Pith. sign in

REVIEW 3 major objections 4 minor 10 cited by

Bidirectional first-person to third-person video translation now works both ways with in-context learning and no camera poses.

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

2026-08-03 19:50 UTC pith:2M443HL6

load-bearing objection A useful task formulation and dataset, but the ego-to-exo evaluation leaks the target video through the reference image, so the headline SOTA claim is not yet established. the 3 major comments →

arxiv 2511.22098 v2 pith:2M443HL6 submitted 2025-11-27 cs.CV

WorldWander: Bridging Egocentric and Exocentric Worlds in Video Generation

Quanjian Song , Yiren Song , Kelly Peng , Yuan Gao , Mike Zheng Shou This is my paper
classification cs.CV
keywords egocentric-exocentric video translationin-context learningvideo diffusion transformerperspective synchronizationcharacter consistencyvideo-to-video translationEgoExo-8Kworld models
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.

This paper tries to establish that a single video diffusion transformer can translate between first-person (egocentric) and third-person (exocentric) perspectives in both directions, using only paired video–reference triplets and no camera geometry. Such a capability would let filmmakers, game developers, and embodied agents switch viewpoints in generated worlds while preserving character identity and motion synchronization. The authors argue that in-context token-level attention plus shared temporal position encoding is enough to learn the cross-view correspondence, and they support this with a new dataset of synchronized egocentric–exocentric triplets. A sympathetic reader would care because prior work only does one direction or requires camera parameters, and this opens a new bidirectional task with a public benchmark.

Core claim

The paper's central claim is that egocentric–exocentric video translation can be formulated as a bidirectional in-context learning task and solved by fine-tuning a pretrained video diffusion transformer with two components: In-Context Perspective Alignment, which concatenates the noise-free conditional latents and the noisy target latents along the token dimension so the model's self-attention sees both perspectives together, and Collaborative Position Encoding, which gives the conditional and target tokens the same temporal position embeddings rather than a single uniform sequence. On the curated EgoExo-8K dataset of roughly 4,000 synthetic and 4,000 real-world triplets, the authors report

What carries the argument

The load-bearing mechanism is collaborative attention over token-concatenated latents: conditional video and reference latents are kept noise-free, the target latent is noised by rectified flow, and all are concatenated along the token dimension so the transformer's self-attention operates jointly over both perspectives. Collaborative Position Encoding then re-applies the same temporal rotary embeddings to each latent's tokens, preserving their frame-level correspondence. Together they let the model learn cross-view mapping without any extra network or camera input.

Load-bearing premise

The real-world training data only varies clothing, not distinct identities, so the claimed generalization to unseen characters in egocentric-to-exocentric translation is not demonstrated.

What would settle it

Train or evaluate the ego-to-exo model on real-world triplets of many different, unseen people (varying body shape, skin tone, hairstyle) and measure character consistency via face or identity retrieval; if identity-similarity drops sharply compared with the clothing-variation setting, the generalization claim is falsified.

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

If this is right

  • Mutual translation: the same framework handles both egocentric-to-exocentric and exocentric-to-egocentric generation, so a single model family can switch a viewer between first- and third-person views of the same scene.
  • No camera pose requirement: because the mapping is learned from triplet data, applications do not need calibrated cameras or pose estimation at inference time.
  • Benchmark resource: EgoExo-8K provides synchronized ego-exo triplets with reference images, giving the research community a common testbed for this new task.
  • Character consistency via reference image: ego-to-exo generation can be conditioned on a back-view character reference, so the synthesized third-person view keeps the same identity across frames.
  • Efficient adaptation: fine-tuning only low-rank adapters on a pretrained large video model is sufficient to acquire the cross-view behavior, lowering the compute barrier to entry.

Where Pith is reading between the lines

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

  • If the method generalizes beyond the dataset's clothing-variation scenarios, this training recipe could be applied to egocentric video from wearables to synthesize third-person views of the wearer, effectively turning any first-person recording into a character-centric world model.
  • The token-concatenation trick is a generic way to inject a second video stream into a transformer-based generator; it could transfer to other paired-video tasks such as depth-to-video, pose-to-video, or multi-view synthesis, where the two streams share a temporal axis.
  • A testable extension would be to remove the reference image during ego-to-exo inference and measure how much of character consistency relies on it, or to add noise to the conditional stream to see whether noise-free conditioning is essential.
  • The failure cases the authors show—physical property violations and inadequate personalization—suggest the next bottleneck is not cross-view alignment but world physics and identity modeling, which may need the model to reason about body geometry rather than just visual correspondences.

