Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Forking neighborhood candidates at a random chunk lets contrastive GRPO align few-step streaming AR video generators without SDE exploration.

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 · grok-4.5

2026-07-13 23:08 UTC pith:7DCTFQTI

load-bearing objection Solid engineering fix for RLHF on few-step streaming AR video: chunk forking + ˆx0 CoPO + semi-on-policy LoRA merge actually moves both VideoAlign and VBench, and the SDE-failure diagnosis is clean. the 3 major comments →

arxiv 2603.17461 v2 pith:7DCTFQTI submitted 2026-03-18 cs.CV

AR-CoPO: Align Autoregressive Video Generation with Contrastive Policy Optimization

Dailan He , Guanlin Feng , Xingtong Ge , Yi Zhang , Bingqi Ma , Guanglu Song , Yu Liu , Hongsheng Li This is my paper
classification cs.CV
keywords autoregressive video generationRLHFcontrastive policy optimizationGRPOfew-step distillationconsistency modelschunk-level alignmentsemi-on-policy
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.

Few-step streaming autoregressive video generators deliver low-latency synthesis, yet standard SDE-based GRPO alignment fails on them: short consistency-model trajectories are nearly deterministic and driven almost entirely by initial noise, so intermediate SDE noise yields no useful exploration. This paper claims that a contrastive reformulation of GRPO, applied only at a randomly chosen pivot chunk via a controlled forking mechanism, can turn sequence-level rewards into localized, stable credit assignment. A complementary semi-on-policy strategy that reuses a replay buffer of reference rollouts under ratio clipping further improves quality where pure on-policy noise exploration collapses. On Self-Forcing the merged model raises VideoAlign overall scores while holding or improving VBench totals, which the authors read as genuine preference alignment rather than reward hacking. Readers who need fast, controllable video generation care because the method keeps inference deterministic and few-step while still enabling RLHF-style post-training.

Core claim

AR-CoPO enables RLHF of streaming few-step AR video generators by constructing neighborhood candidates through controlled forking at a randomly selected pivot chunk, scoring completed sequences, and performing Neighbor-GRPO-style contrastive updates confined to that chunk. Combined with semi-on-policy exploitation over a fixed reference replay buffer and LoRA merging of exploration and exploitation adapters, it improves both in-domain human preference metrics and out-of-domain VBench scores over Self-Forcing, which the authors present as evidence of genuine alignment rather than reward hacking.

What carries the argument

Chunk-level forking with controlled noise sharing: at a random pivot chunk the model branches G neighbors by perturbing only that chunk’s initial noise while sharing all other noises across branches, completes each sequence, assigns sequence-level rewards, and runs a distance-based contrastive GRPO update only on the pivot chunk (using ˆx0 prediction distances for consistency models).

Load-bearing premise

The method assumes that sharing every non-pivot noise across branches makes sequence-level reward differences cleanly attributable only to the pivot chunk’s initial noise, so the local GRPO update is valid credit assignment.

What would settle it

Retrain AR-CoPO with independent sampling of all later-chunk and solver noises per branch instead of controlled sharing; if the shared-noise version no longer improves VideoAlign while holding VBench, or if independent-noise still yields clean dual-benchmark gains, the clean-attribution claim fails.

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

If this is right

  • Few-step streaming AR video models can be post-trained with preference rewards without converting inference to long SDE trajectories.
  • Sequence-level rewards can be credited to individual chunks, avoiding full-sequence multi-branch back-propagation cost.
  • Semi-on-policy replay under ratio clipping can raise quality and text alignment where pure on-policy noise exploration collapses motion quality.
  • Merging separate on-policy and semi-on-policy LoRA adapters lets practitioners trade reward-seeking exploration against quality-preserving exploitation at inference.
  • Requiring joint gains on both an in-domain preference suite and an independent out-of-domain benchmark becomes a practical guard against reward hacking.

