pith. sign in

arxiv: 2603.08403 · v3 · pith:H72LKVXKnew · submitted 2026-03-09 · 💻 cs.CV

SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents

Yu Yang , Yue Liao , Jianbiao Mei , Baisen Wang , Xuemeng Yang , Licheng Wen , Jiangning Zhang , Xiangtai Li
show 6 more authors
Liang Lv Hanlin Chen Botian Shi Yong Liu Shuicheng Yan Gim Hee Lee
This is my paper

Pith reviewed 2026-05-22 11:11 UTC · model grok-4.3

classification 💻 cs.CV
keywords action-conditioned video generationlong-horizon video synthesisplanning agentscritic feedbackself-evolutiontemporal coherenceclosed-loop generation
0
0 comments X

The pith

SPIRAL uses planning and critic agents in a closed loop to generate consistent long-horizon action videos.

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

The paper presents a framework that decomposes complex action instructions into sequential sub-actions, generates video segments conditioned on prior memory, and employs a critic to evaluate and correct errors in each step. This closed-loop process replaces the open-loop approach common in existing models, which often produce incomplete actions or temporal drift over long sequences. If the method works as described, it would allow video generators to maintain scene consistency and follow extended procedural instructions more reliably. The framework also feeds the planning and critique signals back into self-training to improve the underlying generator.

Core claim

SPIRAL instantiates a think-act-reflect process where a PlanAgent decomposes high-level goals into sub-actions that condition a VideoGenerator to synthesize each segment with memory context, while a CriticAgent evaluates the segments to supply corrective feedback, and the resulting signals drive GRPO-based post-training for self-evolution of long-horizon consistency.

What carries the argument

The closed-loop think-act-reflect cycle that combines PlanAgent decomposition, memory-conditioned segment generation, CriticAgent evaluation, and GRPO post-training.

If this is right

  • Consistent gains in action quality and temporal coherence on ActVideoGen-Bench and VBench.
  • Improved performance when the same closed-loop design is applied to multiple different text-to-image-to-video backbones.
  • Further gains from using PlanAgent and CriticAgent signals for GRPO-based self-evolution of the generator.

Where Pith is reading between the lines

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

  • The same reflective loop structure could be tested on other long-sequence generation problems such as audio or robotic motion planning.
  • If the critic feedback proves stable, it might reduce the need for large amounts of human-labeled long-horizon video data.

Load-bearing premise

The critic can reliably spot action errors and drift and turn those detections into training signals that improve the generator without creating new instabilities.

What would settle it

Videos of complex multi-step tasks that still show incomplete action execution or accumulating temporal drift after several critic feedback rounds.

Figures

Figures reproduced from arXiv: 2603.08403 by Baisen Wang, Botian Shi, Gim Hee Lee, Hanlin Chen, Jianbiao Mei, Jiangning Zhang, Liang Lv, Licheng Wen, Shuicheng Yan, Xiangtai Li, Xuemeng Yang, Yong Liu, Yue Liao, Yu Yang.

