{"id":"97259e83-bff5-4950-9f38-74820a16dc07","arxiv_id":"2412.09856","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LinGen replaces self-attention in diffusion transformers with a linear-complexity MATE block, enabling 512p 68-second video generation on a single H100 with quality comparable to Gen-3, LumaLabs, and Kling.","lead":"A new video-generation model swaps quadratic self-attention for a linear-complexity block, letting a single GPU produce 68-second clips that human evaluators rate about as often as Gen-3, LumaLabs, and Kling. The win over its own DiT baseline is large (75.6%), but that baseline may be handicapped by different model dimensions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The only controlled DiT-vs-MATE quality test is at 40K steps/256p, where faster adaptation inflates LinGen; final-convergence quality parity is unverified.","rationale":"The reader's weakest assumption identifies the DiT-4B comparison as a non-valid control for the MATE effect, citing both the width/head mismatch and the early 256p, 40K-step checkpoint. I agree with that diagnosis. The early checkpoint is the more load-bearing issue because the authors themselves explain the win rate as faster adaptation to longer sequences; no controlled comparison at convergence is reported. Without such a comparison, the claim that replacing self-attention with MATE does not compromise quality is not established. The efficiency scaling curves and complexity analysis are sound, and the ablations support the usefulness of MATE's components internally, but the headline quality claim depends on one early-stage human evaluation. A late-checkpoint A/B test, ideally with a matched-width DiT, would settle the concern. Other potential issues, such as VBench scores that are numerically lower than Gen-3/Kling and the absence of released code, are secondary and do not change the verdict. Since the reader already assigned CONDITIONAL and my concern reinforces rather than redirects that verdict, I recommend UNCHANGED.","tokens_in":21742,"tokens_out":5382,"duration_ms":55304,"concrete_test":"Run DiT-4B and LinGen-4B to a late 256p checkpoint (e.g., the full 125K-step text-to-video schedule) and perform the same A/B human evaluation with matched prompts and sample counts; report win rates with confidence intervals. If feasible, add a matched-width DiT baseline (hidden dim 2560, 20 heads) to isolate the MATE effect. If DiT-4B's late-checkpoint win rate approaches 50% or higher, the claim that MATE achieves final-quality parity is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MATE preserves generation quality relative to self-attention rests on the Fig. 8 human evaluation, which compares LinGen-4B and DiT-4B at 256p after only 40K steps (Sec. 4.3). Sec. 4.4 explicitly attributes LinGen's advantage to faster adaptation to longer token sequences, and reports win rates at even earlier 1K-step checkpoints. Thus the 75.6% win rate is a training-dynamics measurement, not a final-quality measurement. No human evaluation at a late checkpoint (e.g., after 125K steps at 256p or after the full 512p recipe) is provided for the DiT-vs-LinGen pair. The possibility that DiT catches up or surpasses at convergence is therefore untested. The comparison is further confounded by model-size mismatch: DiT-4B uses hidden dim 3072 and 24 heads, while LinGen-4B uses 2560 and 20 heads (Sec. 4.1), so the FLOPs/latency savings (15x/11.5x) are not purely architectural. The abstract's 'without compromising quality' is supported only by near-50% win rates against commercial models, which are not controlled for training data or compute. The load-bearing requirement—MATE alone matches self-attention at convergence—is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LinGen, a text-to-video diffusion framework that replaces the quadratic-complexity self-attention layers of a DiT with a linear-complexity MATE block. MATE consists of an MA-branch (bidirectional Mamba2 with a proposed Rotary-Major Scan and review tokens) and a TE-branch (a temporal Swin attention block, TESA). The authors claim that this architecture achieves up to 15x FLOPs and 11.5x latency reductions over a DiT baseline at 512p/68s while preserving generation quality, and that LinGen-4B is comparable to commercial models such as Gen-3, LumaLabs, and Kling in human evaluation. The paper reports FLOPs formulas (Eqs. 1-3), latency measurements on an H100, VBench scores, ablation studies, and human-evaluation win rates, with the central claimed contribution being 'high-resolution minute-length video generation on a single GPU without compromising quality.'","tokens_in":21860,"tokens_out":3836,"duration_ms":41410,"significance":"If the quality-parity claim were fully supported, this would be a substantive efficiency