Pith. sign in

REVIEW 4 major objections 4 minor 30 references

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.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 00:46 UTC pith:4QQNNDFF

load-bearing objection 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. the 4 major comments →

arxiv 2607.26037 v1 pith:4QQNNDFF submitted 2026-07-28 cs.CV cs.GR

Wonder: Video World Model Done Better

Jiacong Xu , Hanwen Jiang , Zhixin Shu , Kalyan Sunkavalli , Vishal M. Patel , Yiqun Mei This is my paper
classification cs.CV cs.GR
keywords video world modelcamera controlautoregressive video diffusionsparse attention memorydistribution matching distillationimage-to-video generationvideo-to-video re-renderingreal-time streaming generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

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.

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 (4)
  1. [§4.2.1 vs §4.3, Eq. (2)] 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
  2. [§3 and §5.1] 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.
  3. [§5, Table 1] 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.
  4. [Table 1 and §4.3] 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.
minor comments (4)
  1. [Table 1] 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.
  2. [Eq. (4)] 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.
  3. [Fig. 9] 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.
  4. [§4.3] 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.

Circularity Check

1 steps flagged

Partial evaluation coupling: camera-control training labels and the RPE metric both come partly from Depth Anything 3, so part of the reported camera-following gain is agreement with the training labeler. The central visual-quality, diversity, and speed claims remain externally benchmarked; the §4.2.1 vs §4.3 memory contradiction is a consistency gap, not an equation-level circularity.

specific steps
  1. fitted input called prediction [§3 Video Labeling & Pose Estimation and §5.1 Metrics]
    "we use DepthAnything3(DA3)(Linetal.,2025)toestimatecameraintrinsicsandextrinsicsfor both real and synthetic videos. The recovered trajectories are converted into discrete camera actions. ... To evaluate camera-following accuracy, we estimate relative camera poses from each generated video using two pose-estimation models, Depth Anything 3 (Lin et al., 2025) and ViPE (Huang et al., 2025a), and average their pose errors for a more robust assessment."

    Training supervision for camera control and the RPE metric used to claim action accuracy are both produced by DA3. Wonder is trained to follow camera trajectories that DA3 extracted from real and synthetic videos, and then DA3 is one of the two estimators that recover trajectories from generated videos for RPE (Umeyama-aligned to the target). If DA3 carries systematic bias, the model that best imitates DA3's own training labels will also score best on the DA3 half of RPE, so part of the reported 'better camera following' measures agreement with the training labeler rather than external camera fidelity. The reduction is partial, not exact: ViPE gives an independent estimate, the two are averaged, and no parameter is fit to the test set; the visual-quality and speed claims are unaffected.

full rationale

Walking the claimed derivation chain: the control, memory, and distillation components are specified by in-paper equations (Eq. 1-9) with independent content. In particular, Eq. (2)'s pooled query-key selection and Eq. (3)'s sparse context forcing implement a deliberately train/test-matched design (the paper says the student is 'warmed up with the same memory constraint it will encounter at inference'), which is methodological alignment, not a prediction reducing to a fitted parameter. No uniqueness theorem is invoked, and no ansatz is smuggled in through citation: the lattice control and sparse full-fidelity memory are derived in-paper. The only genuine partial circular pattern is the camera-action evaluation coupling quoted above: DA3 generates the training camera labels in §3 and is one of the two RPE estimators in §5.1, so the action-accuracy comparison is partially self-referential. The coupling is partial because ViPE is independent, the reported RPE averages both estimators, and no parameter is fit to the evaluation set. Self-citations are not load-bearing: RELIC (co-authored by Yiqun Mei) is cited for progressive-horizon training, replayed back-propagation, and runtime optimization — technique borrowings — and RELIC also appears as an external baseline in the comparison table; Xu et al. 2025 (Freevis, overlapping authors) supports only the motivating claim that video diffusion models learn cross-frame correspondences. One flagged inconsistency is real but is not circularity: §4.2.1 asserts 'we keep the entire KV cache in full fidelity' and rejects sliding-window eviction, while §4.3 adopts 'a rolling/sliding-window caching scheme... the KV cache is organized as a concatenation of [sink chunks, top-k chunks, recent context chunks]'. If non-selected middle-history chunks are evicted, minute-scale revisit coherence is unsupported; this is an evidence gap / omitted proof for the long-horizon claim, weighed as correctness risk rather than as a reduction of a predicted quantity to its input.