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 / 4 minor

Summary. The paper introduces WorldWander, an in-context learning framework for bidirectional egocentric–exocentric video translation. The method builds on the Wan2.2-5B video diffusion transformer and fine-tunes it with LoRA, adding two technical components: In-Context Perspective Alignment (shared latent space, distinct noise levels, and collaborative attention over token-concatenated conditional/target latents) and Collaborative Position Encoding (separate temporal rotary embeddings for conditional and target latents). The authors also curate EgoExo-8K, a dataset of roughly 4,000 synchronized triplet videos from GTA-5 and real-world dual-camera captures, and report automatic and user-study comparisons against AnyV2V, I2VEdit, TrajectoryCrafter, and ReCamMaster. The central claims are that WorldWander achieves superior perspective synchronization, character consistency, and generalization, and that it constitutes a new benchmark for egocentric–exocentric video translation. The supplementary material includes implementation details, additional comparisons, and an explicit limitation statement about limited real-world identity diversity.

Significance. If the claims are validated, the paper would make a useful contribution: it defines a new task formulation, provides a first large-scale synchronized dataset, and proposes a geometry-free in-context learning approach that avoids camera-pose dependence. The code release and the mostly clean training setup are also strengths; the method is trained directly on synchronized triplets and does not hide fitted constants in the evaluation. However, the headline SOTA and generalization claims are currently not established because of an evaluation-protocol leak and a lack of comparison with the most task-relevant prior methods. The core idea remains plausible, so the issues are fixable within the scope of a revision rather than requiring rejection.

major comments (3)
  1. [Sec. 4.5, Table 1] For the real-world subset, the reference image is "randomly sampled from the exocentric video." In the ego→exo task, that exocentric video is exactly the target video the model must generate. The input therefore contains a free ground-truth frame carrying the subject's appearance, clothing, scene, and lighting. The reported CLIP-I, FVD, LPIPS/SSIM, and user-rated character-consistency gains for real-world ego→exo may reflect copying that reference rather than learning a cross-view transformation. Please re-run the real-world ego→exo evaluation with a reference image that is not drawn from the target video (e.g., a separately captured character reference or a reference from a different synchronized recording), and report the same metrics.
  2. [Supp. Sec. 9, Sec. 5.2] The Limitation section concedes that "our real-world dataset has limited diversity in human subjects. Most of the variation comes from changes in clothing, rather than from distinct individuals." Since the test triplets are held-out clips from the same scenes/identities used in training, the claimed generalization across characters in ego-to-exo translation is not established. Please add an identity-split evaluation with unseen subjects (and ideally unseen environments) for both synthetic and real-world subsets, or soften the generalization claim.
  3. [Sec. 5.1, Sec. 2.3] The claimed SOTA is established only against generic video-to-video and perspective re-orientation baselines. The related work cites task-specific methods for exo→ego (Liu et al. 2024; EgoExoGen 2025) and ego→exo (Luo et al. 2024), but none is included in Table 1. Since these methods address the same or closely related tasks, the state-of-the-art claim is not supported. Please add these methods as baselines under the same protocol, or justify their exclusion explicitly.
minor comments (4)
  1. [Sec. 4.1] The two bullets appear to swap the subscripts: "exocentric video V_ego" and "egocentric video V_exo" are inconsistent with the definitions of V_ego and V_exo given earlier in the same paragraph. Please fix the notation.
  2. [Figure 3] The convergence comparison shows loss curves without confidence intervals or multiple seeds. A single run may not support the claim that collaborative attention converges faster than channel-wise concatenation.
  3. [Supp. Sec. 8] The dataset description says "approximately 4,000 triplets" in the main text and "over 400 paired video sequences" for the synthetic subset. Please clarify the exact counts and note the overlap between sequences and triplets.
  4. [Sec. 5.2] The phrase "state-of-the-art" is used for comparisons against the four selected baselines; consider using "best among compared methods" to avoid implying a global SOTA, especially in light of the omitted task-specific baselines.