Figure 1
Figure 1. Figure 1: Action World Models (ActWM): Challenges and Solution. (a) General TI2V handles instructions in a one-shot, open-loop manner, leading to incomplete action execution, hallucinated motions, and temporal drift. (b) We introduce a closed￾loop think–act–reflect formulation, where generation proceeds step by step under explicit planning and feedback, enabling actions to be executed persistently and corrected over… view at source ↗
Figure 2
Figure 2. Figure 2: Framework Overview. (a) Closed-Loop Think-Act-Reflect: PlanAgent decomposes abstract goals into atomic plans for ActWMs execution, while CriticAgent evaluates videos to trigger dual-level feedback (Inner/Outer Loops) for refinement; (b) Progressive￾Evolution GRPO: WorldModel generates group rollouts guided by PlanAgent, leveraging CriticAgent rewards for policy optimization. and real-world settings (Yang e… view at source ↗
Figure 3
Figure 3. Figure 3: Overview and Statistics of ActWM-Dataset. (a) A structured data annotation example featuring Goal, CoT, and step-wise Video-Action-Critic tuples; (b-f) Distribution analysis across video duration, step length, scene types, perspectives, and action keywords. KL divergence, ϵ and β are hyper-parameters controlling policy clipping and regularization strength, respectively. This closed-loop optimization effect… view at source ↗
Figure 4
Figure 4. Figure 4: PlanAgent Robustness to Task Length. Comparison of accuracy (%) across varying horizons; incorporating World Memory (PlanAgent + Mem.) maintains stable performance. Model uses LongLive (T2V) and SVI (I2V) as base genera￾tors; we apply Streaming Long-Tuning on ActWM-Dataset to equip them with action-following capabilities. Evaluation Benchmarks. We evaluate different compo￾nents of our framework on three be… view at source ↗
Figure 6
Figure 6. Figure 6: World Model Performance across Difficulties. Our framework maintains high stability across all levels, whereas base￾lines degrade significantly on long-horizon and complex tasks [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: CriticAgent Discriminative Capability. Incorporating RM (SFT+RM) induces highly polarized scores, providing sharper signals to better penalize failure executions. strating the necessity of structured supervision. DPO further enhances performance (+5.43%), proving particularly effec￾tive in reducing physically infeasible hallucinations. Robustness to Long Horizons. We observe that without memory, performanc… view at source ↗
Figure 8
Figure 8. Figure 8: Closed-loop and GRPO Facilitate Performance Gains. By distilling the Think-Act-Reflect paradigm into intrinsic weights, our framework achieves comprehensive gains across dimensions. motions, structural collapse, or physically implausible tran￾sitions. With our agents, complex behaviors are decomposed into phased sub-tasks with corrective feedback, improving both action completion and global temporal consis… view at source ↗
Figure 9
Figure 9. Figure 9: Visualizations comparing models with and without our framework. One-shot baseline often suffers from structural collapse and incomplete action steps; SPIRAL decomposes complex behaviors into phased action steps with corrective feedback, ensuring both action completion and global temporal consistency. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualizations comparing models with and without our framework. One-shot baseline often suffers from structural collapse and incomplete action steps; SPIRAL decomposes complex behaviors into phased action steps with corrective feedback, ensuring both action completion and global temporal consistency. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparative analysis of action sequences with and without GRPO training. The left columns depict issues like physical violation, motionlessness, and inconsistencies without GRPO. The right columns demonstrate improved action completion and physical plausibility with GRPO, illustrating enhanced coherence in tasks such as running, cooking, and assembling. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
read the original abstract

Long-horizon action-conditioned video generation aims to synthesize temporally coherent videos that follow complex action instructions over extended horizons, requiring procedural ordering, persistent action execution, and scene consistency beyond conventional TI2V's short-term fidelity. Existing single-shot video generation models typically operate in an open-loop manner, leading to incomplete action execution, hallucinated motions, and temporal drift. To address this, we propose SPIRAL, a closed-loop framework that performs sequential planning and iterative reflection for action-conditioned long-horizon video generation. Specifically, SPIRAL instantiates a think-act-reflect process: a PlanAgent decomposes high-level goals into sub-actions, which condition a VideoGenerator to synthesize each segment alongside a memory context, while a CriticAgent evaluates intermediate video segments to provide corrective feedback for iterative refinement. This closed-loop design further supports self-evolution by utilizing PlanAgent-proposed actions and CriticAgent-derived rewards for GRPO-based post-training to enhance the video generator's long-horizon consistency. Moreover, we introduce ActVideoGen-Dataset for task-specific training, and establish ActVideoGen-Bench as a dedicated evaluation suite for measuring action quality and temporal coherence. Experiments across multiple TI2V backbones alongside the self-evolving strategy show consistent gains on ActVideoGen-Bench and VBench, demonstrating the effectiveness of SPIRAL.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes SPIRAL, a closed-loop framework for long-horizon action-conditioned video generation. It decomposes goals via a PlanAgent, generates memory-conditioned video segments with a VideoGenerator, uses a CriticAgent for iterative feedback and refinement, and applies GRPO-based self-evolution for post-training to improve consistency. The work introduces ActVideoGen-Dataset and ActVideoGen-Bench, reporting consistent gains in action quality and temporal coherence across multiple TI2V backbones on the new benchmark and VBench.

