Pith. sign in

REVIEW 3 major objections 5 minor 51 references

StreamHOI shows that long-horizon human-object interaction video generation can be preserved under bounded latency by specializing historical memory per transformer block.

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 10:33 UTC pith:YGDSSBTE

load-bearing objection StreamHOI earns a serious referee: the core idea is clean and the ablations are thoughtful, but the profiling-stability assumption and missing artifacts need to be pressed. the 3 major comments →

arxiv 2607.20174 v1 pith:YGDSSBTE submitted 2026-07-22 cs.CV cs.AI

StreamHOI: Interaction-aware Temporal Memory Adaptation for Streaming HOI Video Generation

Zejing Rao , Haoxian Zhang , Xiaoqiang Liu , Yiping Meng , Guoxin Zhang , Pengfei Wan , Fan Tang , Tong-Yee Lee This is my paper
classification cs.CV cs.AI
keywords human-object interaction video generationstreaming video generationlong video generationKV cache memorysink-local memorydiffusion transformermemory distance scalinglow-latency 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.

StreamHOI is trying to show that the reason streaming human-object interaction videos degrade over time is not a lack of memory but a mismatch between memory layout and where each transformer block looks. The paper finds that shallow blocks attend mostly to the human, hands, object, and contact region, while deeper blocks attend to the surrounding scene; yet standard streaming generators give every block the same split between early 'sink' frames and recent local frames. Its proposed training assigns more early-sink memory to interaction-focused blocks and more recent-frame memory to scene-focused blocks, and rescales temporal positional encoding so interaction blocks can reach far-away early states. If the claims hold, long interactive videos can stay coherent for a minute or more at interactive speeds (17.6 frames per second, 0.75-second first chunk) without enlarging the total KV-cache budget.

Core claim

The central discovery is that a streaming image-to-video generator has block-specific historical-memory preferences that are stable across interaction scenarios, and exploiting them resolves the sink-local trade-off. For each block the paper defines an offline bias score: average historical-memory attention mass on the HOI region minus that on the surrounding region. Blocks with positive bias are trained with more sink memory; blocks with negative bias with more local-window memory, all under a fixed total budget. A learned memory-distance-scaling factor then shortens the effective temporal distance to sink tokens in interaction-focused blocks. The paper reports that this preserves object id

What carries the argument

The load-bearing objects are the per-block spatial bias score b_l (the difference between historical-memory attention on the HOI region and on the surrounding region, Eqs. 1-2), the bias-normalized sink-size allocation s_l (Eqs. 4-5), and the learnable memory-distance-scaling coefficient lambda_l = sigma(rho_l), which rescales the temporal RoPE frequency for HOI-biased blocks (Eqs. 7-8). They carry the argument by converting an observed attention property into a training layout: profiling fixes which blocks get more sink memory, B-MST trains lightweight low-rank adapters under that layout, and MDS makes the extra sink tokens actually retrievable at large temporal distances.

Load-bearing premise

The heaviest assumption is that the block-type grouping computed on the untrained base generator remains the correct semantic layout after the LoRA and distance-scaling training changes the model's attention behavior.

What would settle it

Re-run the offline profiling with the same segmentation procedure after B-MST training and compare per-block bias scores; if many blocks flip sign or the HOI/surrounding margin narrows materially, the fixed grouping is stale. A second falsifier: find a long rollout where random or reverse grouping matches the semantic grouping's quality, which would undo the claim that matching memory to profiled bias is causal.

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

If this is right

  • Under the fixed KV-cache budget, 60-second HOI generation keeps object and interaction-coherence scores close to their 5-second values, instead of the sharp drop seen in streaming baselines.
  • Equal-step uniform-memory fine-tuning gives only marginal gains, so the improvement is attributable to block-specific memory assignment, not to extra training.
  • Reverse or random block grouping degrades long-term interaction consistency, implying the profiled semantic grouping is the operative cause.
  • Removing memory-distance scaling causes a clear metric drop, and the learned coefficients beat manually fixed ones, indicating that memory capacity and temporal-distance access are complementary.
  • The method is also demonstrated on a second streaming generator and improves its 60-second HOI metrics, suggesting the recipe is not tied to a single backbone.

