Pith. sign in

REVIEW 3 major objections 7 minor 114 references

A single video editor can detect a first-person trigger, leave earlier frames alone, and apply the edit only afterward—without a separate detector.

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 ↗

A single diffusion editor jointly detects egocentric events and applies post-trigger edits, beating standard and cascaded baselines on Ego4D with a new 106K-pair dataset and event-aware protocol.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Real task + data + protocol for event-triggered egocentric editing; headline Ego4D ranking is useful but partly judge-coupled and weaker on negatives. the 3 major comments →

arxiv 2607.24560 v1 pith:DIKIBC43 submitted 2026-07-27 cs.CV cs.AI

EgoPlay: Event-Triggered Video Editing for Egocentric Streams

Jinjie Mai , Gordon Guocheng Qian , Willi Menapace , Arpit Sahni , Chaoyang Wang , Ashkan Mirzaei , Runjia Li , Sergey Tulyakov
show 3 more authors
Bernard Ghanem Peter Wonka Rameen Abdal
This is my paper
classification cs.CV cs.AI
keywords video editingegocentric videoevent-triggered editingdiffusion modelsstreaming videoaugmented realityvideo-to-video
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.

The reading

This paper argues that wearable, always-on video effects should fire from what the wearer actually does, not from a pre-scripted timeline or a hand-drawn mask. It introduces EgoPlay, which takes a monocular egocentric clip and a rule of the form “when X happens, do Y,” decides whether and when X occurs, copies the pre-event stream unchanged, and synthesizes the edited continuation only after the trigger. The same model is trained to ignore fabricated triggers and to pick the right edit among several event–edit pairs. Supporting this is a large event-conditioned training set built mainly from first-person footage, a bidirectional diffusion editor that learns the joint behavior end-to-end, and a causal chunk-by-chunk variant for streaming. On a new event-aware benchmark that scores pre-event preservation, post-event edit quality, and false-trigger restraint separately, the unified model beats both a strong instruction editor that always edits and a detector-then-edit cascade, while using far less memory than the cascade.

Core claim

Event-triggered egocentric video editing can be learned jointly inside one video-to-video diffusion model: given only pixels and a natural-language “when X, do Y” rule, the model infers the trigger boundary, preserves pre-trigger frames, applies the edit only afterward, and rejects absent or distractor triggers—outperforming both always-on instruction editors and explicit detector–editor pipelines on averaged event-aware metrics without receiving ground-truth timestamps at inference.

What carries the argument

Event-triggered V2V supervision: positive targets keep the pre-event source, blend a short VAE latent transition at the boundary, then show an edited post-event suffix; negative targets are identical to the source; multi-event prompts mix true and fabricated triggers. A diffusion transformer is fine-tuned on this signal so localization and editing emerge together; a block-causal student with diffusion forcing then enables chunk-by-chunk streaming.

Load-bearing premise

The main ranking rests on an automatic video-language judge of temporally split clips being a faithful score of edit timing and false-trigger rejection, even though that judge can still give perfect scores when an edit wrongly appears on a no-trigger clip.

What would settle it

Re-score the same Ego4D four-mode suite with human raters (or a judge proven not to reward false edits on multiple-negative clips) and check whether the unified model still leads the detector–editor cascade and the always-edit baseline on averaged editing quality, visual quality, and background consistency without ground-truth boundaries.

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

If this is right

  • Always-on AR rules such as “when I upturn my palm, add a fireball” can run from a single RGB stream without masks, timestamps, or per-clip scripts.
  • Joint training can replace a resident detector plus editor stack and cut GPU memory by more than half at similar latency.
  • A causal chunked student nearly matches the bidirectional teacher, so offline event-triggered quality can transfer to streamable inference.
  • Negative and multi-event prompt supervision becomes a standard way to teach when not to edit, not only how to edit.
  • Event-aware scoring that splits pre-event, event, and post-event segments becomes the right protocol for this task class.

Where Pith is reading between the lines

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

  • If joint localization-and-edit training generalizes, the same recipe could condition wearable effects on multi-step routines (open fridge, then grab milk) rather than single completion points.
  • The remaining gap on hard negatives suggests hybrid systems that keep a cheap rejector only for high-stakes no-edit cases while generating with the unified model on positives.
  • Eviction-aware causal training called out for long streams is the natural bridge from five-second clips to all-day first-person capture.
  • Smooth learned transitions at the trigger may matter as much for user comfort in AR as raw edit fidelity.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 7 minor

