{"id":"15df563f-6922-4fe0-a8fc-7719a5fdf187","arxiv_id":"2506.13138","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"STAGE is a streaming generative world model that generates up to 600-frame driving videos on NuScenes with stable quality, using temporal feature transfer across denoising steps and a multi-stage training schedule.","lead":"STAGE generates driving videos frame by frame and keeps them stable for hundreds of frames by passing denoising features from earlier frames into later ones. This matters because long, consistent simulated driving footage could reduce reliance on expensive real-world data for training self-driving systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Segment-wise FVD does not measure the long-horizon stability the SOTA claim depends on; the 600-frame negligible-degradation claim is asserted, not quantified.","rationale":"The reader's weakest assumption is the one I would also flag: segment-wise FVD is not a long-horizon metric. I agree with the conditional verdict. The concern is load-bearing because the paper's headline quantitative advantage exists only in the long-horizon row of Table I, and that row uses an evaluation protocol that cannot see the failure mode the method is designed to address. The short-horizon comparison is actually unfavorable to STAGE on FVD (242.79 vs 94.84), so the SOTA claim hangs entirely on the long-horizon comparison. The ablation study provides credible internal evidence that HTFT and Stage 3 improve FID/FVD within the authors' own pipeline, so this is an evaluation gap rather than a fundamental flaw. The suggested test is feasible: the authors already have the generated videos and the FVD code; they need only to report the metric as a function of segment index and to make the comparison protocol identical across methods. If the trend is flat, the central claim would be substantially supported and the conditional could be lifted; if the trend rises, the claim of stable long-horizon generation over 600 frames would need to be weakened. No other concern (undefined Stage 4, missing code, missing error bars) is as directly tied to the strongest claim; those are important but secondary. Overall, no change to the reader's conditional verdict is needed.","tokens_in":8899,"tokens_out":7289,"duration_ms":78193,"concrete_test":"Re-run the long-horizon evaluation on the same 60 validation sequences under a single protocol: generate full 240-frame videos for STAGE and MagicDriveDiT, and compute FVD with the same VideoGPT features both on the full sequence and per 16-frame segment. Report FVD/FID by segment index (1-16, 17-32, ..., 225-240). If the last-segment FVD is substantially above the first-segment FVD, or the per-segment trend has a significant positive slope, the negligible-degradation claim fails. Also compute per-segment FVD on the 600-frame demo using only the initial frame and predicted conditions; a flat slope with confidence intervals would validate the unlimited-length claim, while a positive slope would invalidate it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that STAGE achieves SOTA long-horizon driving-video generation, halving the long-horizon FVD of MagicDriveDiT (280.34 vs 585.89) and sustaining stable 600-frame generation. The load-bearing assumption is that the reported long-horizon FVD measures quality over the full generated sequence. It does not. Section IV-B states that each long video is divided into 16-frame segments for FVD computation, so Table I's long-term FVD is an average over short clips, not a measure of cross-segment drift, boundary discontinuities, or error accumulation across autoregressive chunks. HTFT's stated purpose is exactly to suppress error accumulation, yet no time-indexed FVD/FID is reported: there is no per-segment trend showing that segment 15 looks as good as segment 1. Section IV-E's 600-frame demo is supported only by selected frames and the assertion of negligible degradation. The comparison is also protocol-mixed: MagicDriveDiT's 585.89 is taken from the original paper, with no evidence the same segment length, FVD extractor, conditioning schedule, or frame count was used; STAGE's own short-horizon FVD (242.79) is worse than MagicDriveDiT's short-horizon FVD (94.84), so the entire SOTA claim rests on this incomparable long-horizon row. This is not a claim that the method is wrong; it is a claim that the headline result is not yet measured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STAGE, a streaming autoregressive generative world model for driving scenes. It generates frames one at a time, conditioned on an anchor frame, HD map, and bounding boxes, and uses a Hierarchical Temporal Feature Transfer (HTFT) mechanism that fuses U-Net denoising features from previous frames through a fixed-length streaming buffer. Training is split into stages, with the later stage simulating autoregressive inference by feeding generated condition frames, and a MILE-based module enables continuing beyond the annotated NuScenes sequences. On nuScenes the model reports FID 11.04 and FVD 242.79 for short-horizon generation and FID 23.70 and FVD 280.34 for long-horizon generation, plus a 600-frame qualitative demonstration. The central claim is state-of-the-art long-horizon generation with roughly half the long-horizon FVD of MagicDriveDiT.","tokens_in":9218,"tokens_out":6385,"duration_ms":60200,"significance":"If the long-horizon evaluation were trustworthy, this would be a notable step: frame-by-frame streaming offers flexible and potentially unbounded length, HTFT is a reasonable mechanism for cross-frame feature propagation, and the ablation in Table II attributes a substantial FVD drop (508.29 to 245.11) to HTFT, which is concrete evidence that the core design choice helps. The 600-frame qualitative result, if backed by quantitative per-horizon metrics, would strengthen the claim of practical long-horizon simulation. However, as submitted, the headline SOTA claim is not yet measured: the long-horizon metric is computed on 16-frame segments, so it cannot detect the cross-chunk drift the method is designed to prevent, and the comparison numbers for competing methods are taken from different papers without a demonstrated common protocol. The contribution is plausible and the components are well motivated, but the evaluation needs substantial rework before the claims can be accepted.","major_comments":[{"comment":"The long-horizon evaluation does not measure long-horizon consistency. Section IV-B states 'Each video is divided into 16-frame segments for FVD computation,' so the long-term FVD of 280.34 in Table I is an average over short clips and cannot detect drift between autoregressive chunks, boundary discontinuities, or error accumulation across the 600-frame sequence. The 'negligible degradation' claim in Section IV-E is therefore asserted, not measured. Please report per-segment FVD/FID as a function of time (e.g., segments 1, 2, ..., 37 for a 600-frame video), compare the first and last segments, and include metrics computed on windows that straddle chunk boundaries.","section":"IV-B"},{"comment":"The SOTA comparison mixes evaluation protocols. Numbers for competing methods are taken from their original papers, with no evidence that segment length, FVD feature extractor, conditioning schedule, frame count, or evaluation subset match those used for STAGE. This is load-bearing because STAGE's short-horizon FVD (242.79) is worse than MagicDriveDiT's (94.84); the only number supporting the SOTA claim is the long-horizon row, which is exactly the row that is protocol-mixed. Please rerun MagicDriveDiT and Vista under the same long-horizon protocol, or at minimum provide a side-by-side listing of protocol details and justify why the reported numbers are commensurable.","section":"IV-B/Table I"},{"comment":"The claim of 'only minimal performance degradation' is not supported by the reported numbers. The short-horizon and long-horizon entries in Table I come from different evaluation sets and different conditioning regimes (entire validation set with a condition frame every 16 frames versus 60 sampled sequences), so the difference between 242.79 and 280.34 conflates protocol differences with temporal degradation. Report degradation within a single protocol, for example FVD on the same 60 sequences at horizons of 16, 64, 240, and 600 frames.","section":"IV-C"},{"comment":"No uncertainty estimates are reported for any metric. FVD and FID are known to be noisy, especially with only 60 long-horizon sequences, and the incremental Stage-3 gain in Table II (FVD 245.11 to 242.79) may be within run-to-run variability. Please provide multiple seeds or bootstrap confidence intervals for Tables I and II, and state the number of generated sequences and random seeds used for each entry.","section":"IV-D/Table II"}],"minor_comments":[{"comment":"The training strategy is described as three stages, but Section IV-A reports time spent on 'Stage 4' and Section IV-E refers to a 'stage 4 model.' Please define Stage 4 explicitly and state which stage(s) produced the quantitative results in Table I.","section":"III-C/IV-A"},{"comment":"The notation f_t^T is used in the HTFT equations but is never defined; define it as the U-Net feature map at denoising step t of frame T, and reconcile it with the x_T^t notation in the preceding text.","section":"III-B"},{"comment":"The text says 'It selects K frames from the previous N frames' but K is not defined; given S = {-1,-5,-10}, K = 3, so please state this explicitly and clarify how the frame-skipping strategy relates to the selection set.","section":"III-B"},{"comment":"The set S = {-1,-5,-10} is described as selecting the '1st, 5th and 10th previous frames,' which is confusing with negative indices; use positive indices such as {1,5,10} or explicitly explain the sign convention.","section":"III-B"},{"comment":"The phrase 'we generate four sets of outputs, combine them, and then compute both FID and FVD' is ambiguous; state whether FVD is computed over 64 frames per sequence or by pooling features, and report the total number of frames used for each metric.","section":"IV-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is suitable in scope for a computer vision/generative modeling journal. My recommendation of major_revision is driven by evaluation methodology, not by doubts about the method's plausibility. The stress-test concern in the review note is accurate: the segment-wise FVD does not measure the long-horizon stability that the SOTA claim rests on. I would ask the editor to require matched-protocol comparisons and horizon-indexed metrics before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that the core mechanism is interesting and the training schedule is thoughtful, but the headline claim about long-horizon superiority is not actually supported by the numbers as reported. The paper deserves a serious referee, but it needs an honest revision of the evaluation section.\n\nWhat's new: HTFT transfers U-Net features from selected previous frames at the same denoising step, which is a clean way to keep streaming generation coherent without full temporal attention. The three-stage training, where the model is decoupled from HTFT first and then trained on its own predictions, is a sensible response to the train/inference gap. The internal ablation shows the steps help: FVD drops from 508 to 245 when HTFT is added, and further to 243 with the inference-simulation stage. That is real evidence that the design choices do something within the authors' pipeline.