Where Pith is reading between the lines

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

  • Editorial inference: the same profile-then-specialize loop could be extended to multi-person interactions if segmentation masks are pooled across all people, since the paper's limitation section flags single-person training data as the current boundary.
  • Editorial inference: because MDS is a per-block learnable rescaling of RoPE, its phase correction is distance-dependent; one could test whether head-level or layer-group-level rescaling captures interaction anchors even more cheaply than block-level.
  • Editorial inference: the stability claim for block bias implies profiling is a one-time cost per base generator, but the paper does not re-profile after training; re-profiling after B-MST would directly test whether the fixed grouping survives the intervention.

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 / 5 minor

Summary. The paper proposes StreamHOI, a streaming framework for long-duration human-object interaction (HOI) video generation built on an image-to-video streaming generator. It first diagnoses a sink-local memory trade-off under a fixed KV budget, then shows through attention profiling that different transformer blocks exhibit stable HOI-biased or surrounding-biased spatial attention over historical memory. StreamHOI uses this offline profiling to assign block-specific sink/local memory sizes and trains the generator, keeping the base model frozen and inserting LoRA adapters, with bias-guided memory-specialized training (B-MST). It also introduces learnable memory distance scaling (MDS) in HOI-biased blocks to compress temporal RoPE distances to early sink frames. The paper reports comparisons with streaming long-video baselines and bidirectional HOI generation baselines at 5/30/60 seconds, ablations including an equal-step Uniform-Mem FT control and reverse/random grouping controls, a cross-backbone transfer to LongLive-5B*, user studies, and runtime measurements (17.6 FPS, 0.75 s first-chunk latency). The central claim is that matching memory composition to profiled block behavior preserves long-horizon interaction consistency without increasing the total KV-cache budget.

Significance. If the central claim is correct, the paper makes a useful empirical and practical contribution. The proposed block-wise memory specialization under a fixed budget is a clean mechanism, and the evidence base is strong on the ablation side: the equal-step Uniform-Mem FT control directly addresses the 'fine-tuning alone' confound, and the reverse/random grouping controls test the semantic content of the grouping. Cross-dataset evaluation on unseen GeoHOI/HOMA benchmarks and the cross-backbone transfer to LongLive-5B* support generalization. However, the mechanism's key invariance — that the offline block classification remains valid after B-MST changes the generator — is not directly verified, and the main quantitative metrics are self-defined and reported without uncertainty estimates. These are fixable within the manuscript's scope, so I view the paper as promising but requiring revision.

major comments (3)
  1. [Sec. 4.2, Eqs. (1)-(2); Sec. 4.3, Eqs. (9)-(10); Table 4] The central mechanism assumes that the offline block classification remains valid after B-MST. The paper states the bias is stable 'across interaction scenarios within the same base generator,' but B-MST inserts rank-256 LoRA adapters in every block and learns per-block RoPE scaling in HOI-biased blocks; these interventions can plausibly shift which blocks rely on sink versus local context. The reverse/random grouping controls in Table 4 are trained with fixed groupings, so they only show that the original grouping helps conditional on that grouping; they do not establish that the trained model's post-training spatial bias matches the original grouping. Please re-run the Sec. 4.2 profiling on the B-MST-trained model and report block-type consistency (e.g., overlap or correlation with the original assignments). If the assignments shift, the mechanism should be reinterpreted as an imposed
  2. [Sec. 5.1.4, Tables 1, 2, 4, 5] The headline HOI-specific comparisons rely on InternVL3-38B-based binary metrics and Obj-CLIP, but no variance, confidence intervals, or significance tests are reported. Many differences are small; for example, Table 2 GeoHOI test set shows Obj-CLIP 0.8993 vs 0.9015 and InternVL I 0.892 vs 0.898. Without bootstrap or per-example variance it is difficult to assess whether the claimed improvements exceed noise. Please report mean ± std or 95% CIs over sampled videos/prompts and provide the precise metric protocol in the main text or a clearly accessible appendix. This is needed to support the central quantitative claims.
  3. [Sec. 5.1.3, Tables 1 and 6] The 1.3B baselines use released weights, but the 5B variants Causal Forcing-5B* and LongLive-5B* are internal reproductions. The comparison against streaming long-video baselines depends on these reproductions being faithful. Please release the reproduction configurations (training steps, data, memory budgets, inference settings) or make the weights available, so the claimed improvements over these baselines are independently verifiable.
minor comments (5)
  1. [Page 1, author line] Typo in author name: 'Tong-Y ee Lee' should be 'Tong-Yee Lee'. Also, 'In contrast,StreamHOI' in Section 2.1 is missing a space.
  2. [Eq. (8)] The sign in the phase perturbation expression is ambiguous. Since λ_l ∈ (0,1), the perturbation relative to the unscaled phase should be (λ_l − 1)(τ_t − τ_j)ω; if Eq. (8) reports an absolute magnitude, this should be stated explicitly.
  3. [Sec. 5.1.4] For Obj-CLIP, please state explicitly which segmentation source is used for the object regions in the generated videos: reference-object masks, SAM2 masks, or detection-based masks. This is important for reproducibility.
  4. [Table 4] The 'w/o B-MST' and 'Uniform-Mem FT' rows are numerically very close. Please clarify whether these are independent runs and whether the near-identical scores indicate that the 300-step DMD stage has fully converged under the uniform-memory design.
  5. [Sec. 5.4, Fig. 15] The user study reports preference rates without confidence intervals or significance tests. Given the modest number of participants (28), please add uncertainty estimates or a statistical test before claiming consistent preference advantages.