Axiom & Free-Parameter Ledger

7 free parameters · 7 axioms · 1 invented entities

The central contribution rests on pretrained-teacher quality, DA3 pseudo-labels, a hand-designed lattice conditioning representation, sparse retrieval similarity, and long-horizon extrapolation assumptions. None of these are machine-checked or backed by released artifacts, so the reader is asked to accept several domain assumptions on faith.

free parameters (7)
  • top-k memory budget k = not reported
    The sparse attention mechanism in Eq. 2 selects top-k historical chunks, but the exact value of k is never given. It directly controls the constant-size memory and latency claims.
  • recent-context window r = 2
    Set in §4.2.1 to preserve local motion smoothness; this is a hand-chosen architectural constant.
  • optional memory drop probability p_l = not reported
    Eq. 3 samples attention masks with Bernoulli keep probability 1-p_l per layer, with 'stronger drop rate for longer target horizons'; the schedule is not specified.
  • GAN control loss weight λ_ctrl = not reported
    The final generator loss L_DMD + λ_ctrl L_ctrl in §4.2.3 depends on a scalar weight that is never disclosed.
  • high-noise range for control discriminator τ = not reported
    Eq. 5 samples τ from a 'high-noise range'; the bounds of that range are not provided.
  • Mixture-of-students stage assignment = G1/G2/G3/G3
    The 4-step sampler in Eq. 4 assigns one generator to the first step, one to the second, and another to the last two steps; this hand-designed allocation is not compared against alternatives.
  • Lattice scaffold and environment-map design = not released
    The pixel-space coordinate field in §4.1.1 is a hand-designed synthetic scene (dense 3D scaffold + colored infinite environment map). Its geometry, density, and color scheme are free design choices on which the control representation depends.
axioms (7)
  • domain assumption The pre-trained Wan2.1-I2V-14B bidirectional diffusion model provides a strong, adaptable teacher and backbone.
    §4.1.2 builds all stages on Wan's VAE and DiT; if this base cannot be made causal and streamable, the system fails.
  • domain assumption Depth Anything 3 returns sufficiently accurate camera intrinsics/extrinsics for real and synthetic videos.
    §3 uses DA3 to label all training trajectories, and §5.1 uses DA3 as a primary RPE evaluator. The entire camera-control evidence chain depends on this estimator.
  • domain assumption VLM-based captioning and quality filtering do not systematically bias training data toward certain motions or scenes.
    §3 filters the raw pool with VLM quality checks and hierarchical captions; biased filtering could make both training and benchmarking easier.
  • ad hoc to paper Rendering a synthetic 3D lattice and infinite environment map yields an unambiguous, learnable pixel-aligned encoding of translation and rotation.
    §4.1.1 claims lattice projections induce parallax and orientation cues; the model's ability to decode them is asserted and demonstrated only through internal benchmarks.
  • domain assumption Pooled query-key similarities capture chunk-level relevance well enough for memory retrieval.
    Eq. 2 selects memory chunks by similarity of pooled key summaries; no retrieval accuracy analysis or ablation is provided.
  • domain assumption DMD/self-forcing score residuals remain valid when the student rollout drifts far from the teacher's training distribution.
    §4.2.3 motivates GAN control regularization because teacher guidance degrades under drift; standard DMD assumptions are carried over from Yin et al., 2024b.
  • domain assumption RoPE extrapolation with YaRN and rolling-window frame remapping preserve positional semantics beyond the 20s training horizon.
    §4.1.3 and §4.3 use YaRN and frame-index remapping for minute-scale rollouts; this is an extrapolation assumption with no long-horizon quantitative evaluation.
invented entities (1)
  • Pixel-space coordinate field (lattice representation) no independent evidence
    purpose: A synthetic 3D lattice plus infinite environment map rendered along the target camera trajectory and concatenated with noisy latents, providing visual camera-control cues for translation, rotation, and parallax.
    The representation is introduced in §4.1.1 and validated only through this paper's internal benchmarks. No released code, ablations, or external experiments verify that the rendered cues are uniquely decodable by the model.

pith-pipeline@v1.3.0-alltime-deepseek · 19216 in / 14982 out tokens · 152411 ms · 2026-08-01T00:46:22.006458+00:00 · methodology