Significance. If the empirical results prove robust, the closed-loop think-act-reflect design with self-evolution could meaningfully advance long-horizon video synthesis beyond open-loop TI2V limitations. The dedicated dataset, benchmark, and integration of planning/critique agents represent useful contributions that may influence agentic generative modeling.

major comments (2)
  1. [§3.3 (CriticAgent and reward formulation)] The central claim that CriticAgent feedback supplies reliable, corrective rewards for GRPO post-training (and thereby drives the reported gains in temporal coherence) is load-bearing. No quantitative validation of CriticAgent accuracy—such as precision/recall against ground-truth action labels, human ratings, or inter-annotator agreement—is provided to confirm that detected errors are not spurious or biased.
  2. [§5 (Experiments and results tables)] Experiments report consistent gains on ActVideoGen-Bench and VBench, yet the description lacks ablations isolating the contribution of CriticAgent feedback versus PlanAgent decomposition or memory conditioning alone. Without these controls or error bars, it is difficult to attribute improvements specifically to the closed-loop self-evolution mechanism.
minor comments (2)
  1. [§3.2] Clarify the precise memory update rule and conditioning mechanism when passing context between consecutive segments in the VideoGenerator.
  2. [§4.1] Add implementation details for GRPO reward scaling/clipping hyperparameters and training schedule to support reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. The points raised highlight important aspects for strengthening the claims around the CriticAgent and the attribution of gains in our closed-loop framework. We respond to each major comment below and will incorporate revisions as indicated.

read point-by-point responses

  1. Referee: [§3.3 (CriticAgent and reward formulation)] The central claim that CriticAgent feedback supplies reliable, corrective rewards for GRPO post-training (and thereby drives the reported gains in temporal coherence) is load-bearing. No quantitative validation of CriticAgent accuracy—such as precision/recall against ground-truth action labels, human ratings, or inter-annotator agreement—is provided to confirm that detected errors are not spurious or biased.

    Authors: We acknowledge that direct quantitative validation of CriticAgent accuracy would provide stronger support for the reliability of its feedback as rewards in GRPO. The current manuscript supports the overall effectiveness through end-to-end gains on ActVideoGen-Bench and VBench, but does not include explicit metrics for the CriticAgent itself. In the revised manuscript, we will add an evaluation of CriticAgent performance, reporting precision and recall against ground-truth action labels from the ActVideoGen-Dataset, along with human ratings on a sampled subset of outputs to assess bias or spurious detections. revision: yes

  2. Referee: [§5 (Experiments and results tables)] Experiments report consistent gains on ActVideoGen-Bench and VBench, yet the description lacks ablations isolating the contribution of CriticAgent feedback versus PlanAgent decomposition or memory conditioning alone. Without these controls or error bars, it is difficult to attribute improvements specifically to the closed-loop self-evolution mechanism.

    Authors: We agree that isolating the contributions of individual components would strengthen attribution of the observed gains specifically to the CriticAgent feedback and GRPO self-evolution. The existing experiments compare the full SPIRAL system against open-loop baselines across backbones, but lack targeted ablations. We will add these controls in the revision, including variants without CriticAgent feedback and without the self-evolution stage, while also reporting error bars or standard deviations from multiple runs for statistical robustness. revision: yes

Circularity Check

0 steps flagged

No circularity: SPIRAL applies standard agent/RL components to new video domain with independent empirical validation

full rationale

The paper's core contribution is a closed-loop think-act-reflect architecture using PlanAgent decomposition, memory-conditioned VideoGenerator segments, CriticAgent feedback, and GRPO post-training on a newly introduced ActVideoGen-Dataset and ActVideoGen-Bench. No equations, fitted parameters, or self-citations are presented as deriving the reported gains in action quality or temporal coherence; the improvements are shown empirically across multiple TI2V backbones on ActVideoGen-Bench and VBench. The derivation chain relies on externally standard planning, reflection, and RL techniques applied to long-horizon video generation without reducing any central claim to a tautology or self-referential fit. This is the expected non-circular outcome for an applied systems paper whose results remain falsifiable via the provided benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The framework depends on the effectiveness of newly introduced agent components whose performance is not independently verified outside this work and on the assumption that GRPO training yields stable long-horizon improvements.