Summary. The paper introduces event-triggered video-to-video editing for egocentric streams: given a prompt of the form "when X happens, do Y," the model must infer whether/when X occurs from pixels, preserve pre-event frames, and apply edit Y only afterward. The authors build a 106K-pair dataset from Ego4D plus an auxiliary common-video subset using a multi-pass VLM pipeline (positive, fabricated-negative, and multi-event prompts; targets synthesized by running pretrained EgoEdit on post-event segments with VAE-interpolated transition frames), fine-tune a bidirectional diffusion transformer with rectified-flow matching, and derive a causal blockwise variant via diffusion forcing. They propose an event-aware evaluation protocol that decomposes clips at ground-truth boundaries and scores pre/post-event behavior with a blinded Qwen3-VL-30B judge. On the Ego4D benchmark, EgoPlay reports the best four-mode average among non-oracle methods (8.01 vs. 6.99 for a VLM-guided cascade and 6.84 for EgoEdit), corroborated by a 10-participant human preference study on positive samples, an edit-start localization ablation, and an efficiency comparison showing roughly half the cascade's GPU memory.

Significance. If the results hold, this is a useful contribution: it defines a genuinely new task (trigger-conditioned temporal restraint in generative editing), ships a 106K-pair dataset with positive/fabricated-negative/multi-event modes, provides an event-aware evaluation protocol, and demonstrates a causal streamable variant that stays close to the bidirectional teacher in scored quality. The promised public release of the benchmark, prompts, and scoring protocol, plus the honest documentation of evaluator failures (Appendix A.1/Fig. 7) and of the auxiliary-split ranking inversion, are real strengths that make the work falsifiable and re-scorable as better judges appear. The significance is somewhat bounded by reliance on the proprietary EgoEdit backbone (no weight release) and by the evaluation-instrument concerns above.

major comments (3)
  1. [§4.1, §F.2, §C.3, Table 1] §C.3 vs. §4.1/§F.2 (Table 1): the generated training pairs were filtered by Qwen3-VL-235B on 'editing quality, visual quality, and background consistency' — the same three criteria, from the same model family, as the Qwen3-VL-30B-A3B evaluator that produces the entire Table 1 ranking. EgoPlay was therefore trained on data selected to maximize the exact judge metrics on which it is later ranked, while neither baseline was. This is not definitional circularity, but it is a correctness risk for the central claim: part of the +17.7%/16.9%/16.4% margins could be judge-family alignment rather than method superiority. The human study (Table 3) mitigates this only for positive modes. A concrete fix: re-score the Table 1 outputs with a judge from a different model family (e.g., a Gemini- or GPT-class video model), or human-score a stratified subsample covering all four modes, and report whether t
  2. [Appendix A.1, Fig. 7, Table 1 (negative rows)] The paper's own Appendix A.1/Fig. 7 documents the evaluator awarding 9/9 when both EgoPlay variants hallucinate the conditioned edit on a multiple-negative sample. This matters quantitatively, not just qualitatively: negative modes constitute half of the four-mode average, and the VLM-guided cascade's 8.99 negative scores are judge-proof by construction (it copies the source, which any judge scores perfectly), whereas EgoPlay's 8.33/8.14 negative scores are exposed to exactly this false-positive failure. The documented failure mode therefore asymmetrically inflates EgoPlay's averaged score. The authors should quantify the judge's false-positive rate on negative samples (e.g., a human audit of all 300 negative-mode outputs, or at least a subsample) and report corrected negative-mode and averaged scores, or bounds on the correction. This is load-bearing for the 'best averaged performance'
  3. [§B.1, Table 6 vs. §4.2] Table 6 shows the four-mode averaged ranking inverting on the auxiliary common-video split: VLM-guided 7.14 vs. EgoPlay 6.76 (and EgoPlay-Causal 6.29). The inversion is driven by negative modes, where copy-the-source behavior dominates under equal mode weighting. The headline 'best averaged event-triggered performance' is thus benchmark- and weighting-dependent, and the main text (§4.2, abstract, conclusion) does not disclose this — the auxiliary split is deferred to the supplement with the statement that results 'do not change the main conclusion.' Given that the equal weighting of positive and negative modes is a design choice of the authors' own protocol, the main text should (a) acknowledge the auxiliary-split inversion explicitly, and (b) justify the mode weighting or report sensitivity (e.g., positive-only average, where EgoPlay's lead is robust across both splits).
minor comments (7)
  1. [§4.5, Table 5] Table 5: the efficiency comparison is measured on a single positive sample ('averaged over multiple inference runs' of one clip). Latency and memory across a benchmark sample would be more convincing. Also clarify whether the cascade's 102.98 GB is intrinsic or an artifact of keeping the 8B VLM resident without offloading — the 'less than half' memory claim depends on this implementation choice.
  2. [§4.3, Table 3] The human study covers only 30 positive samples drawn from just 5 source videos, with 10 participants. It corroborates the positive-mode ranking but says nothing about negative modes, which is precisely where the VLM judge is least trustworthy. At minimum this scope limitation should be stated alongside Table 3.
  3. [§1, §4.5, Table 4] Table 4: EgoPlay-Causal's Acc@0.1 is 10.0 vs. 45.5 for the bidirectional model. The introduction's claim that the causal variant 'nearly matches' the bidirectional one should be explicitly qualified as referring to editing-quality scores, not edit-start timing, since the timing gap is large.
  4. [§3.1, Eq. (2)] Eq. (2) and §3.1: the transition length M is used in Eq. (2) but only given a value (8 frames, via Qwen-Image-2512 VAE spherical interpolation) in §C.3. Please state M=8 at first use in §3.1.
  5. [§4.4, Table 2] Table 2: EgoPlay slightly underperforms the EgoEdit backbone on standard event-free editing (7.60 vs. 7.76 VLM on EgoEditBench; TC 94.92 vs. 96.70). A sentence acknowledging this small regression from event-triggered fine-tuning would improve completeness.
  6. [Figures 4, 5, 8] Fig. 4, 5, 8: the prompt text in the qualitative figures is very small and partially garbled in places (e.g., 'sink s metal grate', '3D printer s white casing' — dropped apostrophes). Please increase font size and fix the text rendering.
  7. [§D.3, §F.2] §D.3: VLM totals are reported from a single run with ~±0.2 variation; given that several Table 1 margins (e.g., negative modes) are within a few tenths, single-run scoring noise should be quantified for the main table as well, e.g., by repeating the judge with different seeds/sampling and reporting variance.