Circularity Check

0 steps flagged

No circular derivation: offline block profiling and B-MST allocation are empirical design choices, not fitted inputs renamed as predictions; self-citations are non-load-bearing baselines.

full rationale

The claimed derivation chain is: (1) observe the sink-local trade-off and block-wise attention bias; (2) classify blocks via Eq. (2) from the measured bias Eq. (1); (3) assign block-specific memory via Eq. (4) and train LoRA/MDS with this layout; (4) evaluate on GeoHOI/HOMA and long-video benchmarks. No step reduces to its own input by construction. Eq. (4) is a deterministic function of the profiled bias, but it is a training configuration, not a prediction; the paper never claims the allocation is derived from the evaluation metrics. The benefit of the allocation is tested with external metrics and controlled ablations (Table 4: uniform-memory FT, reverse/random grouping, w/o B-MST, w/o MDS), and the learnable MDS coefficients are compared against a fixed variant, so the central mechanism has independent empirical content. Self-citations (e.g., GeoHOI [8], HOMA [6], AnchorCrafter [3], and related Kling-AI memory works with overlapping authors) appear as baselines or related work and are not used to justify the core profiling-to-memory claim; no uniqueness theorem or ansatz is imported from the authors' prior work. The only genuinely load-bearing unverified step is the assumption that block bias profiled on the untrained base generator persists after B-MST; the paper states stability 'across interaction scenarios within the same base generator' but never re-profiles the trained model. This is an empirical correctness risk, not circularity: it is not an equation-to-equation reduction and does not make any evaluation a renamed input. Accordingly, no circular step is flagged.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No new physical or model entities are postulated. 'HOI-biased' and 'surrounding-biased' are labels assigned by Eq. 2 from existing attention patterns, and the memory distance scaling module is a parameter rescaling rather than a new entity. The central dependence is on profiled bias scores and several hand-chosen memory-budget hyperparameters.