free parameters (1)
  • GRPO reward scaling and clipping parameters
    Training hyperparameters that must be chosen or fitted to stabilize the self-evolution step.
axioms (1)
  • domain assumption Current multimodal models can serve as reliable PlanAgent and CriticAgent without additional architectural changes.
    Invoked when the paper states that the agents decompose goals and evaluate segments.
invented entities (2)
  • PlanAgent no independent evidence
    purpose: Decomposes high-level action goals into ordered sub-actions that condition the video generator.
    New component introduced to enable sequential planning.
  • CriticAgent no independent evidence
    purpose: Evaluates generated video segments and supplies corrective feedback for refinement.
    New component introduced to close the reflection loop.

pith-pipeline@v0.9.0 · 5814 in / 1435 out tokens · 39411 ms · 2026-05-22T11:11:42.475049+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

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

  1. AEL: Agent Evolving Learning for Open-Ended Environments

    cs.CL 2026-04 conditional novelty 7.0

    AEL uses a fast-timescale bandit for memory policy selection and slow-timescale LLM reflection for causal insights, achieving a Sharpe ratio of 2.13 on a 208-episode portfolio benchmark while showing that added mechan...

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages · cited by 1 Pith paper · 13 internal anchors

  1. [1]

    Cosmos World Foundation Model Platform for Physical AI

    Agarwal, N., Ali, A., Bala, M., Balaji, Y ., Barker, E., Cai, T., Chattopadhyay, P., Chen, Y ., Cui, Y ., Ding, Y ., et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Qwen3-VL Technical Report

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

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    SkyReels-V2: Infinite-length Film Generative Model

    Chen, G., Lin, D., Yang, J., Lin, C., Zhu, J., Fan, M., Zhang, H., Chen, S., Chen, Z., Ma, C., et al. Skyreels- v2: Infinite-length film generative model.arXiv preprint arXiv:2504.13074, 2025a. Chen, J., Zhao, Y ., Yu, J., Chu, R., Chen, J., Yang, S., Wang, X., Pan, Y ., Zhou, D., Ling, H., et al. Sana-video: Efficient video generation with block linear d...

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Thinking-while- generating: Interleaving textual reasoning throughout vi- sual generation.arXiv preprint arXiv:2511.16671, 2025a

    Guo, Z., Zhang, R., Li, H., Zhang, M., Chen, X., Wang, S., Feng, Y ., Pei, P., and Heng, P.-A. Thinking-while- generating: Interleaving textual reasoning throughout vi- sual generation.arXiv preprint arXiv:2511.16671, 2025a. 9 SPIRAL: A Closed-Loop Framework for Self-Improving Action World Models via Reflective Planning Agents Guo, Z., Zhang, R., Tong, C....

    work page arXiv

  5. [5]

    R-Zero: Self-Evolving Reasoning LLM from Zero Data

    Huang, C., Yu, W., Wang, X., Zhang, H., Li, Z., Li, R., Huang, J., Mi, H., and Yu, D. R-zero: Self- evolving reasoning llm from zero data.arXiv preprint arXiv:2508.05004,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

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

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Editthinker: Unlocking iterative reasoning for any image editor.arXiv preprint arXiv:2512.05965, 2025a

    Li, H., Zhang, M., Zheng, D., Guo, Z., Jia, Y ., Feng, K., Yu, H., Liu, Y ., Feng, Y ., Pei, P., et al. Editthinker: Unlocking iterative reasoning for any image editor.arXiv preprint arXiv:2512.05965, 2025a. Li, W., Pan, W., Luan, P.-C., Gao, Y ., and Alahi, A. Stable video infinity: Infinite-length video generation with error recycling.arXiv preprint arX...

    work page arXiv

  8. [8]

    X., Wan, X., Nakhost, H., Lee, C.-Y ., Pfister, T., and Arık, S

    Long, D. X., Wan, X., Nakhost, H., Lee, C.-Y ., Pfister, T., and Arık, S. ¨O. Vista: A test-time self-improving video generation agent.arXiv preprint arXiv:2510.15831,

    work page arXiv

  9. [9]

    Dreamforge: Motion-aware autoregressive video generation for multi-view driving scenes.arXiv preprint arXiv:2409.04003,

    Mei, J., Hu, T., Yang, X., Wen, L., Yang, Y ., Wei, T., Ma, Y ., Dou, M., Shi, B., and Liu, Y . Dreamforge: Motion-aware autoregressive video generation for multi-view driving scenes.arXiv preprint arXiv:2409.04003,

    work page arXiv

  10. [10]