Circularity Check

0 steps flagged

No derivation-chain circularity: EgoPlay is an empirical fine-tune with held-out benchmarks; synthetic EgoEdit targets and same-family VLM judging are methodological coupling, not self-definitional reduction.

full rationale

This is an empirical systems paper, not a first-principles derivation. The central claim is that a single end-to-end model fine-tuned on event-triggered pairs outperforms (i) the instruction editor EgoEdit and (ii) a VLM detector–editor cascade on a held-out Ego4D event-mode split, with supporting human preference and external-editor spot checks. Positive targets T+ are synthesized by running pretrained EgoEdit on post-event segments plus VAE blends (Sec. 3.1, Eqs. 1–2); EgoPlay is initialized from EgoEdit and trained to map event-conditioned prompts to those targets, including negatives T−=S and multi-event compositions. That is standard synthetic-supervision / teacher initialization, not a claim that a fitted parameter predicts a quantity forced by the fit. At inference the model receives no event timestamp and must localize triggers from pixels; Table 1, the human study (Table 3), edit-start Acc@r (Table 4), and Ditto/VIVA/Runway checks are external comparisons, not rearrangements of the training objective. Same-family VLM filtering of training pairs (App. C.3) and VLM evaluation (App. A.1) raise judge-alignment and validity concerns, but they do not make the reported ranking equal to its inputs by construction—the paper even documents judge failures that hurt its own negative-mode story. No uniqueness theorem, ansatz-via-self-citation, or renamed known law carries the result. Score 1 only for mild self-lineage (EgoEdit backbone and authors) that is not load-bearing for a forced identity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central ranking rests on synthetic event-triggered supervision, frozen teacher editing quality, VLM curation/evaluation, and standard diffusion training choices—not on a closed-form theory. Free parameters are training/inference hyperparameters and data-construction constants; axioms are domain assumptions about Ego4D triggers and judge validity; invented entities are task/dataset constructs rather than physical objects.