free parameters (6)
  • s0 (initial sink size) = 4 latent frames
    Chosen for the uniform memory design in Sec. 5.1.1; anchors the block-specific allocation in Eq. 4.
  • B (total per-block memory budget = s0 + w0) = 12 latent frames
    Fixed budget from s0=4 and w0=8; Eq. 5 keeps every block's total at B.
  • smin, wmin (minimum sink/local frames) = unspecified
    Clip bounds in Eq. 4 chosen by hand; they constrain allocation for blocks with extreme bias scores.
  • epsilon in Eq. 3 = small constant
    Numerical stabilizer in bias normalization; value not reported.
  • lambda_l = sigmoid(rho_l) (learnable MDS coefficients) = not reported
    Learned during B-MST for each HOI-biased block; central to the MDS gain in Table 4 but final values are never reported.
  • lambda_c (critic loss weight in Eq. 10) = unspecified
    Weight of the DMD critic term in the B-MST objective; not stated in the main text.
axioms (6)
  • domain assumption Wan2.2-TI2V-5B provides a strong image-to-video prior whose frozen blocks can be specialized by LoRA without full fine-tuning.
    All experiments assume this backbone and that block-level cache changes do not require updating base weights.
  • domain assumption DMD distillation (Eq. 10) is a valid training signal for streaming generation quality.
    The paper relies on the DMD objective from [30] as its training target; this is an established algorithm, not proved here.
  • domain assumption SAM2 masks define HOI and surrounding regions accurately enough to measure block-wise historical-memory attention.
    Profiling in Sec. 4.2 segments human and object with SAM2, then pools attention; inaccurate masks would corrupt bias scores.
  • ad hoc to paper InternVL3-38B structured binary evaluation faithfully measures object fidelity, human quality, and interaction plausibility.
    This is the paper's own evaluation framework; its validity is asserted rather than demonstrated against human labels.
  • ad hoc to paper Block-wise spatial bias measured before B-MST is stable across datasets and invariant to the B-MST training intervention.
    The paper claims stability for the base generator (Sec. 4.2) but does not re-profile after training; this is the weakest load-bearing premise.
  • standard math Temporal RoPE frequency rescaling only strongly affects distant sink keys, as claimed from Eq. 8.
    The distance-dependent correction follows from the RoPE phase formula; the paper uses this to justify applying MDS only to HOI-biased blocks.

pith-pipeline@v1.3.0-alltime-deepseek · 19884 in / 11993 out tokens · 123910 ms · 2026-08-01T10:33:25.369114+00:00 · methodology

0 comments
read the original abstract

Existing human--object interaction (HOI) video generation methods are largely limited to offline short-video generation with complex driving conditions, making them unsuitable for real-time interactive applications. We present \emph{StreamHOI}, a low-latency streaming framework for long-duration HOI video generation. Instead of converting heavily conditioned HOI pipelines into streaming systems, we study how an image-to-video streaming generator should organize historical memory to preserve interactions under bounded latency. We find that the standard sink-local memory design faces a trade-off in streaming HOI generation, and different transformer blocks show different historical-memory preferences for HOI regions and surrounding regions. To match memory composition with block behavior, StreamHOI performs offline HOI-aware block profiling and applies bias-guided memory-specialized training to adapt the generator to block-specific memory layouts. We further introduce a memory distance scaling module to strengthen long-range access to early interaction states. Extensive comparisons with both long-video baselines and recent HOI generation methods demonstrate that StreamHOI achieves strong interaction plausibility, object fidelity, human quality and efficiency, reaching 17.6 FPS with 0.75s first-chunk latency.

Figures

Figures reproduced from arXiv: 2607.20174 by Fan Tang, Guoxin Zhang, Haoxian Zhang, Pengfei Wan, Tong-Yee Lee, Xiaoqiang Liu, Yiping Meng, Zejing Rao.

Figure 1
Figure 1. Figure 1: Conceptual comparison of HOI video generation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the sink-local memory trade-off in streaming HOI generation under a fixed memory budget. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Block-wise historical-memory attention across two [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pipeline of offline HOI-aware block profiling. We [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-sample stability of the block-wise spatial atten [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Pipeline of bias-guided memory-specialized training (B-MST). To resolve the sink-local memory trade-off in [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison against state-of-the-art long-video generation frameworks. While baseline methods exhibit [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Further comparison with Longcat-Video-Avatar 1.5 [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison with SOTAs on the GeoHOI-testset [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison with SOTAs on the HOMA-testset [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual ablation of B-MST and MDS variants. Re [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Mitigation of temporal degradation using learnable [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Ablation of block grouping strategies. Mismatch [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: User study results. Pie charts show first-choice preference rates on Long-Video, GeoHOI, and HOMA benchmarks under four perceptual criteria: Human Consistency, Object Consistency, Interaction Naturalness, and Overall Video Quality. 》》 20000 Frame Ref 》》 “The young man sits on a beige couch, wearing a plaid shirt and cap, strumming a black guitar with his fingers in a cozy, warmly lit living room.” “A man … view at source ↗
Figure 17
Figure 17. Figure 17: Dynamic prompt switching during streaming gen [PITH_FULL_IMAGE:figures/full_fig_p013_17.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