Where Pith is reading between the lines

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

  • The same forking-plus-contrastive pattern may transfer to other near-deterministic few-step autoregressive generators (images, audio) where intermediate SDE noise is uninformative.
  • If residual decoder or solver non-determinism still confounds rewards, stronger noise isolation or noise-free decoding would tighten credit assignment further.
  • The LoRA-merge scale that jointly improves both benchmarks may serve as a diagnostic for over-optimization in other preference-aligned generative models.

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 AR-CoPO for RLHF of few-step streaming autoregressive video generators (e.g., Self-Forcing). It adapts Neighbor GRPO's contrastive view of SDE-GRPO: neighborhood candidates are built by forking at a randomly chosen pivot chunk via initial-noise perturbation, sequence-level rewards are assigned to completed branches, and a localized GRPO update is performed only on that chunk using a distance-induced surrogate policy (in ˆx0 space for consistency models). A semi-on-policy variant freezes rollouts from a reference policy into a replay buffer and applies the same clipped contrastive objective for exploitation; on-policy and semi-on-policy LoRA adapters are trained separately and merged. On Self-Forcing, the merged model raises VideoAlign Overall (7.76 o8.22) while holding VBench Total (82.15 o82.17); ablations show SDE-GRPO fails, pure on-policy TA optimization collapses MQ/VQ, and ratio clipping is required for semi-on-policy stability. Causal-Forcing results appear in the appendix.

Significance. If the dual-benchmark gains hold under broader evaluation, the work supplies a practical post-training recipe for a class of models (few-step causal AR video generators) that existing SDE-based GRPO methods cannot usefully align. The controlled noise-substitution study (Fig. 6) and the on-/semi-/off-policy TA ablation (Table 2, Fig. 8) are concrete, falsifiable contributions that clarify why intermediate SDE exploration is ineffective and why pure on-policy TA optimization is unstable. The dual-criterion selection rule (improve both VideoAlign and VBench) is a useful guard against reward hacking and is reported transparently via the merge-scale table. The method is empirical rather than theoretical, but the design choices (chunk forking, controlled noise sharing, ˆx0 distances, LoRA merge) are clearly specified and reproducible in principle.

major comments (3)
  1. §4.1 / Table 1 and §4.4 / Table 3: the central claim of 'genuine alignment rather than reward hacking' rests on a single dual-benchmark comparison (VideoAlign Overall 7.76 o8.22 with VBench Total 82.15 o82.17 at merge scale 0.8). The merge-scale table shows a clear trade-off; scale 1.0 improves VideoAlign further but drops VBench below baseline. The paper should report variance (multiple seeds or bootstrap CIs) for the selected scale and, ideally, a human preference study on a held-out prompt set, so that the dual-criterion result is not a single-point observation.
  2. §3.2 (Controlled noise sharing) and Algorithm 1: the design asserts that sharing all non-pivot initial noises and all CM solver noises makes reward differences cleanly attributable to the pivot-chunk initial noise. Residual decoder non-determinism, KV-cache numerical drift, or any unshared randomness would confound the advantages. A short verification (e.g., identical-noise re-rollout variance of r(i), or an ablation that deliberately unshares later-chunk noise) would strengthen the credit-assignment rationale; without it the localized update remains heuristically motivated.
  3. §4 Experiments: main results are reported only for Self-Forcing; Causal-Forcing appears only in the appendix (Table 4) and LongLive is used only as a baseline. Given that the abstract claims a general framework for streaming AR generators, at least one additional full main-table evaluation (or an explicit statement that Causal-Forcing is the sole secondary check) is needed to support generality.
minor comments (5)
  1. Fig. 3 caption and Algorithm 1: the surrogate policy is written with ˆx0 distances for CMs, but the main text also states that the original intermediate-xt distance applies for FM-based AR generators; a single sentence clarifying which distance is used in the Self-Forcing experiments would avoid ambiguity.
  2. §3.3 and Fig. 4: 'semi-on-policy' is used for a fixed-reference replay buffer with ratio clipping; a brief note distinguishing it from classical off-policy RL terminology would help readers from the RL community.
  3. Appendix §5.1: ratio clip threshold is listed as 1e-4, which is unusually tight relative to standard PPO/GRPO practice; a one-sentence justification or sensitivity note would be useful.
  4. Typos / formatting: 'Self-Forcingdemonstrate' (abstract), missing spaces after some figure references, and inconsistent capitalization of 'VideoAlign' / 'VBench' in a few places.
  5. Related work: Neighbor GRPO is correctly cited as prior work by overlapping authors; a short explicit statement of what is new (chunk forking + semi-on-policy + CM ˆx0 distances) versus what is reused would improve novelty disclosure.