Circularity Check

0 steps flagged

No significant circularity: the method is supervised fine-tuning with explicit ablations; the real-world reference-from-target sampling is an evaluation leak (soundness), not a derivation-circular step, and self-citations are not load-bearing.

full rationale

WorldWander's claim is empirical: it fine-tunes Wan2.2-5B with LoRA on synchronized ego/exo triplets, using two proposed components (collaborative attention and collaborative position encoding). There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via citation. The ablations in Table 2 compare the proposed components against channel-concatenation and uniform position encoding on held-out synthetic triplets, so the core architectural claims have independent empirical support. The most notable issue is in Sec. 4.5: 'The reference image is then randomly sampled from the exocentric video,' while the ego→exo condition includes this reference and the target is the same exocentric video. This supplies a frame of the target as input and can inflate real-world CLIP-I and character-consistency numbers. That is a genuine evaluation leak and should be scored under soundness/benchmark design, but it is not a derivation-circular step: the model must still generate the full target video, and the exo→ego direction plus the synthetic ego→exo setting are not affected in the same way. The supplementary limitation (Sec. 9) explicitly concedes that real-world subject diversity is low and that generalization across characters 'remains restricted,' which weakens the abstract's generalization claim but is an honest data limitation rather than circularity. The paper contains many self-citations, but none is load-bearing: the backbone (Wan2.2) and all baselines are external, and the task formulation rests on the newly collected dataset rather than on a self-citation chain. Overall, no significant circularity in the derivation chain; score 2 reflects the mild self-referential elements (self-built benchmark and non-load-bearing self-citations) without elevating the evaluation leak to circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No free parameters are fitted to data beyond standard training hyperparameters. The central claim depends on domain assumptions about learnability from paired videos, backbone transfer, and the reference-image sampling choice. No new physical or theoretical entities are introduced; EgoExo-8K is a dataset, not a postulated entity.

axioms (4)
  • domain assumption Synchronized ego-exo video triplets are sufficient to learn cross-view correspondence without camera poses.
    Section 4.1 explicitly discards camera pose dependency; the entire method rests on paired-video correspondence being learnable from data alone.
  • domain assumption Pretrained Wan2.2-5B provides priors that transfer to egocentric-exocentric translation via LoRA fine-tuning.
    Section 4.4; if the backbone priors do not transfer, fine-tuning on roughly 8,000 clips would be insufficient for the claimed generalization.
  • ad hoc to paper The reference image can be sampled from the target exocentric video without materially leaking the target into the input.
    Section 4.5 samples I_ref from V_exo, which is the target for ego-to-exo generation. This is a paper-specific choice that can leak a ground-truth frame into the test input.
  • standard math Rectified-flow / flow-matching loss with clean conditional latents and noised target latents is a valid learning objective for this translation task.
    Section 3, Eq. (2); standard diffusion/flow-matching formulation used as the training objective.

pith-pipeline@v1.3.0-alltime-deepseek · 14251 in / 13356 out tokens · 115063 ms · 2026-08-03T19:50:15.219217+00:00 · methodology

0 comments
read the original abstract

Recent advances in video world models enable interactive environments with free navigation, making translation between first-person (egocentric) and third-person (exocentric) perspectives increasingly important. However, existing studies focus on unidirectional exocentric-to-egocentric translation, overlooking reference-guided exocentric perspective synthesis. This capability is crucial for gaming and embodied AI applications. Motivated by this, we present WorldWander, an in-context learning framework tailored for translating between egocentric and exocentric worlds in video generation. Building upon advanced video diffusion transformers, WorldWander integrates (i) In-Context Perspective Alignment and (ii) Collaborative Position Encoding to model cross-view synchronization and character consistency. To support our task, we curate EgoExo-8K, a dynamic and scene-rich dataset containing synchronized egocentric-exocentric triplets from both synthetic and real-world scenarios. Experiments demonstrate that WorldWander achieves superior perspective synchronization, character consistency, and generalization, setting a new benchmark for egocentric-exocentric video translation.

Figures

Figures reproduced from arXiv: 2511.22098 by Kelly Peng, Mike Zheng Shou, Quanjian Song, Yiren Song, Yuan Gao.

Figure 1
Figure 1. Figure 1: Gallery of WorldWander. It bridges the egocentric and exocentric worlds in video generation, enabling immersive exploration. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall pipeline of WorldWander. The backbone Wan2.2-5B [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of fine-tuning loss across different ap [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Showcase of our curated EgoExo-8K. It features diverse [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: User study of different methods on both the exocentric [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of different methods on both exocentric-to-egocentric and egocentric-to-exocentric video translations. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative ablation about two components: [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional examples of our curated EgoExo-8K. It features diverse synthetic and real-world indoor and outdoor scenarios. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Failure cases of our WorldWander. It includes (a) [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional qualitative comparison across approaches on the exocentric-egocentric video translations task. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional qualitative comparison across approaches on the exocentric-egocentric video translations task. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional gallery of our WorldWander. It supports egocentric-to-exocentric translation for out-of-domain characters. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 10 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Mirror Learning

    cs.LG 2026-07 conditional novelty 7.0

    Fine-tuning a video diffusion model to perform cross-view perspective transfer, then labeling the generated first-person videos with an inverse dynamics model, yields behavior-cloning data that improves driving policies.