contribution to video diffusion: the MATE block is a concrete, linear-complexity replacement for self-attention, and the paper's efficiency evidence is comparatively strong. The FLOPs expressions in Eqs. (1)-(3) are explicit and plausible, the H100 latency measurements at 512p/17s/34s/68s are consistent with the claimed complexity scaling, and the ablations (Table 2, Figs. 11-12) provide useful attribution for TESA, RMS, review tokens, and hybrid training. The weakness is in the quality evidence for the central architectural claim: the only direct LinGen-4B vs DiT-4B quality comparison is at an early training checkpoint with a width-mismatched baseline, and the comparison against commercial models is reported without a full human-evaluation protocol. These issues are addressable in revision, so the contribution is defensible but not yet fully established.","major_comments":[{"comment":"The controlled MATE-vs-self-attention comparison (LinGen-4B vs DiT-4B) is performed at 256p after only 40K training steps, and Sec. 4.4 explicitly explains LinGen's advantage by its faster adaptation to longer token sequences, with human evaluations taken at 1K-step checkpoints. This is a training-dynamics measurement rather than a final-quality measurement. The paper never reports a human evaluation of this pair after convergence (e.g., after the 125K-step 256p stage or after the full 512p recipe), so the possibility that DiT catches up or surpasses at convergence is untested. Because the abstract's 'without compromising quality' claim rests on MATE matching self-attention at final quality, please add a late-checkpoint human evaluation with matched training data and compute.","section":"Sec. 4.3, Fig. 8"},{"comment":"The DiT-4B baseline is not matched in model capacity: DiT-4B uses hidden dimension 3072 with 24 heads, while LinGen-4B uses hidden dimension 2560 with 20 heads. Consequently, the headline efficiency numbers (15x FLOPs, 11.5x latency in Fig. 6) and the Fig. 8 win rates do not isolate the architectural effect of the MATE block; they also include a width and head-count difference. Please report a matched-width DiT baseline, or otherwise control for parameter count and width, in both the efficiency and quality comparisons.","section":"Sec. 4.1 (Models)"},{"comment":"The human-evaluation protocol is under-specified: the paper reports win rates and states that 'the variance of human evaluation is 3%,' but it does not state the number of evaluators, number of prompts, number of videos per model, tie handling, or confidence intervals. Without this information, near-50% win rates against Gen-3, LumaLabs, and Kling do not by themselves establish 'comparable quality.' Please provide the full protocol and the relevant statistics, or soften the claim accordingly.","section":"Sec. 4.3 (Human Evaluation Results)"}],"minor_comments":[{"comment":"The baseline name 'Runaway Gen3' should be 'Runway Gen3'.","section":"Sec. 4.1"},{"comment":"The word 'incidates' should be 'indicates'; also 'commerical' should be 'commercial.'","section":"Sec. 4.3"},{"comment":"The caption contains 'adapation' and 'faster adapation'; these should be corrected.","section":"Fig. 8 caption / Sec. 4.4"},{"comment":"The caption refers to 'the pre-training recipe of LVGen'; this should be 'LinGen.'","section":"Supp. Table 7"},{"comment":"The axis labels such as '256!,17s' appear garbled in the rendered version and should be formatted as '256p, 17s' (similarly for other resolution/length labels).","section":"Fig. 6"},{"comment":"The products (T/Tw)(H/Sw)(W/Sw) implicitly assume exact divisibility; please clarify that these are integer ceiling/floor counts of windows.","section":"Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nLinGen is a credible linear-complexity replacement for self-attention in video DiTs. The MATE block — bidirectional Mamba2, rotary-major scan, review tokens, and temporal Swin attention — is a sensible integration of known pieces, and the scaling analysis is sound: the FLOPs formulas are standard, and the measured latency on H100 shows the cost growing slowly with sequence length. The ablations support each component, and the 68-second single-GPU generation is a genuine engineering data point. The authors also deserve credit for being transparent in Sec. 4.4 that LinGen's advantage over DiT comes from faster adaptation, not necessarily higher final quality.