free parameters (5)
  • Transition length M (eight VAE spherical-interpolation frames) = M=8 frames
    Chosen in data synthesis to avoid hard cuts; shapes the learned soft edit onset the model is rewarded for reproducing.
  • Quality-filter thresholds (aesthetic, DOVER, blur, VMAF motion, brightness, saturation, jitter, LPIPS) = e.g. aesthetic>3.5, DOVER>0.3, blur mean<6.4, motion<20, brightness[60,130], etc.
    Hand-set gates that determine which Ego4D clips enter the 106K set and thus the learned trigger distribution.
  • Bidirectional/causal fine-tune schedules and LRs = 30k steps; 40-step inference
    30k steps, LR 3e-5 (bi) / 1e-5 (causal), EMA 0.9999, 40 denoising steps—standard but claim-dependent training knobs.
  • Causal block size / block count = N=7 blocks × 3 latent frames
    21 latent frames split into 7 blocks of 3; defines streaming granularity and attention mask.
  • Region weights in VLM evaluator aggregation = 0.5 / 0.5 / 1.0 on pre / event / post
    Positive samples weight pre-event and event preservation at half and post-event at full; changes the headline total.
axioms (5)
  • domain assumption Rectified-flow / diffusion transformer video editing is an adequate substrate for joint implicit event localization and pixel editing without an explicit timestamp input.
    Sec. 3.2 trains only on constructed targets T+/T−; localization is assumed to emerge from the editing objective.
  • domain assumption Ego4D narrations filtered to discrete, visually completable first-person actions are valid proxies for wearable trigger events.
    Sec. 3.1 and Appendix C.1 define the retained event class that the benchmark measures.
  • domain assumption Fabricated non-occurring events paired with edits are sufficient supervision for false-trigger restraint.
    Negative and multi-event construction in Sec. 3.1; Table 1 negative modes test this.
  • ad hoc to paper A larger held-out VLM judge, blinded to trigger text and scored on GT temporal segments, ranks methods in line with human event-triggered preference.
    Sec. 4.1 protocol; partially supported by the human study but contradicted in documented failure cases (Fig. 7).
  • standard math Standard flow-matching / diffusion-forcing mathematics (Lipman et al.; Chen et al.) applies unchanged under channel-wise source conditioning and block-causal masks.
    Eqs. (3)–(4); borrowed training objectives.
invented entities (3)
  • Event-triggered V2V task (prompt form if X then Y with pre-event identity constraint) independent evidence
    purpose: Defines the capability the model and benchmark optimize for.
    Framed as absent from prior editors; operationalized via dataset and metrics rather than a new physical object.
  • 106K event-triggered clip–prompt dataset (positive, fabricated-negative, multi-event) no independent evidence
    purpose: Supplies joint supervision for detection, restraint, and editing.
    Constructed artifact; value depends on filter/VLM synthesis choices; benchmark subset to be released.
  • Event-aware evaluation protocol (pre / event / post decomposition + negative robustness) independent evidence
    purpose: Separates timing errors from edit quality and false triggers.
    New measurement design; still relies on VLM scalars 0–3 per axis.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of EgoPlay: Event-Triggered Video Editing for Egocentric Streams." pith.science (2026). https://pith.science/paper/DIKIBC43

@misc{pith2026260724560,
  author       = {Pith},
  title        = {Pith review of: EgoPlay: Event-Triggered Video Editing for Egocentric Streams},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIKIBC43}},
  note         = {Machine review of arXiv:2607.24560}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce EgoPlay, an event-triggered video-to-video editor for egocentric streams, obtained by fine-tuning a pretrained V2V diffusion transformer on event-conditioned data built primarily from Ego4D. Given a monocular video and an event-triggered prompt of the form "when X happens, do Y," EgoPlay infers whether and when event X occurs, preserves pre-event frames, and applies edit Y only to the post-event continuation. Rather than cascading a separate event detector with an editor, EgoPlay learns event recognition, temporal restraint, and pixel-level editing jointly in a single end-to-end model, while also handling negative and multi-event prompts. To support this, we construct a large-scale dataset of 106K event-triggered clip-prompt pairs spanning positive triggers, fabricated-trigger negatives, and multi-event prompts. We then train a bidirectional video diffusion editor with event-triggered supervision and derive a causal variant for chunk-by-chunk streamable inference. We further introduce an event-aware evaluation protocol that separately measures post-trigger editing quality, pre-trigger preservation, and false-trigger robustness. On the Ego4D benchmark, EgoPlay substantially outperforms EgoEdit, the state-of-the-art instruction-based egocentric video editing baseline, with relative gains of 17.7%, 16.9%, and 16.4% in editing quality, visual quality, and background consistency. It also surpasses a VLM-guided detector-editor baseline by 15.7%, 14.5%, and 13.5% on the same metrics, while using less than half the GPU memory.

Figures

Figures reproduced from arXiv: 2607.24560 by Arpit Sahni, Ashkan Mirzaei, Bernard Ghanem, Chaoyang Wang, Gordon Guocheng Qian, Jinjie Mai, Peter Wonka, Rameen Abdal, Runjia Li, Sergey Tulyakov, Willi Menapace.