  2. TryOnCrafter: Unleashing Camera Trajectories for Realistic Video Virtual Try-on via a Renderable 4D Try-on Proxy

    cs.CV 2026-06 unverdicted novelty 7.0

    TryOnCrafter is the first DiT-based framework for camera-controllable video virtual try-on via a renderable 4D try-on proxy distilled from 2D priors into 3DGS avatar animated with SMPL-X.

  3. Soap2Soap: Long Cinematic Video Remaking via Multi-Agent Collaboration

    cs.CV 2026-05 unverdicted novelty 7.0

    Soap2Soap uses a multi-agent system with dual-bridge consistency via JSON screenplays and visual anchors plus batch keyframe generation to achieve better long-term consistency in cinematic video remaking than commercial APIs.

  4. StreamingEffect: Real-Time Human-Centric Video Effect Generation

    cs.CV 2026-05 unverdicted novelty 7.0

    StreamingEffect enables real-time 720p human-centric video effect generation on one GPU via teacher-student distillation, keyframe control, and a new 130K video dataset.

  5. TryOnCrafter: Unleashing Camera Trajectories for Realistic Video Virtual Try-on via a Renderable 4D Try-on Proxy

    cs.CV 2026-06 conditional novelty 6.0

    A 4D try-on proxy (3DGS avatar + SMPL-X + background points) anchors a DiT so virtual try-on videos can follow arbitrary camera trajectories with consistent garments and scene structure.

  6. PAI-Studio: Cinematic Video Background Replacement with Camera-Aware Motion

    cs.CV 2026-05 unverdicted novelty 6.0

    PAI-Studio reformulates cinematic background replacement as in-context conditional generation inside a Diffusion Transformer with bidirectional attention, trained on a new 30K film-sourced dataset, and reports better ...

  7. VISTA: Triplet-Supervised Video Style Transfer with Diffusion Transformers

    cs.CV 2026-05 unverdicted novelty 6.0

    VISTA introduces a new synthetic triplet dataset and diffusion-transformer framework with style adapter that jointly models style, content, and motion to achieve state-of-the-art video style transfer.

  8. FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization

    cs.CV 2026-05 unverdicted novelty 6.0

    FashionChameleon achieves interactive multi-garment video customization at 23.8 FPS via in-context teacher models, streaming distillation, and training-free KV cache rescheduling while using only single-garment data.

  9. FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization

    cs.CV 2026-05 unverdicted novelty 6.0

    FashionChameleon achieves interactive multi-garment video customization in real time by training a teacher model with in-context learning on single-garment pairs, applying streaming distillation, and using training-fr...

  10. OmniHumanoid: Streaming Cross-Embodiment Video Generation with Paired-Free Adaptation

    cs.CV 2026-05 unverdicted novelty 6.0

    OmniHumanoid factorizes transferable motion learning from embodiment-specific adaptation to enable scalable cross-embodiment video generation without paired data for new humanoids.

Reference graph

Works this paper leans on

63 extracted references · 27 linked inside Pith · cited by 8 Pith papers

  1. [1]

    Recammaster: Camera-controlled generative ren- dering 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 ren- dering from a single video. InInt. Conf. Comput. Vis., 2025. 1, 2, 6, 12

    2025

  2. [2]

    Navigation world models

    Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. InIEEE Conf. Com- put. Vis. Pattern Recog., 2025. 1

    2025

  3. [3]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 2

    Pith/arXiv arXiv 2023

  4. [4]

    Align your latents: High-resolution video synthesis with la- tent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models. InIEEE Conf. Comput. Vis. Pattern Recog., 2023. 2

    2023

  5. [5]

    Transanimate: Taming layer diffusion to generate rgba video.arXiv preprint arXiv:2503.17934, 2025

    Xuewei Chen, Zhimin Chen, and Yiren Song. Transanimate: Taming layer diffusion to generate rgba video.arXiv preprint arXiv:2503.17934, 2025. 1

    Pith/arXiv arXiv 2025

  6. [6]

    Relationadapter: Learning and transferring vi- sual relation with diffusion transformers.arXiv preprint arXiv:2506.02528, 2025