    Open-Sora 2.0: Training a Commercial-Level Video Generation Model in $200k

    URL https://openai.com/ index/sora-2. Peng, X., Zheng, Z., Shen, C., Young, T., Guo, X., Wang, B., Xu, H., Liu, H., Jiang, M., Li, W., et al. Open-sora 2.0: Training a commercial-level video generation model in $200 k.arXiv preprint arXiv:2503.09642,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Hunyuan-gamecraft- 2: Instruction-following interactive game world model

    Tang, J., Liu, J., Li, J., Wu, L., Yang, H., Zhao, P., Gong, S., Yuan, X., Shao, S., and Lu, Q. Hunyuan-gamecraft- 2: Instruction-following interactive game world model. arXiv preprint arXiv:2511.23429,

    work page arXiv

  12. [12]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Yu, F., Zhao, H., Yang, J., et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    VideoVerse: Does Your T2V Generator Have World Model Capability to Synthesize Videos?

    Wang, Z., Wei, X., Li, B., Guo, Z., Zhang, J., Wei, H., Wang, K., and Zhang, L. Videoverse: How far is your t2v generator from a world model?arXiv preprint arXiv:2510.08398,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    EvolveR: Self-Evolving LLM Agents through an Experience-Driven Lifecycle

    10 SPIRAL: A Closed-Loop Framework for Self-Improving Action World Models via Reflective Planning Agents Wu, R., Wang, X., Mei, J., Cai, P., Fu, D., Yang, C., Wen, L., Yang, X., Shen, Y ., Wang, Y ., et al. Evolver: Self-evolving llm agents through an experience-driven lifecycle.arXiv preprint arXiv:2510.16079, 2025a. Wu, T., Yang, S., Po, R., Xu, Y ., Li...

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    DanceGRPO: Unleashing GRPO on Visual Generation

    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,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    LongLive: Real-time Interactive Long Video Generation

    Yang, S., Huang, W., Chu, R., Xiao, Y ., Zhao, Y ., Wang, X., Li, M., Xie, E., Chen, Y ., Lu, Y ., et al. Longlive: Real- time interactive long video generation.arXiv preprint arXiv:2509.22622, 2025a. Yang, Y ., Liang, A., Mei, J., Ma, Y ., Liu, Y ., and Lee, G. H. X-scene: Large-scale driving scene generation with high fidelity and flexible controllabili...

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models

    Zeng, A., Lv, X., Zheng, Q., Hou, Z., Chen, B., Xie, C., Wang, C., Yin, D., Zeng, H., Zhang, J., et al. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models. arXiv preprint arXiv:2508.06471, 2025a. Zeng, Q., Cai, K., Chen, R., Lv, Q., and Wang, K. Coagent: Collaborative planning and consistency agent for coher- ent video generation.arXiv prepri...

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Absolute Zero: Reinforced Self-play Reasoning with Zero Data

    Zhao, A., Wu, Y ., Yue, Y ., Wu, T., Xu, Q., Lin, M., Wang, S., Wu, Q., Zheng, Z., and Huang, G. Absolute zero: Reinforced self-play reasoning with zero data.arXiv preprint arXiv:2505.03335,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    ForSFT Data, we implement a Teacher-Student Distillation pipeline based on the VideoVerse benchmark (Wang et al., 2025)

    Reward Data Construction.We employ a hybrid data construction strategy that leverages distinct benchmarks to ensure both reasoning depth and discriminative sensitivity. ForSFT Data, we implement a Teacher-Student Distillation pipeline based on the VideoVerse benchmark (Wang et al., 2025). Adhering to the VideoVerse protocol, we synthesize a diverse corpus...

    work page 2025

  20. [20]

    Inspired by recent stochastic sampling theories, we introduce a diffusion term into the flow matching process

    to adopt a reverse-timeStochastic Differential Equation (SDE)formulation. Inspired by recent stochastic sampling theories, we introduce a diffusion term into the flow matching process. The reverse SDE for generation is given by: dzτ = uτ (zτ )− 1 2 η2 τ ∇z logp τ (zτ ) | {z } Drift Term dτ+η τ dw|{z} Diffusion Term (7) where uτ is the velocity predicted b...

    work page 2025