51 extracted references · 19 linked inside Pith

  1. [1]

    Graspdiff: Grasping generation for hand-object interaction with multimodal guided diffusion,

    B. Zuo, Z. Zhao, W. Sun, X. Yuan, Z. Yu, and Y. Wang, “Graspdiff: Grasping generation for hand-object interaction with multimodal guided diffusion,”IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 9, pp. 5642–5654, 2024

    2024

  2. [2]

    Hoimo- tion: Forecasting human motion during human-object interactions using egocentric 3d object bounding boxes,

    Z. Hu, Z. Yin, D. Haeufle, S. Schmitt, and A. Bulling, “Hoimo- tion: Forecasting human motion during human-object interactions using egocentric 3d object bounding boxes,”IEEE Transactions on Visualization and Computer Graphics, vol. 30, no. 11, pp. 7375–7385, 2024

    2024

  3. [3]

    Anchorcrafter: Animate cyberanchors saling your products via human-object interacting video genera- tion,

    Z. Xu, Z. Huang, J. Cao, Y. Zhang, X. Cun, Q. Shuai, Y. Wang, L. Bao, J. Li, and F. Tang, “Anchorcrafter: Animate cyberanchors saling your products via human-object interacting video genera- tion,”arXiv preprint arXiv:2411.17383, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Make-your-video: Customized video gen- eration using textual and structural guidance,

    J. Xing, M. Xia, Y. Liu, Y. Zhang, Y. Zhang, Y. He, H. Liu, H. Chen, X. Cun, X. Wanget al., “Make-your-video: Customized video gen- eration using textual and structural guidance,”IEEE transactions on visualization and computer graphics, vol. 31, no. 2, pp. 1526–1541, 2024

    2024

  5. [5]

    Motioncrafter: Plug-and-play motion guid- ance for diffusion models,

    Y. Zhang, W. Dong, F. Tang, N. Huang, H. Huang, C. Ma, P . Wan, T.-Y. Lee, and C. Xu, “Motioncrafter: Plug-and-play motion guid- ance for diffusion models,”IEEE Transactions on Visualization and Computer Graphics, 2025

    2025

  6. [6]

    Hunyuanvideo-homa: Generic human- object interaction in multimodal driven human animation,

    Z. Huang, Z. Zhou, J. Cao, Y. Ma, Y. Chen, Z. Rao, Z. Xu, H. Wang, Q. Lin, Y. Zhouet al., “Hunyuanvideo-homa: Generic human- object interaction in multimodal driven human animation,”arXiv preprint arXiv:2506.08797, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Hoi-swap: Swap- ping objects in videos with hand-object interaction awareness,

    Z. S. Xue, R. Luo, C. Chen, and K. Grauman, “Hoi-swap: Swap- ping objects in videos with hand-object interaction awareness,” Advances in Neural Information Processing Systems (NeurIPS), vol. 37, pp. 77 132–77 164, 2024

    2024

  8. [8]

    Geohoi: Geometry-enhanced human-object interaction video generation via hierarchical multi-modal injection,

    Z. Xu, Z. Rao, J. Cao, X. Liu, Z. Fang, H. Zhang, S. Tang, and F. Tang, “Geohoi: Geometry-enhanced human-object interaction video generation via hierarchical multi-modal injection,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 3739–3748

    2026

  9. [9]

    Making avatars interact: To- wards text-driven human-object interaction for controllable talk- ing avatars,

    Y. Zhang, Z. Zhou, Z. Yu, Z. Huang, T. Hu, S. Liang, G. Zhang, Z. Peng, S. Li, Y. Chenet al., “Making avatars interact: To- wards text-driven human-object interaction for controllable talk- ing avatars,”arXiv preprint arXiv:2602.01538, 2026

    arXiv 2026

  10. [10]