    Yan Gong, Yiren Song, Yicheng Li, Chenglin Li, and Yin Zhang. Relationadapter: Learning and transferring vi- sual relation with diffusion transformers.arXiv preprint arXiv:2506.02528, 2025. 2

    Pith/arXiv arXiv 2025

  7. [7]

    Animatediff: Animate your personalized text-to- image diffusion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to- image diffusion models without specific tuning. InInt. Conf. Learn. Represent., 2024. 2

    2024

  8. [8]

    Cameractrl: En- abling camera control for video diffusion models

    Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: En- abling camera control for video diffusion models. InInt. Conf. Learn. Represent., 2025. 1, 2, 4

    2025

  9. [9]

    Matrix-game 2.0: An open-source, real-time, and streaming interactive world model.arXiv preprint arXiv:2508.13009, 2025

    Xianglong He, Chunli Peng, Zexiang Liu, Boyang Wang, Yifan Zhang, Qi Cui, Fei Kang, Biao Jiang, Mengyin An, Yangyang Ren, et al. Matrix-game 2.0: An open-source, real-time, and streaming interactive world model.arXiv preprint arXiv:2508.13009, 2025. 1, 2

    Pith/arXiv arXiv 2025

  10. [10]

    Training-free camera control for video generation

    Chen Hou, Guoqiang Wei, Yan Zeng, and Zhibo Chen. Training-free camera control for video generation. InInt. Conf. Learn. Represent., 2024. 2

    2024

  11. [11]

    Lora: Low-rank adaptation of large language models.Int

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Int. Conf. Learn. Represent., 2022. 5

    2022

  12. [12]

    Vbench: Comprehensive bench- mark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive bench- mark suite for video generative models. InIEEE Conf. Com- put. Vis. Pattern Recog., 2024. 6

    2024

  13. [13]

    Reangle- a-video: 4d video generation as video-to-video translation

    Hyeonho Jeong, Suhyeon Lee, and Jong Chul Ye. Reangle- a-video: 4d video generation as video-to-video translation. InInt. Conf. Comput. Vis., 2025. 1, 2

    2025

  14. [14]

    Personalized vision via visual in-context learning.arXiv preprint arXiv:2509.25172, 2025

    Yuxin Jiang, Yuchao Gu, Yiren Song, Ivor Tsang, and Mike Zheng Shou. Personalized vision via visual in-context learning.arXiv preprint arXiv:2509.25172, 2025. 2

    arXiv 2025

  15. [15]

    Vace: All-in-one video creation and editing.arXiv preprint arXiv:2503.07598, 2025

    Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu. Vace: All-in-one video creation and editing.arXiv preprint arXiv:2503.07598, 2025. 3

    Pith/arXiv arXiv 2025

  16. [16]

    Text2video-zero: Text- to-image diffusion models are zero-shot video generators

    Levon Khachatryan, Andranik Movsisyan, Vahram Tade- vosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text- to-image diffusion models are zero-shot video generators. In Int. Conf. Comput. Vis., 2023. 2

    2023

  17. [17]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    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, 2024. 1, 2

    Pith/arXiv arXiv 2024

  18. [18]

    Anyv2v: A tuning-free framework for any video-to- video editing tasks.arXiv preprint arXiv:2403.14468, 2024

    Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to- video editing tasks.arXiv preprint arXiv:2403.14468, 2024. 3, 6, 7, 12

    Pith/arXiv arXiv 2024

  19. [19]

    Hunyuan-gamecraft: High-dynamic interactive game video generation with hybrid history condition.arXiv preprint arXiv:2506.17201, 2025

    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, 2025. 1, 2, 3

    Pith/arXiv arXiv 2025

  20. [20]

    Flow matching for generative mod- eling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling. InInt. Conf. Learn. Represent., 2023. 1, 2, 3, 4

    2023

  21. [21]

    Stylecrafter: Taming artistic video diffusion with reference-augmented adapter learning.ACM Trans

    Gongye Liu, Menghan Xia, Yong Zhang, Haoxin Chen, Jinbo Xing, Yibo Wang, Xintao Wang, Ying Shan, and Yu- jiu Yang. Stylecrafter: Taming artistic video diffusion with reference-augmented adapter learning.ACM Trans. Graph.,

  22. [22]

    Exocentric-to-egocentric video genera- tion

    Jia-Wei Liu, Weijia Mao, Zhongcong Xu, Jussi Keppo, and Mike Zheng Shou. Exocentric-to-egocentric video genera- tion. InAdv. Neural Inform. Process. Syst., 2024. 2