\n\nThe soft spots are mostly in the evaluation. The long-horizon FVD is computed on 16-frame segments, so it cannot detect drift across segments, boundary discontinuities, or error accumulation over the 600-frame generation. The claim of \"negligible degradation\" is asserted, not quantified; there is no time-indexed FVD showing segment 15 is as good as segment 1. The comparison with MagicDriveDiT uses a number from the original paper, with no guarantee that the segment length, FVD extractor, conditioning, or number of frames match. Also, in short-horizon FVD, STAGE is worse than MagicDriveDiT (242.79 vs 94.84), so the SOTA claim rests entirely on the incomparable long-horizon row. There are also smaller issues: the paper mentions a \"Stage 4\" in the implementation details but only describes three stages in the method section, and no public code or checkpoints are released. Error bars are missing throughout.\n\nNone of these are fatal to the method. The streaming feature transfer and staged training are plausible and the ablations support them. But the paper currently overstates what is measured. The remedy is straightforward: report per-segment FVD/FID over time, use a unified evaluation protocol, and release code or at least detailed hyperparameters and random seeds.\n\nThis paper is for people working on driving world models and long-video generation with diffusion. A skeptical reader should not accept the SOTA claim as-is, but the method is worth a serious look and the authors' thinking is clear. I'd send it to peer review with a strong request to fix the evaluation.","headline":"Promising streaming driving-world-model architecture, but the headline long-horizon SOTA claim rests on metrics that don't measure what they claim.","tokens_in":9754,"tokens_out":1818,"would_cite":false,"duration_ms":19253,"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":"STAGE generates stable 600-frame driving videos frame by frame, halving long-horizon FVD versus the prior one-shot method.","keywords":["driving world model","long-horizon video generation","streaming autoregressive generation","latent diffusion","temporal feature transfer","autonomous driving simulation","NuScenes","video generation evaluation"],"falsifier":"Compute full-sequence FVD or a frame-to-frame consistency measure such as optical flow at chunk boundaries on STAGE's 600-frame outputs, and check whether the quality gap versus the one-shot baseline persists; alternatively, plot FVD over sliding 16-frame windows as a function of frame index, because late-window degradation would show error accumulation even if the averaged segment-wise FVD stays low.","tokens_in":8693,"feed_emoji":"🚗","tokens_out":7188,"duration_ms":72611,"temperature":0.7,"pith_summary":"The paper presents STAGE, a generative world model that produces driving-scene video frame by frame rather than in one shot. Its goal is to show that a streaming design can hold temporal consistency over hundreds of frames, where prior autoregressive methods accumulate errors and one-shot methods lack flexibility. The reported results on the NuScenes driving dataset show a long-horizon FVD of 280.34, roughly half of the 585.89 of the strongest one-shot baseline, and the model sustains a 600-frame sequence. If right, this makes synthetic driving video available for training and testing autonomous-driving agents at lengths that exceed real annotated clips.","feed_headline":"Streaming diffusion model generates 600-frame driving videos","feed_subtitle":"Frame-by-frame world model halves long-horizon FVD versus one-shot baselines.","key_machinery":"The load-bearing component is Hierarchical Temporal Feature Transfer (HTFT), which separates the temporal axis from the denoising axis and, at denoising step t of frame T, pulls feature maps of previous frames from a StreamingBuffer into the current frame's U-Net via a selector, linear projection, cross-attention, residual connection, and dropout. The buffer keeps features from the 10 most recent frames, and only frames -1, -5, and -10 are used, balancing temporal context against compute. The second component is the three-stage training schedule, whose final stage feeds the model its own generated frames as conditions, while a data-augmentation scheme removes high-frequency content from the conditioning frame with a DCT filter and adds noise or dropout. Together these suppress error accumulation, the usual failure mode of streaming generation.","core_discovery":"STAGE's central assertion is that long-horizon driving video can be generated stably by treating generation as a stream: at each new frame, a latent diffusion U-Net denoises the frame while cross-attending to selected feature maps of earlier frames from a fixed-length FIFO buffer, using frames at offsets -1, -5, and -10 per denoising step. Training proceeds in three phases: first, learn the streaming conditional base model without feature transfer; second, add the temporal feature transfer; third, replace the conditioning frame with the model's own output to simulate inference-time quality loss. The paper reports that this combination yields the lowest long-horizon FID and FVD among compared methods (FID 23.70, FVD 280.34 versus 585.89 for the one-shot baseline), with only mild degradation from short-horizon performance, and enables 600-frame generation beyond the annotated length of the training data.","pith_inferences":["Beyond the paper's reported numbers, a stronger test of the streaming