    From slow bidirectional to fast autore- gressive video diffusion models,

    T. Yin, Q. Zhang, R. Zhang, W. T. Freeman, F. Durand, E. Shecht- man, and X. Huang, “From slow bidirectional to fast autore- gressive video diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 22 963–22 974

    2025

  11. [11]

    Self forcing: Bridging the train-test gap in autoregressive video diffusion,

    X. Huang, Z. Li, G. He, M. Zhou, and E. Shechtman, “Self forcing: Bridging the train-test gap in autoregressive video diffusion,” arXiv preprint arXiv:2506.08009, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Causal forcing: Autoregressive diffusion distillation done right for high- quality real-time interactive video generation,

    H. Zhu, M. Zhao, G. He, H. Su, C. Li, and J. Zhu, “Causal forcing: Autoregressive diffusion distillation done right for high- quality real-time interactive video generation,”arXiv preprint arXiv:2602.02214, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    Longlive: Real-time interactive long video generation,

    S. Yang, W. Huang, R. Chu, Y. Xiao, Y. Zhao, X. Wang, M. Li, E. Xie, Y. Chen, Y. Lu, and S. H. Y. Chen, “Longlive: Real-time interactive long video generation,” 2025

    2025

  14. [14]

    Make-your-anchor: A diffusion-based 2d avatar generation framework,

    Z. Huang, F. Tang, Y. Zhang, X. Cun, J. Cao, J. Li, and T.-Y. Lee, “Make-your-anchor: A diffusion-based 2d avatar generation framework,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 6997–7006

    2024

  15. [15]

    Make-your-anchor+: Temporal consistent 2d avatar generation via video diffusion prior,

    Z. Huang, F. Tang, J. Cao, Y. Zhang, X. Cun, Y. Bo, J. Li, and T.-Y. Lee, “Make-your-anchor+: Temporal consistent 2d avatar generation via video diffusion prior,”IEEE Transactions on Visualization & Computer Graphics, vol. 32, no. 04, pp. 2883–2897, Apr. 2026. [Online]. Available: https://doi.ieeecomputersociety. org/10.1109/TVCG.2026.3655478

    arXiv 2026

  16. [16]

    Animate anyone 2: High-fidelity character image animation with environment affordance,

    L. Hu, G. Wang, Z. Shen, X. Gao, D. Meng, L. Zhuo, P . Zhang, B. Zhang, and L. Bo, “Animate anyone 2: High-fidelity character image animation with environment affordance,”arXiv preprint arXiv:2502.06145, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Mimo: Controllable char- acter video synthesis with spatial decomposed modeling,

    Y. Men, Y. Yao, M. Cui, and L. Bo, “Mimo: Controllable char- acter video synthesis with spatial decomposed modeling,” in Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025, pp. 21 181–21 191

    2025

  18. [18]

    Re-hold: Video hand object interaction reenactment via adaptive layout-instructed diffusion model,

    Y. Fan, Q. Yang, K. Wang, H. Zhou, Y. Li, H. Feng, E. Ding, Y. Wu, and J. Wang, “Re-hold: Video hand object interaction reenactment via adaptive layout-instructed diffusion model,” inProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025, pp. 17 550–17 560

    2025

  19. [19]

    idit-hoi: Inpainting-based hand object interaction reenactment via video diffusion transformer,

    Z. Shen, C. Wu, J. Zhou, C. Zhao, K. Wang, H. Zhou, Y. Li, H. Feng, W. He, and J. Wang, “idit-hoi: Inpainting-based hand object interaction reenactment via video diffusion transformer,” arXiv preprint arXiv:2506.12847, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Svimo: Synchronized diffusion for video and motion generation in hand- object interaction scenarios,

    L. Dang, R. Shao, H. Zhang, W. Min, Y. Liu, and Q. Wu, “Svimo: Synchronized diffusion for video and motion generation in hand- object interaction scenarios,”arXiv preprint arXiv:2506.02444, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Harmohoi: Harmonizing appearance and 3d motion 15 for multi-view hand-object interaction synthesis,

    L. Dang, J. Li, Z. Li, H. Zhang, L. An, W. Min, Y. Liu, and Q. Wu, “Harmohoi: Harmonizing appearance and 3d motion 15 for multi-view hand-object interaction synthesis,” 2026. [Online]. Available: https://arxiv.org/abs/2607.17097