0 comments
read the original abstract

We present Wonder, a general-purpose video world model for real-time, camera-controllable world exploration. Given an image or a conditional video, Wonder constructs a playable world where users can navigate interactively by moving the camera, discovering unseen regions, and revisiting previously observed areas in real time and over a long-term horizon. Achieving this capability requires a system-level co-design of control method, memory mechanism, and training strategy. We introduce a novel camera conditioning with a dense coordinate field whose renderings provide spatially aligned motion and orientation cues, allowing the model to interpret camera motion directly as visual evidence. To support fast and precise memory retrieval over a growing generation context, we propose an efficient sparse attention-based memory mechanism, enabling the model to selectively attend to a small set of relevant context tokens at inference time, regardless of actual context length. We further develop several techniques to rectify the self-forcing-style distillation pipeline, improving the student model's ability to respect control signals, as well as maintaining diverse generation modes and long-term memory from the teacher. Together, these components enable Wonder to synthesize diverse, minute-scale videos at 16 FPS while preserving coherent geometry, appearance, and dynamics across long rollouts. Beyond image-to-video generation, Wonder naturally supports video-conditioned generation, allowing existing dynamic scenes to be re-shot in real time.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 23 linked inside Pith

  1. [1]

    Recammaster: Camera-controlled generative rendering from a single video

    Jianhong Bai, Menghan Xia, Xiao Fu, Xintao Wang, Lianrui Mu, Jinwen Cao, Zuozhu Liu, Haoji Hu, Xiang Bai, Pengfei Wan, et al. Recammaster: Camera-controlled generative rendering from a single video. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14834–14844, 2025a. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui ...

    Pith/arXiv arXiv 2025

  2. [4]

    Mineworld: a real-time and open-source interactive world model on minecraft.arXiv preprint arXiv:2504.08388,

    Junliang Guo, Yang Ye, Tianyu He, Haoyu Wu, Yushu Jiang, Tim Pearce, and Jiang Bian. Mineworld: a real-time and open-source interactive world model on minecraft.arXiv preprint arXiv:2504.08388,

    Pith/arXiv arXiv

  3. [5]

    Ltx-2: Efficient joint audio-visual foundation model.arXiv preprint arXiv:2601.03233,

    Yoav HaCohen, Benny Brazowski, Nisan Chiprut, Yaki Bitterman, Andrew Kvochko, Avishai Berkowitz, Daniel Shalem, Daphna Lifschitz, Dudu Moshe, Eitan Porat, et al. Ltx-2: Efficient joint audio-visual foundation model.arXiv preprint arXiv:2601.03233,

    Pith/arXiv arXiv

  4. [7]

    Acdit: Interpolating autoregressive conditional modeling and diffusion transformer.arXiv preprint arXiv:2412.07720,

    Jinyi Hu, Shengding Hu, Yuxuan Song, Yufei Huang, Mingxuan Wang, Hao Zhou, Zhiyuan Liu, Wei-Ying Ma, and Maosong Sun. Acdit: Interpolating autoregressive conditional modeling and diffusion transformer.arXiv preprint arXiv:2412.07720,

    arXiv

  5. [8]

    Vipe: Video pose engine for 3d geometric perception

    Jiahui Huang, Qunjie Zhou, Hesam Rabeti, Aleksandr Korovko, Huan Ling, Xuanchi Ren, Tianchang Shen, Jun Gao, Dmitry Slepichev, Chen-Hsuan Lin, Jiawei Ren, Kevin Xie, Joydeep Biswas, Laura Leal-Taixe, and Sanja Fidler. Vipe: Video pose engine for 3d geometric perception. InNVIDIA Research Whitepapers, 2025a. Junchao Huang, Xinting Hu, Boyao Han, Shaoshuai ...

    arXiv

  6. [9]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    Pith/arXiv arXiv

  7. [10]

    Hunyuan- gamecraft: High-dynamic interactive game video generation with hybrid history condition.arXiv preprint arXiv:2506.17201,

    Jiaqi Li, Junshu Tang, Zhiyong Xu, Longhuang Wu, Yuan Zhou, Shuai Shao, Tianbao Yu, Zhiguo Cao, and Qinglin Lu. Hunyuan- gamecraft: High-dynamic interactive game video generation with hybrid history condition.arXiv preprint arXiv:2506.17201,

    Pith/arXiv arXiv

  8. [11]

    Depth anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647,

    Haotong Lin, Sili Chen, Junhao Liew, Donny Y Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647,

    Pith/arXiv arXiv

  9. [12]

    Rolling forcing: Autoregressive long video diffusion in real time.arXiv preprint arXiv:2509.25161,

    Kunhao Liu, Wenbo Hu, Jiale Xu, Ying Shan, and Shijian Lu. Rolling forcing: Autoregressive long video diffusion in real time.arXiv preprint arXiv:2509.25161,

