{"id":"cf827c34-f191-4cc8-bbad-de689625a080","arxiv_id":"2607.26037","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Wonder generates minute-scale, real-time camera-controllable video worlds from a single image or video at 16 FPS, using a rendered coordinate-field control signal, sparse full-fidelity memory, and stage-specialized distillation students.","lead":"Wonder is a video world model that turns one image or a short video into a playable scene: a user moves the camera in real time and the model keeps generating, remembering places it saw before. The paper reports minute-scale rollouts at 16 FPS with better camera following than recent streaming world models, but ships no code, weights, or measured latency.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Internal contradiction: §4.2.1 promises full-fidelity KV retention without sliding-window eviction, but §4.3's rolling/sliding-window caching scheme bounds the cache to sink+top-k+recent chunks, so minute-scale revisit coherence is unproven.","rationale":"The reader's CONDITIONAL verdict is reasonable, and the DA3 loop is a valid concern. However, I see a more load-bearing, internal issue. The entire 'long-horizon memory' contribution is defined by the contrast between full-fidelity retention and sliding-window discarding. Section 4.2.1 says the cache cannot be discarded with a sliding-window strategy and that the entire KV cache is kept in full fidelity. Section 4.3 then says the opposite: to go beyond the training horizon it adopts a rolling/sliding-window caching scheme and describes the KV cache as a concatenation of sink, top-k, and recent chunks. If the latter is the actual deployed cache, then memory is not full-fidelity and 'coherent revisits' only hold within a bounded window; minute-scale rollouts would lose any region not selected into top-k before eviction. This is not a question of consensus; it is an internal inconsistency in the central mechanism. The fix is straightforward: clarify whether non-selected history is stored and retrieved later, and report measured cache size and memory vs. rollout length. If the cache is truly bounded, the paper's strongest memory claim is not supported. Because this is testable and the reader already conditioned acceptance on releasing code/weights and benchmarks, I keep the verdict CONDITIONAL (no change), but the verification should include the cache-instrumentation check above.","tokens_in":19460,"tokens_out":10671,"duration_ms":102564,"concrete_test":"Release the inference runtime and run a single 3-minute rollout whose trajectory revisits an early region only after the sink+top-k+recent window has long passed. Instrument the code to log the number of resident KV chunks each step: if the count saturates at the window size rather than growing with rollout length, non-selected history is evicted. Then compare the initial and revisit observations of that region using the paper's own appearance-consistency criterion; if they diverge once the region's chunks leave the window, the §4.2.1 full-fidelity claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is real-time, minute-scale rollout with coherent revisits. That claim rests on the memory mechanism of §4.2.1, which states: 'we keep the entire KV cache in full fidelity' and explicitly rejects discarding history with a sliding-window strategy. However, §4.3 says: 'To support minutes-long generation beyond our training horizon, we adopt a rolling/sliding-window caching scheme (Xiao et al., 2024). Specifically, the KV cache is organized as a concatenation of [sink chunks, top-k chunks, recent context chunks].' This is the cache described by Eq. (2), not the full history. If non-selected middle-history chunks are evicted, then 'full-fidelity' memory is false: a region visited early and revisited after more than the sink/top-k/recent budget has passed cannot be retrieved. The qualitative revisit tests in Fig. 9 appear to operate over short offsets and do not exercise eviction. The paper reports no memory-vs-length curve and no ablation of top-k or the cache window, so the reader cannot tell whether the bounded cache is an active-attention selection over retained history or an eviction policy. The DA3 training/evaluation loop identified by the reader is a real secondary concern, but this memory contradiction is internal to the method and undermines the long-horizon coherence claim directly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Wonder, a real-time camera-controllable video world model that builds on a Wan2.1-based bidirectional teacher and distills it into a few-step autoregressive student. Three contributions are claimed: a pixel-space coordinate field (rendered lattice plus environment map) for camera control; a sparse attention memory mechanism that retrieves a constant-size set of relevant chunks from a full-fidelity KV cache; and a distillation improvement using a mixture-of-students sampler and a GAN-based control regularizer. The authors evaluate image-to-video and video-to-video world modeling on custom benchmarks, reporting better visual-quality scores and lower relative pose error than several recent streaming world models, and claim minute-scale generation at 16 FPS with stable latency.","tokens_in":19868,"tokens_out":3482,"duration_ms":32497,"significance":"If the central claims hold, Wonder would be a notable systems contribution: it unifies camera-controllable generation, long-horizon memory, and real-time streaming in one framework, and the paper provides detailed equations for the control