\n\nThe soft spot is exactly that: the only controlled DiT-vs-LinGen quality comparison is at 256p, 40K steps, where faster adaptation inflates the win rate. There is no human evaluation at a later, converged checkpoint or at 512p to show MATE matches self-attention when training is complete. The comparison is also confounded by model size: DiT-4B uses hidden dim 3072 and 24 heads, LinGen-4B uses 2560 and 20 heads. So the headline 15x/11.5x savings are not purely architectural, though the linear-vs-quadratic scaling is real and would survive a matched-dimension comparison. The win rates against Gen-3, Luma, and Kling are all within the stated 3% human-eval variance, so 'comparable quality' is a fair but weak statement. The 'for the first time' claim is also a bit strong given Loong and PA-VDM were cited earlier in the same paper.\n\nNone of this makes the paper worthless. It is a serious engineering contribution with a plausible architecture, and the authors are honest about the adaptation effect. What is missing is a converged quality check. That is fixable: a matched-width DiT baseline trained to the same number of steps at 512p, or at least a late-checkpoint human eval.\n\nWho benefits: people designing efficient video backbones, and anyone tracking whether SSMs can replace attention in diffusion models. It deserves a serious referee; the requested revision is exactly the missing converged comparison. I'd accept it with that condition.","headline":"A serious linear-complexity video backbone whose efficiency claims are solid, but whose load-bearing quality parity with DiT is only shown at an early training checkpoint and with mismatched model widths.","tokens_in":22616,"tokens_out":2894,"would_cite":true,"duration_ms":29713,"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":"By replacing self-attention with a linear-complexity MATE block, LinGen makes high-resolution minute-length text-to-video generation practical on a single GPU, with quality comparable to leading commercial models.","keywords":["text-to-video generation","linear computational complexity","Diffusion Transformers","Mamba2","state space models","temporal attention","minute-length video","MATE block"],"falsifier":"Rerun the human A/B evaluation at matched model capacity (same hidden dimension and head count for DiT) and after a longer training schedule; if DiT-4B matches or beats LinGen-4B, the win-rate claim becomes a training-efficiency artifact. For the complexity claim, measure FLOPs and latency while doubling both resolution and duration beyond 68 seconds; if growth is superlinear in the number of pixels, the linear-complexity claim fails in the regime it targets.","tokens_in":21380,"feed_emoji":"🎬","tokens_out":6084,"duration_ms":55092,"temperature":0.7,"pith_summary":"The paper claims that the quadratic cost of self-attention in Diffusion Transformers can be replaced by a linear-complexity block without hurting video quality, and that this makes high-resolution minute-length text-to-video generation possible on a single GPU. It proposes LinGen, which swaps every self-attention layer for a MATE block built from a bidirectional Mamba2 branch and a temporal Swin attention branch. The significance for a general reader is concrete: if the claim holds, long-form video generation stops being a compute privilege of large clusters and becomes a problem whose cost grows linearly with the number of pixels.","feed_headline":"Linear attention swap makes minute-long AI video feasible on one GPU","feed_subtitle":"LinGen cuts FLOPs up to 15x by replacing self-attention, rivaling Gen-3 and Kling in human evals.","key_machinery":"MATE is the load-bearing object: a linear-complexity replacement for self-attention with two branches. The MA-branch stacks a bidirectional Mamba2 block with Rotary-Major Scan (four alternating 3D-to-1D token rearrangements across layers) and review tokens (average-pooled sequence summaries prepended to Mamba2's hidden state). The TE-branch is TEmporal Swin Attention, a shifted-window attention block that gathers short-range spatial and medium-range temporal correlations. The block's job is to keep the generated video consistent across frames while avoiding the adjacency loss and long-range decay that native Mamba introduces.","core_discovery":"LinGen-4B, a 4-billion-parameter model, generates 512p videos up to 68 seconds at 16 fps on a single H100 GPU. The central experimental claim is that replacing DiT self-attention with MATE raises the human-evaluation win rate to 75.6% over a DiT-4B baseline while cutting FLOPs up to 15x and latency up to 11.5x. Against commercial systems, LinGen-4B earns 50.5%, 52.1%, and 49.1% win rates against Gen-3, LumaLabs, and Kling, which the paper reads as comparable quality. The mechanism is that MATE covers short-, medium-, and long-range correlations at linear cost: Mamba2 handles global structure, Rotary-Major Scan preserves spatial and temporal adjacency, review tokens offset long-range decay, and TESA handles adjacent temporal windows.","pith_inferences":["If the linear scaling holds beyond 68 seconds, hour-length generation on a single GPU becomes a plausible engineering target, though the paper only demonstrates