Circularity Check

1 steps flagged

No significant circularity: empirical RLHF procedure with external dual-benchmark evaluation; Neighbor GRPO self-citation is prior machinery, not a load-bearing uniqueness claim.

specific steps
  1. self citation load bearing [§3.1 Preliminary: Neighbor GRPO; Eqs. 1–3; citation [2]]
    "We draw inspiration from the recently proposed Neighbor GRPO [2] perspective, which offers a novel reinterpretation: the SDE-GRPO update can be mathematically reformulated as a distance-driven contrastive objective over a set of neighbor candidate trajectories. ... Neighbor GRPO [2] aligns flow-based generative models by reinterpreting SDE-based policy optimization as a distance-driven contrastive learning objective, and adopts a contrastive policy optimization (CoPO)."

    The surrogate transition πθ(i) and the clipped GRPO objective (Eqs. 2–3) are taken directly from the authors’ prior Neighbor GRPO paper rather than re-derived from first principles here. This is ordinary self-citation of prior machinery; it is not load-bearing for the paper’s strongest claim (dual-benchmark improvement on Self-Forcing), which is measured externally and does not require the Neighbor GRPO equations to be uniquely forced.

full rationale

AR-CoPO is an engineering adaptation of contrastive GRPO to streaming few-step AR video generators. The central claims are empirical: after forking + CoPO + semi-on-policy + LoRA merge, VideoAlign Overall rises (7.76 o8.22) while VBench Total is maintained (82.15 o82.17), which the authors treat as evidence against pure reward hacking (Table 1, §4.4). Those benchmarks are external and not defined by the training objective, so the reported gains are not forced by construction. Neighbor GRPO [2] (overlapping authors) supplies the distance-driven surrogate policy (Eqs. 1–3) and is cited as prior work; the present paper’s contribution is the AR-specific chunk forking, controlled noise sharing, ˆx0-space distance for CMs, and semi-on-policy/LoRA-merge strategy. No uniqueness theorem is imported to forbid alternatives, no fitted parameter is renamed a prediction, and no equation reduces to its own input by definition. The single minor self-citation is therefore non-load-bearing for the dual-benchmark result. Score 1 reflects only that citation; the derivation chain itself is self-contained against external metrics.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard RL and generative-model machinery plus a handful of hand-chosen hyper-parameters and the modeling assumption that sequence-level rewards can be localized to a single pivot chunk. No new physical entities are postulated; the invented constructs are algorithmic.