Figure 1
Figure 1. Figure 1: EgoPlay Event-Triggered video editing. Bidirectional Source Latent Given a source video and a natural￾language instruction, a conventional video editor can ground where to edit, but always applies the edit throughout the clip. In contrast, EgoPlay grounds when to edit from an event-triggered prompt: it preserves frames before the specified event and applies the edit only after the event occurs. of several … view at source ↗
Figure 2
Figure 2. Figure 2: Data pipeline. Given raw videos and human-annotated event spans, we (1) filter low-quality clips and curate events with clear completion points, (2) generate event-triggered prompts via a multi-pass VLM that produces positive, negative, and event-free instructions, and (3) synthesize edited targets by applying an event-free edit to the post-event segment and blending with VAE transition frames to achieve a… view at source ↗
Figure 3
Figure 3. Figure 3: EgoPlay method. EgoPlay takes a source video and an event￾triggered instruction, jointly infers when the trigger event occurs and what edits to apply with the event. The bidirectional model learns this behavior end-to-end from event-triggered V2V supervision, while the causal variant restricts temporal attention for streamable inference. Architecture. We adapt a pretrained text-to-video diffusion trans￾for… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on Ego4D event-triggered video editing. We compare EgoPlay and EgoPlay-Causal against VLM-guided editing and EgoEdit in the single-positive setting. For each example, three frames are shown: before, during (highlighted), and after the edit. Each edit is conditioned on a temporal trigger (When, italic) specifying the moment the edit should take effect, along with a spatial instruction… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison across diverse evaluation modes. We evaluate four methods—VLM-guided, EgoEdit, EgoPlay, and EgoPlay-Causal—on four increasingly challenging conditioning scenarios. (V2V Edit) Standard video editing with no temporal trigger. (Single Negative) The edit is conditioned on a single negative trigger, a moment when the edit should not apply. (Multiple Negative) Multiple negative triggers (m… view at source ↗
Figure 6
Figure 6. Figure 6: EgoPlay-Causal-Realtime. We further train a 4-step streaming model with self-forcing for real-time deployment. The model can produce successful event-triggered edits at 16 FPS (top row), but it is less robust than our 40-step causal model and can sometimes begin editing from the start of the clip (bottom row). ACM Trans. Graph., Vol. 1, No. 1, Article . Publication date: July 2026 [PITH_FULL_IMAGE:figures… view at source ↗
Figure 7
Figure 7. Figure 7: Limitation of VLM-based evaluation. A multiple-negative Ego4D sample. The conditioned edit (“transform the paint brush into a cluster of thin, glowing amber vines”) is gated on a fabricated trigger (“once I pick up the photo of the child and place it on the notebook”) that never occurs, so the correct behavior is to leave the video unedited. Both EgoPlay and EgoPlay-Causal instead hallucinate the edit by t… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on the auxiliary common-video split (top) and Ego4D (bottom). We compare EgoPlay and EgoPlay-Causal against VLM-guided editing and EgoEdit in the single-positive setting. For each example, three frames are shown: before, during (highlighted), and after the edit. Each edit is conditioned on a temporal trigger (When, italic) specifying the moment the edit should take effect, along with… view at source ↗
Figure 9
Figure 9. Figure 9: User-study interface. Representative frames from the second single-positive benchmark example populate the interface. The study displayed synchronized reference and output videos rather than still images. Method identities were hidden, and the four outputs were randomly ordered for each trial. VLM localizer prompt. The detector stage uses the following prompt template with the source video and the original… view at source ↗
Figure 10
Figure 10. Figure 10: Edit-timing annotation interface and learned edit transition. Annotators mark the first edited frame in each predicted video, with the ground￾truth trigger-end as reference. EgoPlay learns a smooth edit transition after the trigger rather than a hard cut, consistent with the transition frames used during data synthesis (see Sec. 3.1, [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

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

114 extracted references · 11 linked inside Pith

  1. [1]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    One-step Diffusion with Distribution Matching Distillation , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  2. [2]

    Advances in Neural Information Processing Systems , pages =

    Improved Distribution Matching Distillation for Fast Image Synthesis , author =. Advances in Neural Information Processing Systems , pages =

  3. [3]

    International Conference on Learning Representations , pages =

    Pyramidal Flow Matching for Efficient Video Generative Modeling , author =. International Conference on Learning Representations , pages =

  4. [4]

    Advances in Neural Information Processing Systems , pages =

    Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion , author =. Advances in Neural Information Processing Systems , pages =

  5. [5]

    Advances in Neural Information Processing Systems , pages =

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion , author =. Advances in Neural Information Processing Systems , pages =

  6. [6]

    Forty-third International Conference on Machine Learning , year =

    Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation , author =. Forty-third International Conference on Machine Learning , year =

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    From Slow Bidirectional to Fast Autoregressive Video Diffusion Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  8. [8]

    and Zhang, Hao and Stoica, Ion , title =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph E. and Zhang, Hao and Stoica, Ion , title =. Proceedings of the 29th Symposium on Operating Systems Principles , year =

  9. [9]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  10. [10]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Ego4d: Around the world in 3,000 hours of egocentric video , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  11. [11]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Generic event boundary detection: A benchmark for event segmentation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  12. [12]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Temporal action localization in the deep learning era: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    2023