    Pith/arXiv arXiv 2026

  22. [22]

    Dreamactor-h1: High-fidelity human- product demonstration video generation via motion-designed dif- fusion transformers,

    L. Wang, Z. Xia, T. Hu, P . Wang, P . Wei, Z. Zheng, M. Zhou, Y. Zhang, and M. Gao, “Dreamactor-h1: High-fidelity human- product demonstration video generation via motion-designed dif- fusion transformers,”arXiv preprint arXiv:2506.10568, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Pyramidal flow matching for efficient video generative modeling,

    Y. Jin, Z. Sun, N. Li, K. Xu, H. Jiang, N. Zhuang, Q. Huang, Y. Song, Y. Mu, and Z. Lin, “Pyramidal flow matching for efficient video generative modeling,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 23 378–23 402

    2025

  24. [24]

    Diffusion forcing: Next-token prediction meets full-sequence diffusion,

    B. Chen, D. Mart ´ı Mons ´o, Y. Du, M. Simchowitz, R. Tedrake, and V . Sitzmann, “Diffusion forcing: Next-token prediction meets full-sequence diffusion,”Advances in Neural Information Processing Systems, vol. 37, pp. 24 081–24 125, 2024

    2024

  25. [25]

    History-guided video diffusion,

    K. Song, B. Chen, M. Simchowitz, Y. Du, R. Tedrake, and V . Sitzmann, “History-guided video diffusion,”arXiv preprint arXiv:2502.06764, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Consistency models,

    Y. Song, P . Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” 2023

    2023

  27. [27]

    Improved techniques for training con- sistency models,

    Y. Song and P . Dhariwal, “Improved techniques for training con- sistency models,” inInternational Conference on Learning Represen- tations, vol. 2024, 2024, pp. 15 078–15 097

    2024

  28. [28]

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,

    Z. Wang, C. Lu, Y. Wang, F. Bao, C. Li, H. Su, and J. Zhu, “Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,”Advances in neural information processing systems, vol. 36, pp. 8406–8441, 2023

    2023

  29. [29]

    Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models,

    W. Luo, T. Hu, S. Zhang, J. Sun, Z. Li, and Z. Zhang, “Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models,”Advances in Neural Information Processing Sys- tems, vol. 36, pp. 76 525–76 546, 2023

    2023

  30. [30]

    One-step diffusion with distribution matching distillation,

    T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Free- man, and T. Park, “One-step diffusion with distribution matching distillation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 6613–6623

    2024

  31. [31]

    Frame context packing and drift prevention in next-frame-prediction video diffusion models,

    L. Zhang, S. Cai, M. Li, G. Wetzstein, and M. Agrawala, “Frame context packing and drift prevention in next-frame-prediction video diffusion models,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  32. [32]

    Rolling forcing: Autoregressive long video diffusion in real time,

    K. Liu, W. Hu, J. Xu, Y. Shan, and S. Lu, “Rolling forcing: Autoregressive long video diffusion in real time,”arXiv preprint arXiv:2509.25161, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Self-forcing++: Towards minute-scale high-quality video generation,

    J. Cui, J. Wu, M. Li, T. Yang, X. Li, R. Wang, A. Bai, Y. Ban, and C.- J. Hsieh, “Self-forcing++: Towards minute-scale high-quality video generation,” inThe Fourteenth International Conference on Learning Representations

  34. [34]

    Context as memory: Scene-consistent interactive long video gen- eration with memory retrieval,

    J. Yu, J. Bai, Y. Qin, Q. Liu, X. Wang, P . Wan, D. Zhang, and X. Liu, “Context as memory: Scene-consistent interactive long video gen- eration with memory retrieval,”arXiv preprint arXiv:2506.03141, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Memory forcing: Spatio-temporal memory for consistent scene generation on minecraft,

    J. Huang, X. Hu, B. Han, S. Shi, Z. Tian, T. He, and L. Jiang, “Memory forcing: Spatio-temporal memory for consistent scene generation on minecraft,”arXiv preprint arXiv:2510.03198, 2025

    arXiv 2025

  36. [36]