representation, memory selection, and distillation losses. The manuscript also offers a concrete, reproducible-looking data curation pipeline and comparisons against five recent baselines. However, the current evidence is not yet convincing. The memory mechanism is described inconsistently across §4.2.1 and §4.3, the camera-following evaluation is partly circular with the training-time pose estimator, the three claimed contributions are never ablated, and Table 1 contains no uncertainty or significance information. The 16 FPS claim is asserted without a latency measurement protocol.","major_comments":[{"comment":"There is an internal contradiction about the memory mechanism. §4.2.1 states that 'we keep the entire KV cache in full fidelity' and explicitly argues against sliding-window strategies, while §4.3 describes a 'rolling/sliding-window caching scheme' whose cache is organized as '[sink chunks, top-k chunks, recent context chunks]'. Under the top-k selection in Eq. (2), middle-history chunks that are neither initial, recent, nor selected are effectively evicted from the active cache. If they are evicted, the claim of full-fidelity long-horizon memory is false; if they are retained but not attended to, the distinction from a sliding window needs to be clarified. The qualitative revisit tests in Fig. 9 do not demonstrate eviction behavior, and no memory-vs-length curve or top-k ablation is provided. Please state exactly which KV entries are stored and which are discarded, and provide a quantit","section":"§4.2.1 vs §4.3, Eq. (2)"},{"comment":"The camera-following evaluation is partly circular. In §3, Depth Anything 3 (DA3) is used to estimate camera intrinsics and extrinsics for all real and synthetic training videos, and these recovered trajectories define the pseudo-ground-truth camera actions. In §5.1, DA3 is also one of the two pose estimators used to compute translational and rotational RPE on generated videos. If DA3 is systematically biased, the model is trained to imitate those biases, and the RPE metric rewards videos that reproduce them. This does not establish accurate camera following in an absolute sense. Please report RPE from the two estimators separately, include an independent geometric baseline (e.g., feature-based SLAM or structure from motion where applicable), and discuss the risk of training/evaluation alignment.","section":"§3 and §5.1"},{"comment":"No ablations isolate the three claimed contributions: the pixel-space coordinate field, the sparse full-fidelity memory mechanism, and the distillation changes (mixture of students, GAN control regularization, sparse context forcing). Table 1 compares only the full system against external baselines, so the quantitative gains cannot be attributed to any specific design choice. This is load-bearing because the paper's central message is that these components jointly enable the claimed performance. Please add an ablation study that removes or replaces each component (e.g., pose embedding instead of lattice control, dense attention or sliding window instead of top-k retrieval, single student instead of mixture-of-students, no control GAN) on a fixed benchmark subset.","section":"§5, Table 1"},{"comment":"The quantitative claims lack statistical support. Table 1 reports average scores and RPE values without error bars, confidence intervals, or significance tests; differences such as 0.8558 vs 0.8415 in Table 1 may be within noise. No sample size per condition is given beyond the benchmark totals. In addition, the abstract and conclusion claim 'minute-scale videos at 16 FPS' and §4.3 lists runtime optimizations, but no latency measurement is reported: no hardware details, no measurement protocol, no FPS-vs-length curve, and no comparison of per-step latency as context grows. Please provide confidence intervals or significance tests for all headline numbers, and report measured latency with a clear methodology, ideally as a function of rollout length.","section":"Table 1 and §4.3"}],"minor_comments":[{"comment":"The numerical entries are poorly formatted and often run together (e.g., '0.7900 0.68250.69150.61200.9893'), making the table hard to read. Please add clear separators and align columns.","section":"Table 1"},{"comment":"The notation in Eq. (4), especially the superscripts on G3 and G2, is confusing. It should be stated explicitly which generator corresponds to which denoising step and which generator is reused for the final two steps.","section":"Eq. (4)"},{"comment":"The qualitative revisit comparison should report the temporal offset between the first observation and the revisit, as well as the total rollout length. Currently the figure only illustrates short-range behavior and does not substantiate the 'long-horizon' memory claim.","section":"Fig. 9"},{"comment":"The sentence about manually restructuring 'some PyTorch operations' is vague. If these optimizations are important for the real-time claim, they should be described concretely or removed.