  13. [13]

    EasyV2V: A High-quality Instruction-based Video Editing Framework , author=

  14. [14]

    IEEE Transactions on Visualization and Computer Graphics , volume=

    Towards Pervasive Augmented Reality: Context-Awareness in Augmented Reality , author=. IEEE Transactions on Visualization and Computer Graphics , volume=. 2017 , publisher=

    2017

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Self forcing: Bridging the train-test gap in autoregressive video diffusion , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Yaron Lipman and Ricky T. Q. Chen and Heli Ben. Flow Matching for Generative Modeling , booktitle =

  17. [17]

    William Peebles and Saining Xie , title =

  18. [18]

    Zhenghao Zhang and Zuozhuo Dai and Long Qin and Weizhi Wang , year =

  19. [19]

    Shoubin Yu and Difan Liu and Ziqiao Ma and Yicong Hong and Yang Zhou and Hao Tan and Joyce Chai and Mohit Bansal , year =

  20. [20]

    Qin, Bosheng and Li, Juncheng and Tang, Siliang and Chua, Tat-Seng and Zhuang, Yueting , booktitle =

  21. [21]

    Ku, Max and Wei, Cong and Ren, Weiming and Yang, Harry and Chen, Wenhu , journal =

  22. [22]

    arXiv , year =

    Zero-Shot Video Editing Using Off-The-Shelf Image Diffusion Models , author =. arXiv , year =

  23. [23]

    Singer, Uriel and Zohar, Amit and Kirstain, Yuval and Sheynin, Shelly and Polyak, Adam and Parikh, Devi and Taigman, Yaniv , journal =

  24. [24]

    Chi Zhang and Chengjian Feng and Feng Yan and Qiming Zhang and Mingjin Zhang and Yujie Zhong and Jing Zhang and Lin Ma , year =

  25. [25]

    Wu, Yuhui and Chen, Liyi and Li, Ruibin and Wang, Shihao and Xie, Chenxi and Zhang, Lei , journal =

  26. [26]

    Zixuan Ye and Xuanhua He and Quande Liu and Qiulin Wang and Xintao Wang and Pengfei Wan and Di Zhang and Kun Gai and Qifeng Chen and Wenhan Luo , year =

  27. [27]

    2025 , archiveprefix =

    Movie Gen: A Cast of Media Foundation Models , author =. 2025 , archiveprefix =

    2025

  28. [28]

    Black Forest Labs and Stephen Batifol and Andreas Blattmann and Frederic Boesel and Saksham Consul and Cyril Diagne and Tim Dockhorn and Jack English and Zion English and Patrick Esser and Sumith Kulal and Kyle Lacey and Yam Levi and Cheng Li and Dominik Lorenz and Jonas Müller and Dustin Podell and Robin Rombach and Harry Saini and Axel Sauer and Luke Sm...

  29. [29]

    OpenAI Team , year =

  30. [30]

    2024 , journal =

    Consistent Video-to-Video Transfer Using Synthetic Dataset , author =. 2024 , journal =

    2024

  31. [31]

    Jiahao Hu and Tianxiong Zhong and Xuebo Wang and Boyuan Jiang and Xingye Tian and Fei Yang and Pengfei Wan and Di Zhang , year =

  32. [32]

    Jaehong Yoon and Shoubin Yu and Mohit Bansal , year =

  33. [33]

    Yuwei Fang and Willi Menapace and Aliaksandr Siarohin and Tsai-Shien Chen and Kuan-Chien Wang and Ivan Skorokhodov and Graham Neubig and Sergey Tulyakov , year =

  34. [34]

    Gang Cheng and Xin Gao and Li Hu and Siqi Hu and Mingyang Huang and Chaonan Ji and Ju Li and Dechao Meng and Jinwei Qi and Penchong Qiao and Zhen Shen and Yafei Song and Ke Sun and Linrui Tian and Feng Wang and Guangyuan Wang and Qi Wang and Zhongjian Wang and Jiayu Xiao and Sheng Xu and Bang Zhang and Peng Zhang and Xindi Zhang and Zhe Zhang and Jingren ...