    Pith/arXiv arXiv

  10. [13]

    Yume: An interactive world generation model.arXiv preprint arXiv:2507.17744,

    XiaofengMao, ShaohengLin, ZhenLi, ChuanhaoLi, WenshuoPeng, TongHe, JiangmiaoPang, MingminChi, YuQiao, andKaipeng Zhang. Yume: An interactive world generation model.arXiv preprint arXiv:2507.17744,

    Pith/arXiv arXiv

  11. [14]

    Yarn: Efficient context window extension of large language models.arXiv preprint arXiv:2309.00071,

    Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. Yarn: Efficient context window extension of large language models.arXiv preprint arXiv:2309.00071,

    Pith/arXiv arXiv

  12. [15]

    Advancing open-source world models.arXiv preprint arXiv:2601.20540,

    Robbyant Team. Advancing open-source world models.arXiv preprint arXiv:2601.20540,

    Pith/arXiv arXiv

  13. [16]

    Flashattention-3: Fast and accurate attention with asynchrony and low-precision.ArXiv, abs/2407.08608, 2024.https://api.semanticscholar.org/CorpusID: 271098045

    Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, and Tri Dao. Flashattention-3: Fast and accurate attention with asynchrony and low-precision.ArXiv, abs/2407.08608, 2024.https://api.semanticscholar.org/CorpusID: 271098045. Joonghyuk Shin, Zhengqi Li, Richard Zhang, Jun-Yan Zhu, Jaesik Park, Eli Schechtman, and Xun Huang. Motionstream...

    Pith/arXiv arXiv 2024

  14. [17]

    Hunyuan-gamecraft-2: Instruction-following interactive game world model.arXiv preprint arXiv:2511.23429,

    Junshu Tang, Jiacheng Liu, Jiaqi Li, Longhuang Wu, Haoyu Yang, Penghao Zhao, Siruis Gong, Xiang Yuan, Shuai Shao, Linfeng Zhang, et al. Hunyuan-gamecraft-2: Instruction-following interactive game world model.arXiv preprint arXiv:2511.23429,

    arXiv

  15. [18]

    Dreamx-world 1.0: A general-purpose interactive world model.arXiv preprint arXiv:2606.16993, 2026a

    DreamX Team, Yancheng Bai, Rui Chen, Xiangxiang Chu, Rujing Dang, Hao Dou, Bingjie Gao, Qiwen Gu, Siyu Hong, Jiachen Lei, et al. Dreamx-world 1.0: A general-purpose interactive world model.arXiv preprint arXiv:2606.16993, 2026a. HunyuanWorld Team, Zhenwei Wang, Yuhao Liu, Junta Wu, Zixiao Gu, Haoyuan Wang, Xuhui Zuo, Tianyu Huang, Wenhuan Li, Sheng Zhang,...

    arXiv

  16. [19]

    Inspatio-world: Areal-time4dworldsimulatorviaspatiotemporalautoregressivemodeling.arXivpreprintarXiv:2604.07209, 2026b

    InSpatio Team, Donghui Shen, Guofeng Zhang, Haomin Liu, Haoyu Ji, Hujun Bao, Hongjia Zhai, Jialin Liu, Jing Guo, Nan Wang, etal. Inspatio-world: Areal-time4dworldsimulatorviaspatiotemporalautoregressivemodeling.arXivpreprintarXiv:2604.07209, 2026b. Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns.IEEE Transa...

    Pith/arXiv arXiv

  17. [21]

    Bullettime: Decoupled control of time and camera pose for video generation

    Yiming Wang, Qihang Zhang, Shengqu Cai, Tong Wu, Jan Ackermann, Zhengfei Kuang, Yang Zheng, Frano Rajič, Siyu Tang, and Gordon Wetzstein. Bullettime: Decoupled control of time and camera pose for video generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18319–18330, 2026a. Zhouxia Wang, Ziyang Yuan, Xinta...

    2024

  18. [22]

    Matrix-game 3.0: Real-time and streaming interactive world model with long-horizon memory.arXiv preprint arXiv:2604.08995, 2026b

    Zile Wang, Zexiang Liu, Jiaxing Li, Kaichen Huang, Baixin Xu, Fei Kang, Mengyin An, Peiyu Wang, Biao Jiang, Yichen Wei, et al. Matrix-game 3.0: Real-time and streaming interactive world model with long-horizon memory.arXiv preprint arXiv:2604.08995, 2026b. 20 Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language...