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The reader's strongest concern about DA3 circularity is valid, but the more serious issue is the internal contradiction between §4.2.1 and §4.3 on whether the KV cache is full-fidelity or sliding-window. The paper's headline claims require resolving this contradiction and adding ablations and latency measurements. I recommend major revision rather than rejection because the issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: Wonder is a genuine system-level effort on camera-controllable video world models, with a real new idea in the pixel-space coordinate field (synthetic lattice + environment map rendered along the trajectory). That conditioning does something none of the cited pose/ray/point-cloud methods do: it turns camera motion into dense, frame-aligned visual evidence that survives few-step distillation. If that holds up in the released model, it is a useful contribution. The sparse retrieval mechanism (pooled query-key summaries over full-fidelity chunks) is also sensible, and the mixture-of-students plus GAN control regularizer are reasonable engineering additions. The paper is well written and the training pipeline is described in unusual detail.\n\nThe soft spots, in order. First, the memory mechanism as written is internally inconsistent. Section 4.2.1 says 'we keep the entire KV cache in full fidelity' and explicitly rejects sliding-window eviction. Section 4.3 then says for minute-scale generation they adopt a 'rolling/sliding-window caching scheme' where the KV cache is organized as sink + top-k + recent chunks. That means non-selected middle history is evicted, so a region visited early and revisited after the top-k/recent budget has passed cannot be retrieved. The qualitative revisit test in Fig. 9 appears to use a short trajectory that does not exercise eviction. Without a memory-vs-length curve or a top-k ablation, the long-horizon revisit claim is unproven. The stress-test note correctly identifies this.\n\nSecond, the DA3 training/evaluation loop is a real, if secondary, concern. DA3 generates the pseudo-ground-truth camera trajectories for training and is also one of the two pose estimators used in the RPE evaluation. The model is trained to reproduce DA3's biases, and the metric partially rewards doing so. ViPE as a second estimator helps but does not eliminate the circularity. Third, Table 1 has no error bars or significance tests, there are no ablations isolating the three claimed contributions, and the 16 FPS / stable-latency / minute-scale claims are asserted rather than benchmarked. These are fixable in a revision but they matter.\n\nWho gets value: researchers building interactive video generators will want this as a baseline and for the control representation. It deserves a serious referee. My recommendation: send to peer review, but require the authors to clarify the caching policy (what is actually stored vs. attended), provide a memory curve and top-k sensitivity, release code/weights so the 16 FPS and RPE numbers can be independently checked, and ideally decouple the DA3 dependence in evaluation.","headline":"Serious system paper with a promising camera-control representation, but the memory mechanism is described inconsistently and the headline throughput/latency/revisit numbers are asserted rather than measured.","tokens_in":20323,"tokens_out":2421,"would_cite":true,"duration_ms":25414,"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":"Wonder claims a single image can become a playable world: pixel-rendered camera motion, full-fidelity sparse memory, and stabilized distillation together yield real-time, minute-scale, control-accurate video world generation.","keywords":["video world model","camera control","autoregressive video diffusion","sparse attention memory","distribution matching distillation","image-to-video generation","video-to-video re-rendering","real-time streaming generation"],"falsifier":"Re-run the image-to-video benchmark on synthetic scenes with known camera trajectories — the paper's own UE and Blender renders — so target poses are ground truth rather than estimates from an external pose model, and check whether the reported translation and rotation error hold up. Also extend the revisit test beyond the training horizon: if a camera re-enters a region first generated more than a minute earlier and scene identity degrades while per-step latency stays flat, the claim that sparse full-fidelity memory preserves revisits at constant cost is falsified.","tokens_in":19367,"feed_emoji":"🎥","tokens_out":12219,"duration_ms":101505,"temperature":0.7,"pith_summary":"Wonder is a video world model built to do three things simultaneously — follow user camera controls precisely, remember a scene across minute-long exploration, and generate video in real time — and its central claim is that prior systems fail because control, memory, and training were optimized in isolation rather than as one coupled system. The paper's answer to camera control is a pixel-space coordinate field: a synthetic lattice scaffold plus a colored environment map is rendered along the target camera trajectory and fed to the generator as frame-aligned visual evidence, so the model can read translation and rotation directly from pixels instead of inferring them from abstract pose embeddings. Memory is kept lossless but cheap: the full history stays in the KV cache, yet each step attends only to a constant-size set of chunks selected by lightweight pooled summaries, and a sparse-context-forcing training schedule teaches the student to generate under exactly that retrieval condition. Distillation is stabilized by splitting the four denoising steps across three stage-specialized students and adding a camera-aware adversarial regularizer that supervises low-frequency camera consistency. If the claims hold, this is the first system reported to combine accurate camera following, persistent revisits, and stable 16 FPS latency over minute-scale rollouts in a single model.","feed_headline":"One image becomes an explorable world at 16 FPS","feed_subtitle":"Pixel-rendered camera cues