  35. [35]

    Jinshu Chen and Xinghui Li and Xu Bai and Tianxiang Ma and Pengze Zhang and Zhuowei Chen and Gen Li and Lijie Liu and Songtao Zhao and Bingchuan Li and Qian He , year =

  36. [36]

    Xuan Ju and Tianyu Wang and Yuqian Zhou and He Zhang and Qing Liu and Nanxuan Zhao and Zhifei Zhang and Yijun Li and Yuanhao Cai and Shaoteng Liu and Daniil Pakhomov and Zhe Lin and Soo Ye Kim and Qiang Xu , year =

  37. [37]

    Chong Mou and Qichao Sun and Yanze Wu and Pengze Zhang and Xinghui Li and Fulong Ye and Songtao Zhao and Qian He , year =

  38. [38]

    CVPR , year =

    Generative Video Propagation , author =. CVPR , year =

  39. [39]

    Ahmet Berke Gokmen and Yigit Ekin and Bahri Batuhan Bilecen and Aysegul Dundar , year =

  40. [40]

    Lucy Edit: Open-Weight Text-Guided Video Editing , author =

  41. [41]

    Chenyang Qi and Xiaodong Cun and Yong Zhang and Chenyang Lei and Xintao Wang and Ying Shan and Qifeng Chen , year =

  42. [42]

    CVPR , year =

    Liu, Shaoteng and Zhang, Yuechen and Li, Wenbo and Lin, Zhe and Jia, Jiaya , title =. CVPR , year =

  43. [43]

    Bojia Zi and Penghui Ruan and Marco Chen and Xianbiao Qi and Shaozhe Hao and Shihao Zhao and Youze Huang and Bin Liang and Rong Xiao and Kam-Fai Wong , booktitle =

  44. [44]

    Ju, Xuan and Zeng, Ailing and Bian, Yuxuan and Liu, Shaoteng and Xu, Qiang , journal =

  45. [45]

    Xuan Ju and Weicai Ye and Quande Liu and Qiulin Wang and Xintao Wang and Pengfei Wan and Di Zhang and Kun Gai and Qiang Xu , year =

  46. [46]

    Cai, Qi and Chen, Jingwen and Chen, Yang and Li, Yehao and Long, Fuchen and Pan, Yingwei and Qiu, Zhaofan and Zhang, Yiheng and Gao, Fengbin and Xu, Peihan and others , journal =

  47. [47]

    Chunting Zhou and Lili Yu and Arun Babu and Kushal Tirumala and Michihiro Yasunaga and Leonid Shamis and Jacob Kahn and Xuezhe Ma and Luke Zettlemoyer and Omer Levy , year =

  48. [48]

    2024 , url =

    Video generation models as world simulators , author =. 2024 , url =

    2024

  49. [49]

    Kong, Weijie and Tian, Qi and Zhang, Zijian and Min, Rox and Dai, Zuozhuo and Zhou, Jin and Xiong, Jiangfeng and Li, Xin and Wu, Bo and Zhang, Jianwei and others , journal =

  50. [50]

    GitHub repository , howpublished =

    Mochi 1 , author =. GitHub repository , howpublished =. 2024 , publisher =

    2024

  51. [51]

    Yang, Zhuoyi and Teng, Jiayan and Zheng, Wendi and Ding, Ming and Huang, Shiyu and Xu, Jiazheng and Yang, Yuanming and Hong, Wenyi and Zhang, Xiaohan and Feng, Guanyu and others , journal =

  52. [52]

    Team Wan and Ang Wang and Baole Ai and Bin Wen and Chaojie Mao and Chen-Wei Xie and Di Chen and Feiwu Yu and Haiming Zhao and Jianxiao Yang and Jianyuan Zeng and Jiayu Wang and Jingfeng Zhang and Jingren Zhou and Jinkai Wang and Jixuan Chen and Kai Zhu and Kang Zhao and Keyu Yan and Lianghua Huang and Mengyang Feng and Ningyi Zhang and Pandeng Li and Ping...

  53. [53]

    Yu Gao and Lixue Gong and Qiushan Guo and Xiaoxia Hou and Zhichao Lai and Fanshi Li and Liang Li and Xiaochen Lian and Chao Liao and Liyang Liu and Wei Liu and Yichun Shi and Shiqi Sun and Yu Tian and Zhi Tian and Peng Wang and Rui Wang and Xuanda Wang and Xun Wang and Ye Wang and Guofeng Wu and Jie Wu and Xin Xia and Xuefeng Xiao and Zhonghua Zhai and Xi...

  54. [54]

    Team Seedream and : and Yunpeng Chen and Yu Gao and Lixue Gong and Meng Guo and Qiushan Guo and Zhiyao Guo and Xiaoxia Hou and Weilin Huang and Yixuan Huang and Xiaowen Jian and Huafeng Kuang and Zhichao Lai and Fanshi Li and Liang Li and Xiaochen Lian and Chao Liao and Liyang Liu and Wei Liu and Yanzuo Lu and Zhengxiong Luo and Tongtong Ou and Guang Shi ...