    2024

  23. [23]

    Easytext: Controllable diffusion trans- former for multilingual text rendering.arXiv preprint arXiv:2505.24417, 2025

    Runnan Lu, Yuxuan Zhang, Jiaming Liu, Haofan Wang, and Yiren Song. Easytext: Controllable diffusion trans- former for multilingual text rendering.arXiv preprint arXiv:2505.24417, 2025. 2

    arXiv 2025

  24. [24]

    Intention-driven ego-to-exo video generation.arXiv preprint arXiv:2403.09194, 2024

    Hongchen Luo, Kai Zhu, Wei Zhai, and Yang Cao. Intention-driven ego-to-exo video generation.arXiv preprint arXiv:2403.09194, 2024. 2

    Pith/arXiv arXiv 2024

  25. [25]

    Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation

    Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. InSIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. 2

    2024

  26. [26]

    Follow-your-click: Open-domain regional image animation via motion prompts

    Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Leqi Shen, Chenyang Qi, Jixuan Ying, Chengfei Cai, Zhifeng Li, Heung-Yeung Shum, et al. Follow-your-click: Open-domain regional image animation via motion prompts. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 6018–6026, 2025

    2025

  27. [27]

    Follow-your-emoji-faster: To- wards efficient, fine-controllable, and expressive freestyle portrait animation.arXiv preprint arXiv:2509.16630, 2025

    Yue Ma, Zexuan Yan, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, et al. Follow-your-emoji-faster: To- wards efficient, fine-controllable, and expressive freestyle portrait animation.arXiv preprint arXiv:2509.16630, 2025. 2

    arXiv 2025

  28. [28]

    I2vedit: First-frame-guided video editing via image-to- video diffusion models

    Wenqi Ouyang, Yi Dong, Lei Yang, Jianlou Si, and Xingang Pan. I2vedit: First-frame-guided video editing via image-to- video diffusion models. InSIGGRAPH Asia, 2024. 3, 6, 7, 12

    2024

  29. [29]

    Fonts: Text rendering with typography and style controls.arXiv preprint arXiv:2412.00136, 2024

    Wenda Shi, Yiren Song, Dengming Zhang, Jiaming Liu, and Xingxing Zou. Fonts: Text rendering with typography and style controls.arXiv preprint arXiv:2412.00136, 2024. 2

    Pith/arXiv arXiv 2024

  30. [30]

    Wordcon: Word-level ty- pography control in scene text rendering.arXiv preprint arXiv:2506.21276, 2025

    Wenda Shi, Yiren Song, Zihan Rao, Dengming Zhang, Ji- aming Liu, and Xingxing Zou. Wordcon: Word-level ty- pography control in scene text rendering.arXiv preprint arXiv:2506.21276, 2025. 2

    Pith/arXiv arXiv 2025

  31. [31]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1, 2

    Pith/arXiv arXiv 2010

  32. [32]

    Univst: A unified framework for training-free localized video style transfer.arXiv preprint arXiv:2410.20084, 2024

    Quanjian Song, Mingbao Lin, Wengyi Zhan, Shuicheng Yan, Liujuan Cao, and Rongrong Ji. Univst: A unified framework for training-free localized video style transfer.arXiv preprint arXiv:2410.20084, 2024. 1, 2

    arXiv 2024

  33. [33]

    Lightmotion: A light and tuning-free method for simulating camera motion in video generation.arXiv preprint arXiv:2503.06508, 2025

    Quanjian Song, Zhihang Lin, Zhanpeng Zeng, Ziyue Zhang, Liujuan Cao, and Rongrong Ji. Lightmotion: A light and tuning-free method for simulating camera motion in video generation.arXiv preprint arXiv:2503.06508, 2025. 2

    Pith/arXiv arXiv 2025

  34. [34]

    Scenedecorator: Towards scene-oriented story generation with scene planning and scene consistency.arXiv preprint arXiv:2510.22994, 2025

    Quanjian Song, Donghao Zhou, Jingyu Lin, Fei Shen, Jiaze Wang, Xiaowei Hu, Cunjian Chen, and Pheng-Ann Heng. Scenedecorator: Towards scene-oriented story generation with scene planning and scene consistency.arXiv preprint arXiv:2510.22994, 2025. 1

    arXiv 2025

  35. [35]

    Score-based generative modeling through stochastic differential equa- tions.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions.arXiv preprint arXiv:2011.13456, 2020. 1