    Pith/arXiv arXiv 2024

  19. [23]

    World-env: Leveraging world model as a virtual environment for vla post-training.arXiv preprint arXiv:2509.24948,

    Junjin Xiao, Yandan Yang, Xinyuan Chang, Ronghan Chen, Feng Xiong, Mu Xu, Wei-Shi Zheng, and Qing Zhang. World-env: Leveraging world model as a virtual environment for vla post-training.arXiv preprint arXiv:2509.24948,

    Pith/arXiv arXiv

  20. [24]

    Freevis: Training-free video stylization with inconsistent references.arXiv preprint arXiv:2510.01686,

    Jiacong Xu, Yiqun Mei, Ke Zhang, and Vishal M Patel. Freevis: Training-free video stylization with inconsistent references.arXiv preprint arXiv:2510.01686,

    arXiv

  21. [25]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. InInternational Conference on Learning Representations, volume 2025, pages 83048–83077,

    2025

  22. [26]

    Context as memory: Scene-consistent interactive long video generation with memory retrieval.arXiv preprint arXiv:2506.03141,

    Jiwen Yu, Jianhong Bai, Yiran Qin, Quande Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Context as memory: Scene-consistent interactive long video generation with memory retrieval.arXiv preprint arXiv:2506.03141,

    Pith/arXiv arXiv

  23. [27]

    Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048,

    Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048,

    Pith/arXiv arXiv

  24. [28]

    Matrix-game: Interactive world foundation model.arXiv preprint arXiv:2506.18701,

    Yifan Zhang, Chunli Peng, Boyang Wang, Puyi Wang, Qingcheng Zhu, Fei Kang, Biao Jiang, Zedong Gao, Eric Li, Yang Liu, et al. Matrix-game: Interactive world foundation model.arXiv preprint arXiv:2506.18701,

    Pith/arXiv arXiv

  25. [29]

    Pytorch fsdp: experiences on scaling fully sharded data parallel.arXiv preprint arXiv:2304.11277,

    Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel.arXiv preprint arXiv:2304.11277,

    Pith/arXiv arXiv

  26. [2002]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Pith/arXiv arXiv

  27. [2023]

    Sana-wm: Efficient minute-scale world modeling with hybrid linear diffusion transformer.arXiv preprint arXiv:2605.15178, 2026a

    Haoyi Zhu, Haozhe Liu, Yuyang Zhao, Tian Ye, Junsong Chen, Jincheng Yu, Tong He, Song Han, and Enze Xie. Sana-wm: Efficient minute-scale world modeling with hybrid linear diffusion transformer.arXiv preprint arXiv:2605.15178, 2026a. Hongzhou Zhu, Min Zhao, Guande He, Hang Su, Chongxuan Li, and Jun Zhu. Causal forcing: Autoregressive diffusion distillation...

    Pith/arXiv arXiv

  28. [2024]

    Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining.ArXiv, abs/2304.09151, 2023.https://api

    Hyung Won Chung, Noah Constant, Xavier García, Adam Roberts, Yi Tay, Sharan Narang, and Orhan Firat. Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining.ArXiv, abs/2304.09151, 2023.https://api. semanticscholar.org/CorpusID:258187051. Justin Cui, Jie Wu, Ming Li, Tao Yang, Xiaojie Li, Rui Wang, Andrew Bai, Yuanhao B...

    Pith/arXiv arXiv 2023

  29. [2025]

    Ca2-vdm: Efficient autoregressive video diffusion model with causal generation and cache sharing.arXiv preprint arXiv:2411.16375,

    Kaifeng Gao, Jiaxin Shi, Hanwang Zhang, Chunping Wang, Jun Xiao, and Long Chen. Ca2-vdm: Efficient autoregressive video diffusion model with causal generation and cache sharing.arXiv preprint arXiv:2411.16375,

    Pith/arXiv arXiv

  30. [2026]

    Cameractrl: Enablingcameracontrol for video diffusion models

    HaoHe, YinghaoXu, YuweiGuo, GordonWetzstein, BoDai, HongshengLi, andCeyuanYang. Cameractrl: Enablingcameracontrol for video diffusion models. InThe Thirteenth International Conference on Learning Representations, 2025a. Xianglong He, Chunli Peng, Zexiang Liu, Boyang Wang, Yifan Zhang, Qi Cui, Fei Kang, Biao Jiang, Mengyin An, Yangyang Ren, et al. Matrix-g...

    Pith/arXiv arXiv