free parameters (5)
  • group size G = 12
    Number of neighborhood candidates per fork; set to 12 following Neighbor GRPO, directly affects advantage estimates and compute.
  • initial-noise perturbation strength σ = 0.5
    Controls exploration radius of the forked noises (Eq. 1); set to 0.5 for all experiments.
  • LoRA merge scale for on-policy adapter = 0.8
    Selected post-hoc by dual-benchmark criterion (Table 3); final reported model uses 0.8.
  • ratio-clip threshold = 1e-4
    Trust-region width in the GRPO objective; set to 1e-4.
  • temperature τ0 for ˆx0 distances
    Softmax temperature in the surrogate policy for consistency-model alignment (Eq. 4); value not numerically stated beyond “temperature hyperparameter.”
axioms (4)
  • domain assumption Few-step consistency-model / distilled ODE samplers are near-deterministic; sample diversity is governed almost entirely by the initial chunk noise rather than intermediate solver noise.
    Stated and experimentally supported in §4.2 and Fig. 6; load-bearing for why SDE-GRPO fails and why forking only initial noise is sufficient.
  • domain assumption Sequence-level reward differences among branches that share all non-pivot noises are cleanly attributable to the pivot-chunk action.
    “Controlled noise sharing” paragraph in §3.2; required for localized credit assignment.
  • domain assumption Neighbor GRPO’s distance-driven contrastive surrogate (Eqs. 2–3) remains a valid policy-gradient estimator when distances are measured in ˆx0 space for consistency models.
    §3.2 “CoPO for Consistency Model Alignment”; extension of the cited Neighbor GRPO paper.
  • standard math Standard GRPO / PPO-style ratio clipping enforces a sufficient trust region for semi-on-policy updates over a fixed reference buffer.
    Invoked via Schulman et al. TRPO/PPO citations; used to justify stability of the semi-on-policy regime.
invented entities (2)
  • chunk-level forking mechanism no independent evidence
    purpose: Constructs a local action space at a randomly chosen pivot chunk so that sequence-level rewards can drive a localized GRPO update without full-sequence multi-branch back-propagation.
    Core algorithmic novelty of AR-CoPO (§3.2); no independent existence outside this training procedure.
  • semi-on-policy AR-CoPO training with reference replay buffer no independent evidence
    purpose: Complements on-policy exploration with exploitation of fixed high-quality rollouts while ratio clipping keeps the policy inside a trust region.
    Introduced in §3.3 and Fig. 4; the dual-adapter merge is part of the same construct.

pith-pipeline@v1.1.0-grok45 · 17964 in / 3428 out tokens · 33298 ms · 2026-07-13T23:08:49.176051+00:00 · methodology

0 comments
read the original abstract

Streaming autoregressive (AR) video generators combined with few-step distillation achieve low-latency, high-quality synthesis, yet remain difficult to align via reinforcement learning from human feedback (RLHF). Existing SDE-based GRPO methods face challenges in this setting: few-step ODEs and consistency model samplers deviate from standard flow-matching ODEs, and their short, low-stochasticity trajectories are highly sensitive to initialization noise, rendering intermediate SDE exploration ineffective. We propose AR-CoPO (AutoRegressive Contrastive Policy Optimization), a framework that adapts the Neighbor GRPO contrastive perspective to streaming AR generation. AR-CoPO introduces chunk-level alignment via a forking mechanism that constructs neighborhood candidates at a randomly selected chunk, assigns sequence-level rewards, and performs localized GRPO updates. We further propose a semi-on-policy training strategy that complements on-policy exploration with exploitation over a replay buffer of reference rollouts, improving generation quality across domains. Experiments on Self-Forcing demonstrate that AR-CoPO improves both out-of-domain generalization and in-domain human preference alignment over the baseline, providing evidence of genuine alignment rather than reward hacking.

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. TempAct: Advancing Temporal Plausibility in Autoregressive Video Generation via Planner-Executor RL

    cs.CV 2026-06 unverdicted novelty 7.0

    TempAct introduces a planner-executor RL framework with hierarchical group exploration and rewards to improve temporal consistency in autoregressive video diffusion models.

  2. KVPO: ODE-Native GRPO for Autoregressive Video Alignment via KV Semantic Exploration

    cs.CV 2026-05 unverdicted novelty 7.0

    KVPO aligns streaming autoregressive video generators with human preferences via ODE-native GRPO, using KV cache for semantic exploration and TVE for velocity-based policy modeling, yielding gains in quality and alignment.

  3. TempAct: Advancing Temporal Plausibility in Autoregressive Video Generation via Planner-Executor RL

    cs.CV 2026-06 unverdicted novelty 6.0

    TempAct applies hierarchical planner-executor RL with group exploration and multi-level rewards to improve temporal consistency in autoregressive video models.