and sparse full-fidelity memory keep minute-long rollouts coherent in real time.","key_machinery":"Pixel-space coordinate field (the lattice representation): a lightweight renderer projects a synthetic 3D lattice scaffold and a colored environment map at infinity along the target camera trajectory, rendering conditioning frames for the noisy latents that expose translation and rotation as visual evidence — precise like rendering, open-ended like navigation — and that survive few-step distillation better than abstract pose embeddings. Sparse full-fidelity memory keeps every historical key/value chunk but attends only to a constant-size active set (first chunk, recent chunks, top-k similarity chunks chosen by pooled query-key summaries), trained with sparse context forcing. Distillation pai","core_discovery":"Central claim: a camera-controllable video world model can be real-time, long-horizon, and control-accurate at once if control, memory, and distillation are co-designed. The load-bearing choice is the pixel-space coordinate field: a renderer projects a synthetic lattice scaffold plus a colored environment map along the target camera path into frame-aligned visuals — motion becomes pixel evidence, and stays usable beyond the observed view because the scene is synthetic. Memory keeps the full KV cache but attends only to a constant-size retrieved set, trained via sparse context forcing. Distillation adds three stage-specialized students and a camera-aware GAN regularizer, yielding the lowest p","pith_inferences":["The camera-following evidence chain is close to circular: the same external pose estimator that labels the training videos is also one of the two estimators that score the generated videos, so evaluating on synthetic scenes with known ground-truth trajectories — the UE and Blender data the authors built for training — is the decisive test of whether reported control accuracy reflects genuine camer","The 'motion as rendered visual evidence' idea is a general recipe the paper only applies to cameras: object motion, lighting direction, or focal-length changes could be rendered the same way, and the paper's distillation-stability argument predicts such rendered cues would survive few-step training better than embedding-based controls.","The mixture-of-students design is described as stage specialization, but nothing in the argument limits it to denoising steps; splitting capacity by horizon length or scene type behind the same streaming interface, with one shared KV cache, is a natural and testable extension."],"forward_implications":["If the central claim holds, interactive world models become deployable: per-step latency stays flat as the generated history grows, so minute-scale rollouts run at 16 FPS on current hardware rather than slowing linearly with context length.","Camera control no longer depends on reconstructing the input scene first: since the conditioning scaffold is synthetic, generation continues into unseen regions where point-cloud re-rendering would produce empty frames.","One model can both start a world from a single image and re-shoot an existing video along new camera paths while preserving subject motion, opening real-time 4D video re-editing as a direct application.","The sparse-context-forcing training schedule gives other autoregressive video systems a stable way to teach causal students retrieval-based memory, rather than applying sparse attention only at inference and hoping the model adapts."],"fun_headline_variants":["Turn one image into a real-time 16 FPS explorable world","Camera controls turn a single image into a 16 FPS world","Minute-scale video worlds from one image, navigable at 16 FPS","Pixel coordinate fields turn camera moves into video worlds","Sparse memory and camera cues enable explorable video worlds"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole camera-control evidence chain rests on one external pose-estimation model, which supplies the pseudo-ground-truth camera trajectories the model trains on and independently measures half of the evaluation poses, so systematic bias in that estimator would be learned by the model and then rewarded by the camera-following metric.","fun_headline_variants_meta":{"raw":{"variants":["Turn one image into a real-time 16 FPS explorable world","Camera controls turn a single image into a 16 FPS world","Minute-scale video worlds from one image, navigable at 16 FPS","Pixel coordinate fields turn camera moves into video worlds","Sparse memory and camera cues enable explorable video worlds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001119,"raw_usage":{"total_tokens":4498,"prompt_tokens":755,"completion_tokens":3743,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":3668}},"tokens_in":499,"tokens_out":3743,"duration_ms":23215,"temperature":1.0,"reasoning_tokens":3668,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:46:22.006458+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the image-to-video benchmark on synthetic scenes with known camera trajectories — the paper's own UE and Blender renders — so target poses are ground truth rather than estimates from an external pose model, and check whether the reported translation and rotation error hold up. Also extend the revisit test beyond the training horizon: if a camera re-enters a region first generated more than a minute earlier and scene identity degrades while per-step latency stays flat, the claim that sparse full-fidelity memory preserves revisits at constant cost is falsified.","supporting_citations":[],"review_version":1}