    Pith/arXiv arXiv 2011

  36. [36]

    Proces- spainter: Learn painting process from sequence data.arXiv preprint arXiv:2406.06062, 2024

    Yiren Song, Shijie Huang, Chen Yao, Xiaojun Ye, Hai Ci, Jiaming Liu, Yuxuan Zhang, and Mike Zheng Shou. Proces- spainter: Learn painting process from sequence data.arXiv preprint arXiv:2406.06062, 2024. 1

    Pith/arXiv arXiv 2024

  37. [37]

    Makeany- thing: Harnessing diffusion transformers for multi- domain procedural sequence generation.arXiv preprint arXiv:2502.01572, 2025

    Yiren Song, Cheng Liu, and Mike Zheng Shou. Makeany- thing: Harnessing diffusion transformers for multi- domain procedural sequence generation.arXiv preprint arXiv:2502.01572, 2025. 1

    Pith/arXiv arXiv 2025

  38. [38]

    Omniconsis- tency: Learning style-agnostic consistency from paired styl- ization data.arXiv preprint arXiv:2505.18445, 2025

    Yiren Song, Cheng Liu, and Mike Zheng Shou. Omniconsis- tency: Learning style-agnostic consistency from paired styl- ization data.arXiv preprint arXiv:2505.18445, 2025. 2

    Pith/arXiv arXiv 2025

  39. [39]

    To- wards accurate generative models of video: A new metric & challenges.arXiv preprint arXiv:1812.01717, 2018

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. To- wards accurate generative models of video: A new metric & challenges.arXiv preprint arXiv:1812.01717, 2018. 6

    Pith/arXiv arXiv 2018

  40. [40]

    Wan: Open and advanced large-scale video gen- erative models.arXiv preprint arXiv:2503.20314, 2025

    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 gen- erative models.arXiv preprint arXiv:2503.20314, 2025. 1, 2, 3, 5, 12

    Pith/arXiv arXiv 2025

  41. [41]

    Consistent video editing as flow-driven image-to-video generation.arXiv preprint arXiv:2506.07713, 2025

    Ge Wang, Songlin Fan, Hangxu Liu, Quanjian Song, Hewei Wang, and Jinfeng Xu. Consistent video editing as flow-driven image-to-video generation.arXiv preprint arXiv:2506.07713, 2025. 2

    Pith/arXiv arXiv 2025

  42. [42]

    Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023. 2

    Pith/arXiv arXiv 2023

  43. [43]

    Motionctrl: A unified and flexible motion controller for video generation

    Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. InSIGGRAPH, 2024. 1, 2, 4

    2024

  44. [44]

    Diffdecompose: Layer-wise de- composition of alpha-composited images via diffusion trans- formers.arXiv preprint arXiv:2505.21541, 2025

    Zitong Wang, Hang Zhao, Qianyu Zhou, Xuequan Lu, Xi- angtai Li, and Yiren Song. Diffdecompose: Layer-wise de- composition of alpha-composited images via diffusion trans- formers.arXiv preprint arXiv:2505.21541, 2025. 2

    Pith/arXiv arXiv 2025

  45. [45]

    Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025. 12

    Pith/arXiv arXiv 2025

  46. [46]

    Mo- tionbooth: Motion-aware customized text-to-video genera- tion

    Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Mo- tionbooth: Motion-aware customized text-to-video genera- tion. InAdv. Neural Inform. Process. Syst., 2024. 2

    2024

  47. [47]

    Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Int. Conf. Comput. Vis., 2023. 2

    2023

  48. [48]

    Egoexo- gen: Ego-centric video prediction by watching exo-centric videos

    Jilan Xu, Yifei Huang, Baoqi Pei, Junlin Hou, Qingqiu Li, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Egoexo- gen: Ego-centric video prediction by watching exo-centric videos. InInt. Conf. Learn. Represent., 2025. 2

    2025

  49. [49]

    Rerender a video: Zero-shot text-guided video-to-video translation

    Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. InSIGGRAPH Asia, 2023. 2

    2023

  50. [50]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024. 2

    Pith/arXiv arXiv 2024

  51. [51]

    Stylemaster: Stylize your video with artistic generation and translation

    Zixuan Ye, Huijuan Huang, Xintao Wang, Pengfei Wan, Di Zhang, and Wenhan Luo. Stylemaster: Stylize your video with artistic generation and translation. InIEEE Conf. Com- put. Vis. Pattern Recog., 2025. 2

    2025

  52. [52]