  55. [55]

    CVPR , year =

    Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn , title =. CVPR , year =

  56. [56]

    2024 , howpublished =

    Black Forest Labs , title =. 2024 , howpublished =

    2024

  57. [57]

    Imagen-Team-Google , year =

  58. [58]

    ICCV , year =

    Adding Conditional Control to Text-to-Image Diffusion Models , author =. ICCV , year =

  59. [59]

    Jiang, Zeyinzi and Han, Zhen and Mao, Chaojie and Zhang, Jingfeng and Pan, Yulin and Liu, Yu , booktitle =

  60. [60]

    Kulikov, Vladimir and Kleiner, Matan and Huberman-Spiegelglas, Inbar and Michaeli, Tomer , journal =

  61. [61]

    TMLR , year =

    Paul Couairon and Cl. TMLR , year =

  62. [62]

    Brooks, Tim and Holynski, Aleksander and Efros, Alexei A , journal =

  63. [63]

    arXiv , year =

    Prompt-to-prompt image editing with cross attention control , author =. arXiv , year =

  64. [64]

    ICCV , year =

    Cao, Mingdeng and Wang, Xintao and Qi, Zhongang and Shan, Ying and Qie, Xiaohu and Zheng, Yinqiang , title =. ICCV , year =

  65. [65]

    Kai Zhang and Lingbo Mo and Wenhu Chen and Huan Sun and Yu Su , booktitle =

  66. [66]

    Haozhe Zhao and Xiaojian Ma and Liang Chen and Shuzheng Si and Rujie Wu and Kaikai An and Peiyu Yu and Minjia Zhang and Qing Li and Baobao Chang , booktitle =

  67. [67]

    Hui, Mude and Yang, Siwei and Zhao, Bingchen and Shi, Yichun and Wang, Heng and Wang, Peng and Zhou, Yuyin and Xie, Cihang , journal =

  68. [68]

    CVPR , year =

    Emu Edit: Precise Image Editing via Recognition and Generation Tasks , author =. CVPR , year =

  69. [69]

    Wei, Cong and Xiong, Zheyang and Ren, Weiming and Du, Xinrun and Zhang, Ge and Chen, Wenhu , booktitle =

  70. [70]

    Qifan Yu and Wei Chow and Zhongqi Yue and Kaihang Pan and Yang Wu and Xiaoyang Wan and Juncheng Li and Siliang Tang and Hanwang Zhang and Yueting Zhuang , booktitle =

  71. [71]

    Shitao Xiao and Yueze Wang and Junjie Zhou and Huaying Yuan and Xingrun Xing and Ruiran Yan and Chaofan Li and Shuting Wang and Tiejun Huang and Zheng Liu , booktitle =

  72. [72]

    Zechuan Zhang and Ji Xie and Yu Lu and Zongxin Yang and Yi Yang , journal =

  73. [73]

    Bin Xia and Bohao Peng and Yuechen Zhang and Junjia Huang and Jiyang Liu and Jingyao Li and Haoru Tan and Sitong Wu and Chengyao Wang and Yitong Wang and Xinglong Wu and Bei Yu and Jiaya Jia , journal =

  74. [74]

    2508.02324 , archiveprefix =

    Wu, Chenfei and Li, Jiahao and Zhou, Jingren and Lin, Junyang and Gao, Kaiyuan and Yan, Kun and Yin, Sheng-ming and Bai, Shuai and Xu, Xiao and Chen, Yilei and Chen, Yuxiang and Tang, Zecheng and Zhang, Zekai and Wang, Zhengyi and Yang, An and Yu, Bowen and Cheng, Chen and Liu, Dayiheng and Li, Deqing and Zhang, Hang and others , year =. 2508.02324 , arch...

    Pith/arXiv arXiv

  75. [75]

    Ge, Yuying and Zhao, Sijie and Li, Chen and Ge, Yixiao and Shan, Ying , journal =

  76. [76]

    Yang, Ling and Zeng, Bohan and Liu, Jiaming and Li, Hong and Xu, Minghao and Zhang, Wentao and Yan, Shuicheng , journal =

  77. [77]

    Emerging Properties in Unified Multimodal Pretraining: The

    Chaorui Deng and Deyao Zhu and Kunchang Li and Chenhui Gou and Feng Li , journal =. Emerging Properties in Unified Multimodal Pretraining: The

  78. [78]

    arXiv , year =

    Reconstruction Alignment Improves Unified Multimodal Models , author =. arXiv , year =

  79. [79]

    TOG , year =

    Magic fixup: Streamlining photo editing by watching dynamic videos , author =. TOG , year =

  80. [80]

    arXiv , year =

    Recognize Anything: A Strong Image Tagging Model , author =. arXiv , year =

Showing first 80 references.

This paper was first reviewed by grok-4.5 on July 31, 2026.