{"id":"778e35e3-0114-43f9-8b0f-a02730ea07a2","arxiv_id":"2506.04956","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FEAT combines spatial, temporal, and channel attention in a diffusion transformer and reports improved medical video generation metrics with lower parameter counts than Endora.","lead":"FEAT is a transformer for generating medical videos that adds channel-wise attention and a residual guidance module to spatial and temporal attention, aiming for high quality at lower compute. The authors show their small model matches or beats a larger state-of-the-art system with about a quarter of the parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline performance claim rests on single-run metric comparisons without error bars or stated baseline training protocols; FEAT-L's advantage over Endora may be real, but it is not yet established as a robust result.","rationale":"I read the paper as a plausible architectural contribution: sequential spatial-temporal-channel attention with linear-complexity WKV and global channel attention is a reasonable design, and the ablation in Table 3 is consistent with each component helping. The reader's formal weakest assumption concerned whether the modified WKV token-shift preserves global receptive field and linear complexity; I did not find a concrete failure mode there, since the depthwise-convolution shift only adds local mixing before the WKV step and the per-token recurrence remains linear in sequence length. The more load-bearing weakness is the empirical comparison itself. The strongest claim is explicitly quantitative, and Section 3 gives no error bars, no multi-seed estimates, and no baseline training/evaluation protocol details. Small differences such as FVD 72.2 vs 72.3 on Kvasir-Capsule are not interpretable without variance information. FEAT-L's larger margins may survive a proper test, but they currently rest on a single table with no statistical support. I therefore agree with the reader's CONDITIONAL verdict and see no reason to strengthen or weaken it; hence UNCHANGED. The recommended concrete test is deliberately narrow: rerun the key comparisons with seeds and confidence intervals, which would either confirm the headline claim or show that the advantage is within noise.","tokens_in":8637,"tokens_out":8378,"duration_ms":106528,"concrete_test":"Re-run FEAT-S, FEAT-L, Endora, and at least one additional baseline (e.g., LVDM) on both datasets with 5 independent seeds each, using the exact same data preprocessing, sampling schedule, and FVD implementation; report mean and 95% confidence intervals for FVD, CD-FVD, FID, and IS. Concretely, verify whether FEAT-L's observed FVD advantage on Colonoscopic (351.1 vs 460.7) exceeds twice the pooled standard error; if the confidence intervals overlap, the headline superiority claim is not supported. Also report the FVD feature extractor version and the 2048-sample selection procedure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, stated in the Abstract and quantified in Table 1, is that FEAT-S with 23% of Endora's parameters achieves comparable or superior performance and FEAT-L surpasses all comparison methods. This is an empirical claim, and the load-bearing assumption is that the reported metric differences reflect genuine method quality rather than evaluation noise, implementation differences, or protocol mismatch. Section 3.1 states that FVD/CD-FVD/FID/IS are computed with 2048 samples, but no seeds, number of repetitions, confidence intervals, or statistical significance tests are reported. Section 3.2 says the baselines were 'replicated,' yet no training iterations, batch sizes, sampling hyperparameters, number of GPUs, or FVD implementation details are given. On Colonoscopic, FEAT-S beats Endora on FVD (415.4 vs 460.7) but essentially ties on FID (13.3 vs 13.4); on Kvasir-Capsule the FVD difference is 72.2 vs 72.3, which is within expected FVD noise. FEAT-L's larger FVD gains (351.1 vs 460.7 and 59.2 vs 72.3) could be real, but if Endora's numbers were generated under a different evaluation pipeline, or if a different random seed changes which 2048 real/generated samples are used, the comparison is not trustworthy. Additionally, the 'surpasses all comparison methods' framing omits cited endoscopic-video baselines such as EndoMamba [10], so the scope of the claim is narrower than stated. These issues do not invalidate the architecture, but they make the headline quantitative claim currently unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FEAT, a latent diffusion transformer for medical video generation that applies sequential spatial, temporal, and channel attention. Spatial and temporal blocks use WKV attention with modified token-shift mechanisms, the channel block uses global channel attention from Restormer, and a residual value guidance module (ResVGM) injects the input embedding as fine-grained denoising guidance. Experiments are reported on Colonoscopic and Kvasir-Capsule datasets, comparing against StyleGAN-V, LVDM, MoStGAN-V, and Endora on FVD, CD-FVD, FID, and IS, plus a semi-supervised classification downstream task and an ablation study. The headline claim is that FEAT-S, with 158.0M parameters (23% of Endora's 673.7M), achieves comparable or superior performance, while FEAT-L surpasses all evaluated baselines.","tokens_in":9006,"tokens_out":5417,"duration_ms":64371,"significance":"If the empirical results are robust, FEAT is a genuinely parameter-efficient architecture