AnyFlow introduces the first any-step video diffusion distillation framework using two-time flow map formulation. It learns arbitrary time pair transitions through forward flow map training and on-policy distillation, achieving strong few-step performance while continuing to improve with more steps.
论文原始摘要
Few-step video generation has been significantly advanced by consistency distillation. However, the performance of consistency-distilled models often degrades as more sampling steps are allocated at test time, limiting their effectiveness for any-step video diffusion. This limitation arises because consistency distillation replaces the original probability-flow ODE trajectory with a consistency-sampling trajectory, weakening the desirable test-time scaling behavior of ODE sampling. To address this limitation, we introduce AnyFlow, the first any-step video diffusion distillation framework based on flow maps. Instead of distilling a model for only a few fixed sampling steps, AnyFlow optimizes the full ODE sampling trajectory. To this end, we shift the distillation target from endpoint consistency mapping $(z_{t}\rightarrow z_{0})$ to flow-map transition learning $(z_{t}\rightarrow z_{r})$ over arbitrary time intervals. We further propose Flow Map Backward Simulation, which decomposes a full Euler rollout into shortcut flow-map transitions, enabling efficient on-policy distillation that reduces test-time errors (i.e., discretization error in few-step sampling and exposure bias in causal generation). Extensive experiments across both bidirectional and causal architectures, at scales ranging from 1.3B to 14B parameters, demonstrate that AnyFlow achieves performance matches or surpasses consistency-based counterparts in the few-step regime, while scaling with sampling step budgets.
Paper Collector 中文速览
首个基于流图的任意步视频扩散蒸馏框架
方法概述
通过Flow Map Backward Simulation将完整Euler展开分解为快捷流图转换，实现高效在策略蒸馏，减少测试时离散化误差和暴露偏差
核心贡献
提出AnyFlow框架，将蒸馏目标从端点一致性映射改为任意时间间隔的流图转换学习，实现全ODE采样轨迹优化
原始来源与核验范围
核验仅覆盖论文正文中的主张、证据片段和参考文献关系。NGJOO 未独立运行作者代码、重做实验或验证真实部署效果；页面中的可复现性分数是文档完整度评估，不是复现实验结果。
核心问题
How can video diffusion models support flexible any-step generation that maintains strong few-step performance while continuing to improve with more sampling steps, overcoming the degradation problem of existing consistency-based distillation methods?
核心方法
AnyFlow employs a two-stage training pipeline: forward flow map training fine-tunes pretrained models to learn transitions between arbitrary time pairs f_θ(z_t, t, r) → z_r, followed by on-policy distillation using Distribution Matching Distillation with flow map backward simulation. The approach introduces interpolated timestep conditioning, guidance-fused training, and adaptive loss reweighting for stable video diffusion adaptation.
方法组件
- Consistency Models learn direct mappings from intermediate states to endpoints along PF-ODE trajectories.
- Multi-step generation in CMs involves repeated noise injection and iterative denoising.
- Improvements to CMs include annealed time schedules, segmented objectives, and score distillation regularization.
- AnyFlow differs from CM-based approaches by using flow map distillation for any-step generation.
- Flow map models learn transitions between arbitrary time pairs along PF-ODE trajectories.
- The formulation unifies consistency modeling (r=0) and standard flow matching (t=r).
- Training requires accurate transition learning across arbitrary time pairs, which is challenging.
- MeanFlow uses JVPs which are difficult to scale under FSDP, leading to numerical approximations or algebraic reformulations.
- AnyFlow builds on these advances for scalable any-step video diffusion.
论点验证
The paper describes the on-policy distillation stage with DMD-based training in detail (p_47-51), provides Algorithm 2, and shows experimental validation in Table 2 demonstrating performance improvements.
The paper states in p_6 that a single AnyFlow model supports T2V, I2V, and V2V in the causal setting. VBench-I2V scores are reported. However, V2V is mentioned but not explicitly evaluated with metrics.
Specific numerical results are reported in p_6: AnyFlow-FAR reaches 84.05 at 4 NFEs and improves to 84.41. These are concrete VBench scores that can be verified.
Specific numerical comparison provided: 87.87 VBench-I2V score at 4 NFEs vs Wan2.1-I2V-14B at 50×2 NFEs (87.71). Clear quantitative evidence.
Specific numerical comparison: AnyFlow 84.04 at 4 NFEs vs rCM-14B 83.73 at 4 NFEs. Clear quantitative evidence for the 14B bidirectional T2V setting.
This is a straightforward architectural description. The paper clearly describes the two-stage approach throughout Section 4, with detailed explanations in Sections 4.2.1 and 4.2.2.
The design choice is clearly stated in p_10 and validated through experimental results in Table 2 showing AnyFlow's stronger any-step performance compared to consistency-based methods.
The paper provides detailed explanation of the decomposition strategy in p_49, describing how trajectories are decomposed into three segments with shortcut transitions based on composition property.
Experimental validation across both architectures (bidirectional and causal) and multiple NFE settings is provided in Table 2 and throughout Section 5. The paper demonstrates performance across 4, 8, 16, 32 NFEs.
This is a clear methodological description. The paper explicitly states this design choice in p_33 and provides the mathematical formulation.
Clear methodological description stated in p_33. The mathematical formulation fθ(zt, t, r) ↦ zr is provided, showing transitions between arbitrary time pairs rather than just z0.
The paper identifies specific error types (discretization error, exposure bias) and provides evidence in Figure 3 and Table 2. The claim is specific about the types of errors and when they occur.
Clear architectural description with detailed explanations in Sections 4.2.1 and 4.2.2. This is a straightforward methodological claim.
Clear methodological description in p_35. The paper specifies the training data (teacher-synthesized) and the objective (learning stable transition operator).
... 共 63 个论点
可复现性评估
较低可复现性 (0%)
缺失的复现细节
- Code is not available - no implementation details can be verified
- Training data (256K synthetic prompt-video pairs) is not available
- Random seeds are not specified for reproducibility
- Hardware specifications not provided (GPU type, count, memory requirements)
- Training time and compute costs not reported
- AdamW optimizer hyperparameters incomplete (weight decay, beta values)
- Learning rate schedule and warmup configuration not specified
- Gradient clipping settings not mentioned
- Details on synthetic data generation process missing (prompt sources, filtering criteria)
- Exact on-policy objective formulation and weighting not fully specified
局限与证据边界
- Forward flow map training alone (e.g., MeanFlow) is insufficient for strong test-time performance in video generation
- the base distilled model, AnyFlow-FAR-Wan2.1-1.3B, still struggles with identity preservation in robotics videos and trajectory accuracy in driving scenes
- The primary limitation of our method is its reliance on external datasets for flow map training
- Even when synthetic data are used, the training distribution may still differ from that of the base model, which can introduce mild distribution shift, such as smoother textures
- there is still substantial room for improving the training methodologies
本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。
分析时间：2026-05-16T01:22:21+00:00 · 数据来源：Paper Collector