claim would compute FVD in sliding windows along the 600-frame stream and check whether later windows drift; the current segment-wise average could hide progressive degradation.","The frame-by-frame design suggests an extension the authors do not run: measuring FVD as a function of generation length to test whether error accumulation is genuinely bounded rather than merely slow.","If the predicted-condition loop is accurate enough, STAGE could support closed-loop synthetic-data generation where a controller's decisions alter the next conditions; the paper demonstrates the loop but stops at visual quality rather than downstream-task utility.","The frame-skipping selection {-1, -5, -10} is a design choice; ablating buffer size and selected offsets would separate how much stability comes from HTFT versus the three-stage training."],"forward_implications":["If STAGE holds up, synthetic driving clips of 600 frames—beyond the duration of its training clips—become practical for downstream simulation.","The reported gap between short- and long-horizon FVD is small (242.79 vs 280.34), implying the streaming scheme keeps degradation roughly flat as length grows.","Because generation conditions on an anchor frame, HD map, and bounding boxes, the same model can be reused for layout-controlled simulation rather than only free-form video generation.","A planning module that predicts future ego states and BEV labels lets the model continue past the end of annotated data, opening a route to simulation in areas without map or box annotations.","Small distant foreground objects receive higher loss weight through convex-hull area weighting, which should improve the fidelity of vehicles and pedestrians in long streams."],"supporting_citations":[{"why":"Supplies the 585.89 long-horizon FVD one-shot baseline that STAGE claims to halve.","marker":"[3]"},{"why":"Autoregressive baseline whose official checkpoint provides the long-horizon FID/FVD comparison.","marker":"[5]"},{"why":"Provides the 12 Hz annotation protocol and the short-horizon baseline numbers STAGE compares against.","marker":"[14]"},{"why":"Pre-trained latent diffusion backbone whose U-Net features HTFT reuses and whose weights initialize STAGE.","marker":"[23]"},{"why":"Supplies the ego-state and BEV-label prediction used to synthesize future conditions for generation beyond annotated data.","marker":"[25]"},{"why":"Defines the Fréchet Video Distance used for all video-quality comparisons.","marker":"[27]"},{"why":"Defines the Fréchet Inception Distance used for image-quality comparisons.","marker":"[28]"},{"why":"Provides the VideoGPT feature extractor used to compute FVD.","marker":"[29]"}],"fun_headline_variants":["Streaming diffusion model hits 600-frame driving videos","World model generates 600-frame driving scenes in a stream","STAGE: 600-frame driving video simulation via streaming","Long-horizon driving video: streaming world model reaches 600 frames"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claims rely on the assumption that measuring FVD on independent 16-frame chunks of a 600-frame generation reveals whether the video stays consistent across the whole stream; if drift happens between chunks, the reported scores would miss it.","fun_headline_variants_meta":{"raw":{"variants":["Streaming diffusion model hits 600-frame driving videos","World model generates 600-frame driving scenes in a stream","STAGE: 600-frame driving video simulation via streaming","Long-horizon driving video: streaming world model reaches 600 frames"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1477,"prompt_tokens":970,"completion_tokens":507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":438}},"tokens_in":586,"tokens_out":507,"duration_ms":5307,"temperature":1.0,"reasoning_tokens":438,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:35:14.412677+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute full-sequence FVD or a frame-to-frame consistency measure such as optical flow at chunk boundaries on STAGE's 600-frame outputs, and check whether the quality gap versus the one-shot baseline persists; alternatively, plot FVD over sliding 16-frame windows as a function of frame index, because late-window degradation would show error accumulation even if the averaged segment-wise FVD stays low.","supporting_citations":[{"cited_title":"Vista: A generalizable driving world model with high fidelity and versatile controllability,","cited_arxiv_id":null,"evidence_quote":"Autoregressive baseline whose official checkpoint provides the long-horizon FID/FVD comparison."},{"cited_title":"Drivedreamer: Towards real-world-drive world models for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Provides the 12 Hz annotation protocol and the short-horizon baseline numbers STAGE compares against."},{"cited_title":"High-resolution image synthe- sis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Pre-trained latent diffusion backbone whose U-Net features HTFT reuses and whose weights initialize STAGE."},{"cited_title":"Model-based imitation learning for urban driving,","cited_arxiv_id":null,"evidence_quote":"Supplies the ego-state and BEV-label prediction used to synthesize future conditions for generation beyond annotated data."},{"cited_title":"Towards accurate generative models of video: A new metric & challenges,","cited_arxiv_id":null,"evidence_quote":"Defines the Fréchet Video Distance used for all video-quality comparisons."},{"cited_title":"Gans trained by a two time- scale update rule converge to a local nash equilibrium,","cited_arxiv_id":null,"evidence_quote":"Defines the Fréchet Inception Distance used for image-quality comparisons."}],"review_version":1}