for medical video generation: the ablation in Table 3 is internally consistent and shows each proposed component improving FVD/FID, and the parameter-ratio arithmetic (158.0/673.7 ≈ 23%) is correct. The paper also provides code and uses public datasets, which aids reproducibility. However, the central empirical claim is not yet established because the comparisons rest on single-run metrics without uncertainty quantification or a fully specified evaluation protocol; several metric differences are small relative to expected FVD noise. The efficient-attention complexity statement also needs precision before the efficiency claim can be fully assessed.","major_comments":[{"comment":"The central comparison is based on single-run metrics with no error bars, confidence intervals, or significance tests. On Kvasir-Capsule, FEAT-S and Endora differ by only 0.1 in FVD (72.2 vs 72.3) and 0.64 in FID (9.97 vs 10.61); on Colonoscopic, the FID difference is 0.1 (13.3 vs 13.4). These differences are likely within metric noise, so the claim that FEAT-S is \"comparable or superior\" with 23% of Endora's parameters is not yet supported. Please report mean and standard deviation over at least three seeds, and include a significance test (e.g., a paired bootstrap or repeated-measures test) for the headline comparisons.","section":"§3.1, Table 1"},{"comment":"The paper states that baselines were \"replicated\" but provides no training details: no number of training iterations or epochs, no batch size, no learning-rate schedule, no number of GPUs, no sampling steps, and no classifier-free guidance settings. Without these, FEAT-L's larger FVD gains (e.g., 351.1 vs 460.7 on Colonoscopic) could stem from protocol mismatch rather than architectural superiority. Please report the full training and evaluation protocol for every method, including the exact FVD/CD-FVD implementation and how the 2048 evaluation samples are drawn from each dataset.","section":"§3.2"},{"comment":"The claim that FEAT-L \"surpasses all comparison methods\" is narrower than stated because EndoMamba [10], an endoscopic video foundation model, is cited in the references but never included in the comparisons. To support a state-of-the-art claim over endoscopic video generation, either add EndoMamba to Table 1 or explicitly qualify the claim as \"all evaluated baselines.\" In addition, Table 1 leaves parameter and FLOPs entries blank for StyleGAN-V, LVDM, and MoStGAN-V; these should be filled in or marked with a clear explanation.","section":"Abstract and Table 1"},{"comment":"The paper describes the attention mechanisms as achieving \"global dependencies\" with \"linear computational complexity\" in each dimension, but this is imprecise. WKV attention has linear complexity in sequence length but is a recurrent/decay-based formulation rather than full pairwise attention, so its notion of global dependency should be stated explicitly. Global channel attention from [11] has cost O(F·H·W·C²), which is quadratic in channel count C. Please provide the explicit complexity in terms of F, H, W, and C for each of the three blocks, and clarify in what sense WKV attention is global for the spatial and temporal dimensions.","section":"§2.2, Figure 2"},{"comment":"ResVGM is a central contribution and the Table 3 ablation attributes a large improvement to it (FVD 583.6 → 415.4), yet the module is underspecified. The paper says Z is obtained \"via convolution of the input\" without giving the convolution kernel size, stride, or normalization, and it does not state whether the same Z is reused in all 27 blocks or whether λ1_c, λ2_c are per-block. Please clarify the computation of Z, the sharing pattern across blocks, and the initialization of the learnable scaling parameters.","section":"§2.3, Eq. (2)"}],"minor_comments":[{"comment":"Reference [26] is titled \"HyperKvasir,\" but the text refers to the Kvasir-Capsule dataset; please verify that the citation is correct or add the proper Kvasir-Capsule reference.","section":"References"},{"comment":"There are typos in the acknowledgments: \"Proiect\" should be \"Project\" and \"Bejing\" should be \"Beijing.\"","section":"Acknowledgments"},{"comment":"The formatting of the missing FLOPs/parameter entries uses backslashes (\"\\ \\\"), which will not render correctly; please use an em dash and add a table note explaining why these values are omitted.","section":"Table 1"},{"comment":"Reference [16] lists the arXiv ID as \"2404.058923,\" which appears to have an extra digit; reference [27] lists \"2104.112225(14), 6,\" which appears malformed. Please check both.","section":"References"},{"comment":"The qualitative comparison is at low resolution and does not clearly show temporal consistency; consider adding frame-by-frame zoomed crops or a temporal-structure visualization.","section":"Figure 4"},{"comment":"The evaluation description states that FVD follows StyleGAN-V, but the CD-FVD computation is not described; please specify the official implementation used and any preprocessing differences between FVD and CD-FVD.