up to 68 seconds and does not test hour-scale consistency.","The RMS and TESA patterns are general techniques for adapting causal sequence models to 2D and 3D grids; they could transfer to other SSM-based image, audio, or multimodal generative models that face the same adjacency problem.","A matched-capacity comparison (same hidden dimension and head count for the DiT baseline) would determine how much of the reported quality advantage comes from MATE itself versus from the baseline's training-efficiency profile.","Testing MATE as a drop-in for self-attention in an already-trained short-video model, without re-training from scratch, would reveal whether the block can be adopted incrementally."],"forward_implications":["Generating 512p, 68-second video on a single H100 becomes routine; the 15x FLOPs and 11.5x latency savings grow with video length and resolution.","MATE can be dropped into existing DiT-based video models as a self-attention substitute, transferring the linear-scaling benefit to models already trained in the quadratic regime.","The 75.6% win rate over DiT-4B at the same data budget implies the architecture learns text-to-video generation from longer token sequences faster, not just cheaper.","LinGen's win rates near 50% against Gen-3, LumaLabs, and Kling suggest that minute-length generation no longer trades away the quality bar set by short commercial clips.","Because MATE is orthogonal to sampling distillation, combining it with step-reduction methods could produce a further multiplicative speed-up."],"supporting_citations":[{"why":"Supplies the bidirectional Mamba2 block and the linear-complexity formula used in the MA-branch.","marker":"[6]"},{"why":"Defines Mamba, the SSM whose adjacency preservation issue motivates RMS and TESA.","marker":"[12]"},{"why":"Introduces Zigzag scan, the prior scan alternative that RMS matches in quality while reducing latency.","marker":"[19]"},{"why":"Provides Flow Matching, the training objective used for denoising in LinGen.","marker":"[30]"},{"why":"Introduces shifted-window attention, which inspires the TESA temporal window design.","marker":"[31]"},{"why":"Defines the Diffusion Transformer backbone whose self-attention layers MATE replaces.","marker":"[39]"},{"why":"Contributes the Temporal AutoEncoder design, the linear-quadratic t-schedule, and the observed correlation between loss and human-rated quality.","marker":"[41]"},{"why":"Demonstrates the large-scale DiT approach that motivates seeking a linear-complexity alternative.","marker":"[1]"}],"fun_headline_variants":["68-second video on one GPU: linear attention beats DiT 3-to-1","Linear-cost video gen: 15x fewer FLOPs, 68s clips on one H100","LinGen: minute-long video from a single GPU with linear attention","One GPU, minute-long video: linear attention rivals Gen-3"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison that isolates MATE's effect treats DiT-4B (hidden dimension 3072, 24 heads) as a fair self-attention baseline for LinGen-4B (hidden dimension 2560, 20 heads), with the headline quality win measured early in training at 256p and 40K steps.","fun_headline_variants_meta":{"raw":{"variants":["68-second video on one GPU: linear attention beats DiT 3-to-1","Linear-cost video gen: 15x fewer FLOPs, 68s clips on one H100","LinGen: minute-long video from a single GPU with linear attention","One GPU, minute-long video: linear attention rivals Gen-3"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000423,"raw_usage":{"total_tokens":2251,"prompt_tokens":1107,"completion_tokens":1144,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":723,"completion_tokens_details":{"reasoning_tokens":1057}},"tokens_in":723,"tokens_out":1144,"duration_ms":11763,"temperature":1.0,"reasoning_tokens":1057,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:38:38.877821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the human A/B evaluation at matched model capacity (same hidden dimension and head count for DiT) and after a longer training schedule; if DiT-4B matches or beats LinGen-4B, the win-rate claim becomes a training-efficiency artifact. For the complexity claim, measure FLOPs and latency while doubling both resolution and duration beyond 68 seconds; if growth is superlinear in the number of pixels, the linear-complexity claim fails in the regime it targets.","supporting_citations":[{"cited_title":"Swin Transformer: Hierarchical vision transformer using shifted windows","cited_arxiv_id":null,"evidence_quote":"Introduces shifted-window attention, which inspires the TESA temporal window design."},{"cited_title":"Video generation models as world simu- lators","cited_arxiv_id":null,"evidence_quote":"Demonstrates the large-scale DiT approach that motivates seeking a linear-complexity alternative."}],"review_version":1}