    Tra- jectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models

    Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Tra- jectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. InInt. Conf. Comput. Vis., 2025. 1, 2, 6, 7, 12

    2025

  53. [53]

    Recapture: Gener- ative video camera controls for user-provided videos using masked video fine-tuning

    David Junhao Zhang, Roni Paiss, Shiran Zada, Nikhil Kar- nad, David E Jacobs, Yael Pritch, Inbar Mosseri, Mike Zheng Shou, Neal Wadhwa, and Nataniel Ruiz. Recapture: Gener- ative video camera controls for user-provided videos using masked video fine-tuning. InCVPR, 2025. 1

    2025

  54. [54]

    Easycontrol: Adding efficient and flexible control for diffusion transformer.arXiv preprint arXiv:2503.07027, 2025

    Yuxuan Zhang, Yirui Yuan, Yiren Song, Haofan Wang, and Jiaming Liu. Easycontrol: Adding efficient and flexible control for diffusion transformer.arXiv preprint arXiv:2503.07027, 2025. 2

    Pith/arXiv arXiv 2025

  55. [55]

    Tora: Trajectory-oriented diffusion transformer for video genera- tion

    Zhenghao Zhang, Junchao Liao, Menghao Li, Zuozhuo Dai, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video genera- tion. InIEEE Conf. Comput. Vis. Pattern Recog., 2025. 2

    2025

  56. [56]

    Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024. 1, 2

    Pith/arXiv arXiv 2024

  57. [57]

    Magictailor: Component-controllable person- alization in text-to-image diffusion models.arXiv preprint arXiv:2410.13370, 2024

    Donghao Zhou, Jiancheng Huang, Jinbin Bai, Jiaze Wang, Hao Chen, Guangyong Chen, Xiaowei Hu, and Pheng- Ann Heng. Magictailor: Component-controllable person- alization in text-to-image diffusion models.arXiv preprint arXiv:2410.13370, 2024. 1 WorldWander: Bridging Egocentric and Exocentric Worlds in Video Generation Supplementary Material

    Pith/arXiv arXiv 2024

  58. [58]

    To ensure a fair comparison, all methods are first applied at their optimal resolutions and then resized to704×1024, with a fixed frame count ofF= 49

    Implementation Details about Baselines Recall that in the main experiments, we compare World- Wander with four representative baselines: AnyV2V [18], I2VEdit [28], TrajectoryCrafter [52], and RecamMaster [1]. To ensure a fair comparison, all methods are first applied at their optimal resolutions and then resized to704×1024, with a fixed frame count ofF= 4...

  59. [59]

    We present additional paired examples in Figure

    Details and Examples about EgoExo-8K Our EgoExo-8K dataset includes both synthetic and real- world scenarios, covering diverse environments such as deserts, fields, urban areas, and various indoor and outdoor settings. We present additional paired examples in Figure

  60. [60]

    During fine-tuning, these paired videos are segmented into 5-second clips (300 frames each). We describe the data collection process for these scenes in detail below: Synthetic Scenarios.We collect paired first-person and third-person videos from the GTA-5 game by setting up both perspectives and synchronizing the recordings. Both perspectives have a reso...

  61. [61]

    (ii) Our real-world dataset has limited diversity in human subjects

    Limitation and Future Work Although WorldWander shows strong performance on ego- centric–exocentric video translation, as illustrated in Fig- ure 9, it still exhibits some limitations: (i) Since World- Wander is fine-tuned over pretrained backbones, its perfor- mance ceiling is inherently constrained, especially in scenes involving significant physical ch...

  62. [62]

    Our WorldWander outperforms all baselines on both egocentric-to-exocentric and exocentric-to-egocentric video translation tasks, further highlighting its superior performance

    Additional Qualitative Comparisons In addition to the qualitative comparisons presented in the main paper, we provide further comparison results here, as illustrated in Figure 10 and Figure 11. Our WorldWander outperforms all baselines on both egocentric-to-exocentric and exocentric-to-egocentric video translation tasks, further highlighting its superior ...

  63. [63]

    As shown, our WorldWander supports egocentric- to-exocentric translation for out-of-domain characters, en- abling character-centric world exploration

    Additional Visual Gallery To demonstrate the generalization ability of our World- Wander, we provide additional visual gallery in Fig- ure 12. As shown, our WorldWander supports egocentric- to-exocentric translation for out-of-domain characters, en- abling character-centric world exploration. Egocentric-to-Exocentric I2VEditAnyV2VTracjtoryCrafterInput ReC...