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the architectural idea is promising, but the empirical evidence for the headline claim needs substantial strengthening before acceptance. The omission of EndoMamba from the comparison is worth raising explicitly in the decision letter, as it bears on the state-of-the-art claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's actual new contribution is the specific combination: spatial WKV, temporal WKV, and global channel attention in sequential blocks, plus a residual value guidance module that injects the input embedding into attention values. Each component is borrowed from prior work (Vision-RWKV, Restormer, value residual learning), which the authors cite, but the combination applied to medical video generation is not something I've seen. The ablation in Table 3 is consistent and each component helps, and the parameter-efficiency arithmetic is correct: FEAT-S at 158M vs Endora's 673.7M is indeed 23%. Code is promised, which is good.\n\nThe soft spot is the empirical headline. \"FEAT-L surpasses all comparison methods\" rests on single runs with no error bars, no seeds, no confidence intervals, and no stated training iterations or sampling protocol for the baselines. The FVD differences are sometimes tiny: on Kvasir-Capsule, FEAT-S's 72.2 vs Endora's 72.3 is within noise; on Colonoscopic, FEAT-S beats Endora on FVD (415 vs 460) but essentially ties on FID (13.3 vs 13.4). FEAT-L's larger gaps could be real, but without knowing how Endora was replicated, I cannot judge. Also, EndoMamba is cited but not compared; if it is a generation baseline, that is a scope gap; if it is not, the mention is confusing.\n\nThese are exactly the things a good referee can fix. They do not invalidate the architecture; they just mean the central claim is currently unverified. The downstream augmentation experiment is small (one dataset, 40 labels) but the result is suggestive.\n\nWho is this for? Researchers working on efficient video generation, especially in medical imaging, will get value from the design. It deserves a serious referee, but the authors need to provide repeated runs, protocol details, and a more careful scope. I would want to see that before citing it.","headline":"A plausible efficiency-focused architecture for medical video generation, but the headline performance claim is not yet backed by robust statistics.","tokens_in":9513,"tokens_out":2800,"would_cite":false,"duration_ms":31282,"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":"FEAT is a diffusion transformer for medical video generation that replaces quadratic self-attention with linear-time spatial, temporal, and channel attention, and it reports matching a state-of-the-art model with a small variant and…","keywords":["medical video generation","diffusion transformer","linear-complexity attention","WKV attention","global channel attention","residual value guidance","endoscopy","semi-supervised learning"],"falsifier":"Train the same FEAT-S configuration with the original Vision-RWKV token shift in place of the 2D/1D depthwise-convolution shifts, keeping data and hyperparameters fixed; if FVD does not worsen substantially from the reported 415.4 on Colonoscopic, the claimed contribution of the shift modifications is not supported. Alternatively, measure per-token cost as the frame count grows from 16 to 32 at fixed resolution: a superlinear increase would contradict the linear-complexity claim.","tokens_in":8469,"feed_emoji":"🩺","tokens_out":10035,"duration_ms":107167,"temperature":0.7,"pith_summary":"FEAT is a diffusion transformer for medical video generation that replaces quadratic self-attention with three cascaded attention modules—spatial, temporal, and channel—each computed in linear time. The paper's central claim is that modeling dependencies on all three dimensions, and not just space and time, is what produces realistic endoscopy videos, and that the linear design makes this affordable. On 16-frame benchmarks, the small FEAT-S variant reaches quality comparable to the state-of-the-art Endora while using only about 23% of its parameters, and the larger FEAT-L surpasses all compared methods on the reported metrics. If correct, medical video synthesis becomes cheaper to train and deploy, and generated videos are useful enough as unlabeled data to improve downstream semi-supervised disease diagnosis.","feed_headline":"A leaner transformer beats medical-video SOTA at 23% of its size","feed_subtitle":"FEAT-S matches Endora's quality at about a quarter of the compute, and FEAT-L surpasses it.","key_machinery":"The load-bearing mechanism is the weighted key-value (WKV) attention block, a linear-time attention variant, paired with global channel attention that produces a channel-interaction matrix over the spatial and temporal positions. To keep locality, FEAT replaces the original token shift in WKV with 2D depthwise convolution for the spatial dimension and 1D depthwise convolution for the temporal dimension. The three block types are stacked in a sequential spatial-temporal-channel arrangement, and ResVGM adds the input embedding as a residual to the value stream of every block with two learned per-channel scaling parameters. This combination is what lets the paper claim global dependencies across all dimensions at linear computational cost.","core_discovery":"The discovery is an architecture, not a single theorem. FEAT processes latent video tokens with interleaved spatial, temporal, and channel Transformer blocks: the spatial block uses weighted key-value (WKV) attention with a 2D depthwise-convolution token shift, the temporal block uses WKV attention with a 1D depthwise-convolution token shift, and the channel block uses global channel attention. A residual value guidance module (ResVGM) injects the input embedding—which encodes both video content and the current noise pattern—as a residual value