Reference graph

Works this paper leans on

34 extracted references · 17 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Advances in neural information processing systems30(2017)

    Christiano, P.F., Leike, J., Brown, T., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. Advances in neural information processing systems30(2017)

    2017

  2. [2]

    arXiv preprint arXiv:2511.16955 (2025)

    He, D., Feng, G., Ge, X., Niu, Y., Zhang, Y., Ma, B., Song, G., Liu, Y., Li, H.: Neighbor grpo: Contrastive ode policy optimization aligns flow models. arXiv preprint arXiv:2511.16955 (2025)

    arXiv 2025

  3. [3]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  4. [4]

    arXiv preprint arXiv:2506.08009 (2025)

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

    Pith/arXiv arXiv 2025

  5. [5]

    arXiv preprint arXiv:2412.03603 (2024)

    Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., et al.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Labs,B.F.,Batifol,S.,Blattmann,A.,Boesel,F.,Consul,S.,Diagne,C.,Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Müller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., Smith, L.: Flux.1 kontext: Flow matching for in-context image generation and editing in latent space (2025),https://arxiv.org/abs/2...

    Pith/arXiv arXiv 2025

  7. [7]

    arXiv preprint arXiv:2507.21802 (2025)

    Li, J., Cui, Y., Huang, T., Ma, Y., Fan, C., Yang, M., Zhong, Z.: Mixgrpo: Unlock- ing flow-based grpo efficiency with mixed ode-sde. arXiv preprint arXiv:2507.21802 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    arXiv preprint arXiv:2509.06040 (2025)

    Li, Y., Wang, Y., Zhu, Y., Zhao, Z., Lu, M., She, Q., Zhang, S.: Branchgrpo: Stable and efficient grpo with structured branching in diffusion models. arXiv preprint arXiv:2509.06040 (2025)

    arXiv 2025

  9. [9]

    arXiv preprint arXiv:2210.02747 (2022)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

    Pith/arXiv arXiv 2022

  10. [10]

    arXiv preprint arXiv:2505.05470 (2025)

    Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., Wan, P., Zhang, D., Ouyang, W.: Flow-grpo: Training flow matching models via online rl. arXiv preprint arXiv:2505.05470 (2025)

    Pith/arXiv arXiv 2025

  11. [11]

    arXiv preprint arXiv:2501.13918 (2025)

    Liu, J., Liu, G., Liang, J., Yuan, Z., Liu, X., Zheng, M., Wu, X., Wang, Q., Qin, W., Xia, M., et al.: Improving video generation with human feedback. arXiv preprint arXiv:2501.13918 (2025)

    Pith/arXiv arXiv 2025

  12. [12]

    Advances in neural information processing systems35, 5775–5787 (2022)

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in neural information processing systems35, 5775–5787 (2022)

    2022

  13. [13]

    Advances in Neural Information Processing Systems37, 118428–118455 (2024)

    Ma, B., ZONG, Z., Song, G., Li, H., Liu, Y.: Exploring the role of large language models in prompt encoding for diffusion models. Advances in Neural Information Processing Systems37, 118428–118455 (2024)

    2024

  14. [14]

    Advances in neural information processing sys- tems35, 27730–27744 (2022)

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in neural information processing sys- tems35, 27730–27744 (2022)

    2022

  15. [15]

    arXiv preprint arXiv:2410.13720 (2024)

    Polyak, A., Zohar, A., Brown, A., Tjandra, A., Sinha, A., Lee, A., Vyas, A., Shi, B., Ma, C.Y., Chuang, C.Y., et al.: Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720 (2024)

    Pith/arXiv arXiv 2024

  16. [16]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022) AR-CoPO 17

    2022

  17. [17]

    In: International conference on machine learning

    Schulman, J., Levine, S., Abbeel, P., Jordan, M., Moritz, P.: Trust region policy optimization. In: International conference on machine learning. pp. 1889–1897. PMLR (2015)

    2015

  18. [18]

    arXiv preprint arXiv:1707.06347 (2017)

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    Pith/arXiv arXiv 2017

  19. [19]

    arXiv preprint arXiv:2402.03300 (2024)

