":1}