    Memflow: Flowing adaptive memory for consistent and efficient long video narratives,

    S. Ji, X. Chen, S. Yang, X. Tao, P . Wan, and H. Zhao, “Memflow: Flowing adaptive memory for consistent and efficient long video narratives,”arXiv preprint arXiv:2512.14699, 2025

    arXiv 2025

  37. [37]

    Context forcing: Consistent autoregressive video generation with long context,

    S. Chen, C. Wei, S. Sun, P . Nie, K. Zhou, G. Zhang, M.-H. Yang, and W. Chen, “Context forcing: Consistent autoregressive video generation with long context,”arXiv preprint arXiv:2602.06028, 2026

    arXiv 2026

  38. [38]

    Infinity-rope: Action-controllable infinite video genera- tion emerges from autoregressive self-rollout,

    H. Yesiltepe, T. Meral, A. K. Akan, K. Oktay, and P . Ya- nardag, “Infinity-rope: Action-controllable infinite video genera- tion emerges from autoregressive self-rollout,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 40 256–40 265

    2026

  39. [39]

    Deep forcing: Training-free long video generation with deep sink and participative compression,

    J. Yi, W. Jang, P . H. Cho, J. Nam, H. Yoon, and S. Kim, “Deep forcing: Training-free long video generation with deep sink and participative compression,”arXiv preprint arXiv:2512.05081, 2025

    arXiv 2025

  40. [40]

    Head forcing: Long autoregressive video generation via head heterogeneity,

    J. Tian, Y. Wang, G. Yu, and C. Zhang, “Head forcing: Long autoregressive video generation via head heterogeneity,” inarXiv preprint, 2026

    2026

  41. [41]

    Hoigen-1m: A large-scale dataset for human-object interaction video generation,

    K. Liu, Q. Liu, X. Liu, J. Li, Y. Zhang, J. Luo, X. He, and W. Liu, “Hoigen-1m: A large-scale dataset for human-object interaction video generation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 24 001–24 010

    2025

  42. [42]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R¨adle, C. Rolland, L. Gustafsonet al., “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Longcat-video-avatar 1.5 technical report,

    M. L. Team, X. Cai, M. Cheng, F. Gao, Z. Kong, J. Li, L. Li, W. Li, H. Liu, S. Tanet al., “Longcat-video-avatar 1.5 technical report,” arXiv preprint arXiv:2605.26486, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    Qwen2.5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P . Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P . Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin, “Qwen2.5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Unianimate-dit: Human image animation with large- scale video diffusion transformer,

    X. Wang, S. Zhang, L. Tang, Y. Zhang, C. Gao, Y. Wang, and N. Sang, “Unianimate-dit: Human image animation with large- scale video diffusion transformer,”arXiv preprint arXiv:2504.11289, 2025

    arXiv 2025

  46. [46]

    Vace: All- in-one video creation and editing,

    Z. Jiang, Z. Han, C. Mao, J. Zhang, Y. Pan, and Y. Liu, “Vace: All- in-one video creation and editing,”arXiv preprint arXiv:2503.07598, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Humo: Human-centric video generation via collaborative multi-modal conditioning,

    L. Chen, T. Ma, J. Liu, B. Li, Z. Chen, L. Liu, X. He, G. Li, Q. He, and Z. Wu, “Humo: Human-centric video generation via collaborative multi-modal conditioning,”arXiv preprint arXiv:2509.08519, 2025

    Pith/arXiv arXiv 2025

  48. [48]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  49. [49]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models,

    J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, H. Tian, Y. Duan, W. Su, J. Shaoet al., “Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models,”arXiv preprint arXiv:2504.10479, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    Vbench: Comprehensive benchmark suite for video generative models,

    Z. Huang, Y. He, J. Yu, F. Zhang, C. Si, Y. Jiang, Y. Zhang, T. Wu, Q. Jin, N. Chanpaisitet al., “Vbench: Comprehensive benchmark suite for video generative models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 21 807–21 818

    2024

  51. [51]

    Qwen3-vl technical report,

    S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y. Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y. Sun, J. Tang, J. Tu, J. Wan, P . Wang, P . Wang, ...

    Pith/arXiv arXiv 2025