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    Pith/arXiv arXiv 2024

  20. [20]

    arXiv preprint arXiv:2010.02502 (2020)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    Pith/arXiv arXiv 2010

  21. [21]

    Ad- vances in neural information processing systems33, 3008–3021 (2020)

    Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Ad- vances in neural information processing systems33, 3008–3021 (2020)

    2020

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Purushwalkam, S., Ermon, S., Xiong, C., Joty, S., Naik, N.: Diffusion model alignment using direct preference optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8228–8238 (2024)

    2024

  23. [23]

    arXiv preprint arXiv:2508.20751 (2025)

    Wang, Y., Li, Z., Zang, Y., Zhou, Y., Bu, J., Wang, C., Lu, Q., Jin, C., Wang, J.: Pref-grpo: Pairwise preference reward-based grpo for stable text-to-image rein- forcement learning. arXiv preprint arXiv:2508.20751 (2025)

    Pith/arXiv arXiv 2025

  24. [24]

    arXiv preprint arXiv:2511.18870 (2025)

    Wu, B., Zou, C., Li, C., Huang, D., Yang, F., Tan, H., Peng, J., Wu, J., Xiong, J., Jiang, J., et al.: Hunyuanvideo 1.5 technical report. arXiv preprint arXiv:2511.18870 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    In: European Con- ference on Computer Vision

    Wu, X., Hao, Y., Zhang, M., Sun, K., Huang, Z., Song, G., Liu, Y., Li, H.: Deep reward supervisions for tuning text-to-image diffusion models. In: European Con- ference on Computer Vision. pp. 108–124. Springer (2024)

    2024

  26. [26]

    arXiv preprint arXiv:2505.07818 (2025)

    Xue, Z., Wu, J., Gao, Y., Kong, F., Zhu, L., Chen, M., Liu, Z., Liu, W., Guo, Q., Huang, W., et al.: Dancegrpo: Unleashing grpo on visual generation. arXiv preprint arXiv:2505.07818 (2025)

    Pith/arXiv arXiv 2025

  27. [27]

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

    2025

  28. [28]

    In: NeurIPS (2024)

    Yin, T., Gharbi, M., Park, T., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T.: Improved distribution matching distillation for fast image synthesis. In: NeurIPS (2024)

    2024

  29. [29]

    In: CVPR (2024)

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: CVPR (2024)

    2024

  30. [30]

    In: CVPR (2025)

    Yin, T., Zhang, Q., Zhang, R., Freeman, W.T., Durand, F., Shechtman, E., Huang, X.: From slow bidirectional to fast autoregressive video diffusion models. In: CVPR (2025)

    2025

  31. [31]

    arXiv preprint arXiv:2508.15761 (2025)

    Zhang, Y., Yang, H., Zhang, Y., Hu, Y., Zhu, F., Lin, C., Mei, X., Jiang, Y., Peng, B., Yuan, Z.: Waver: Wave your way to lifelike video generation. arXiv preprint arXiv:2508.15761 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    arXiv preprint arXiv:2412.20404 (2024)

    Zheng, Z., Peng, X., Yang, T., Shen, C., Li, S., Liu, H., Zhou, Y., Li, T., You, Y.: Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    arXiv preprint arXiv:2602.02214 (2026) 18 preprint

    Zhu, H., Zhao, M., He, G., Su, H., Li, C., Zhu, J.: Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video genera- tion. arXiv preprint arXiv:2602.02214 (2026) 18 preprint

    Pith/arXiv arXiv 2026

  34. [34]

    In: Forty-second International Conference on Machine Learning (2024) AR-CoPO 19 (a)Prompt: A banana on the top of an apple, front view

    Zong, Z., Jiang, D., Ma, B., Song, G., Shao, H., Shen, D., Liu, Y., Li, H.: Easyref: Omni-generalized group image reference for diffusion models via multimodal llm. In: Forty-second International Conference on Machine Learning (2024) AR-CoPO 19 (a)Prompt: A banana on the top of an apple, front view. (b)Prompt: A boat sailing leisurely along the Seine Rive...

    2024