into every attention block, giving fine-grained, pixel-level denoising guidance in place of coarse timestep embeddings alone. The paper reports that every added component improves FVD monotonically (990.0, 788.4, 583.6, 415.4) on Colonoscopic, and that FEAT-S matches or slightly beats Endora on the tested metrics while FEAT-L leads on all of them across both datasets.","pith_inferences":["An untested extension: because the spatial-temporal-channel decomposition is not specific to endoscopy, the same block design could presumably transfer to other medical video modalities such as ultrasound or cardiac MRI, but FEAT's experiments only cover gastrointestinal datasets.","A testable consequence: the residual value guidance mechanism is generic, so adding the input embedding as a residual value to any video-diffusion transformer should improve generation if coarse timestep guidance is the limiting factor; FEAT does not evaluate this outside its own architecture.","A caveat about the efficiency claim: the linear-complexity argument is asymptotic, and the practical speedup at the fixed 16-frame, 128x128 evaluation size would need wall-clock measurements rather than the reported parameter and FLOP counts alone."],"forward_implications":["Attention cost grows linearly with the number of video tokens, so longer or higher-resolution medical videos become more practical than with quadratic self-attention.","FEAT-S uses 158.0M parameters and 118.7 GFLOPs versus 673.7M and 465.8 GFLOPs for Endora, making the smaller variant materially cheaper while keeping comparable quality.","FEAT-generated colonoscopy videos transfer to downstream diagnosis: semi-supervised polyp F1 rises to 91.3 with FEAT-L, above Endora's 87.0 and a supervised-only 74.5 baseline.","Each proposed component contributes independently, since adding WKV attention, channel attention, and ResVGM in turn lowers FVD from 990.0 to 415.4 on Colonoscopic.","Channel interactions are no longer neglected in the transformer design, addressing a limitation shared by previous spatial-temporal video transformers."],"supporting_citations":[{"why":"Supplies the Endora state-of-the-art model that FEAT-S matches in quality and FEAT-L surpasses on the reported metrics.","marker":"[7]"},{"why":"Provides the 16-frame fixed-interval sampling protocol and the Latte spatial-temporal transformer design that FEAT builds on.","marker":"[9]"},{"why":"Supplies the global channel attention mechanism used in the FEAT channel block.","marker":"[11]"},{"why":"Introduces the original RWKV weighted key-value attention that the FEAT WKV blocks are based on.","marker":"[15]"},{"why":"Supplies the Vision-RWKV WKV attention block that FEAT adapts with 2D and 1D depthwise-convolution token shifts.","marker":"[17]"},{"why":"Motivates the latent diffusion setting in which FEAT diffuses in an encoded space rather than pixel space.","marker":"[21]"},{"why":"Supplies the value-residual attention idea that ResVGM uses to add the input embedding as a residual value.","marker":"[24]"},{"why":"Provides the Colonoscopic dataset used for training, evaluation, and the downstream PolyDiag experiment.","marker":"[25]"},{"why":"Provides the Kvasir-Capsule dataset used as the second evaluation benchmark.","marker":"[26]"},{"why":"Defines the evaluation protocol for FVD computation from 2048 video samples with complete 16-frame structure.","marker":"[31]"}],"fun_headline_variants":["FEAT: full-dimensional attention transformer for medical video at 23% size","Linear-complexity attention lets FEAT beat medical-video SOTA","Residual value guidance sharpens efficient medical video synthesis","Scalable full-dim transformer: FEAT outperforms, shrinks to 23%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument presumes that the borrowed WKV attention, after the 2D and 1D depthwise-convolution token-shift modifications, still gives a genuinely global spatial and temporal receptive field at linear cost on 16-frame latent video tokens; if these shifts shrink the effective receptive field, the efficiency-quality comparison loses its stated mechanism.","fun_headline_variants_meta":{"raw":{"variants":["FEAT: full-dimensional attention transformer for medical video at 23% size","Linear-complexity attention lets FEAT beat medical-video SOTA","Residual value guidance sharpens efficient medical video synthesis","Scalable full-dim transformer: FEAT outperforms, shrinks to 23%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0005,"raw_usage":{"total_tokens":2451,"prompt_tokens":957,"completion_tokens":1494,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1415}},"tokens_in":573,"tokens_out":1494,"duration_ms":12581,"temperature":1.0,"reasoning_tokens":1415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:29:46.663439+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same FEAT-S configuration with the original Vision-RWKV token shift in place of the 2D/1D depthwise-convolution shifts, keeping data and hyperparameters fixed; if FVD does not worsen substantially from the reported 415.4 on Colonoscopic, the claimed contribution of the shift modifications is not supported. Alternatively, measure per-token cost as the frame count grows from 16 to 32 at fixed resolution: a superlinear increase would contradict the linear-complexity claim.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the Endora state-of-the-art model that FEAT-S matches in quality and FEAT-L surpasses on the reported metrics."}],"review_version":1}