acceptodds
Under review as a conference paper at ICLR 2027

FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching

Abstract

Extending the generation horizon of video diffusion models to long sequences remains a long-standing and important challenge. Existing training-free approaches either extend bidirectional models through architecture-specific modifications that degrade over long horizons, or rely on autoregressive generation that accumulates drift through exposure bias and produces repetitive motion. We instead formulate long video generation as an inverse problem: a long video is decomposed into a chain of short chunks, each constrained to be consistent with its neighbor's denoised estimate, which serves as the measurement. Following diffusion inverse solvers, we solve this problem with a manifold-constrained update on the denoised estimates, which we call Tweedie matching; it reduces to a closed-form per-frame solution and enforces temporal consistency while restricting the correction to the clean data manifold of the pretrained model. Because deterministic ODE trajectories tend to revert to their independent paths, we further propose stochastic early-phase sampling, which injects fresh noise after each correction in the high-noise phase to synchronize the windows, before switching to deterministic sampling to preserve fine-grained visual fidelity. The resulting framework is training-free and architecture-agnostic. Applied to various video generation models, it generates videos several times longer than the native window while outperforming both training-free and autoregressive baselines on VBench and long-range drift metrics, and extends without fine-tuning to audio-video joint generation and text-to-3DGS.

Primary area:
generative models
Keywords:
Video generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Rjm3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper does training-free long video generation. It runs a pretrained flow model on overlapping chunks and pulls neighboring chunks together during sampling. There are two pieces: Tweedie matching blends each chunk's x^0∣t\hat{x}_{0|t} prediction with its neighbor's on the overlap. With the linear λj\lambda_j schedule this is really a crossfade. Stochastic early-phase sampling re-noises with fresh noise, then switches to plain ODE. The whole thing is framed as a DDS-style inverse problem. Experiments use Wan2.1-1.3B and LTX-2 at 30s/60s, with VBench, drift metrics, and a user study. There are also demos on audio-video (LTX-2) and text-to-3DGS (VIST3A).

Critical strengths

It's simple and actually plug-and-play. Nothing in the backbone changes, so it runs on LTX-2 audio+video and VIST3A with no retraining. AR distillation methods really can't do that.

The motion results convinced me. At 60s, the AR baselines distilled from Wan drop to 0.41–0.53 dynamic degree, while FlowLong keeps 0.82. Figs. 7–9 back this up. The authors are also upfront that VBench consistency metrics reward near-static videos, which I appreciate.

Memory doesn't grow with length: 21GB vs. 45–51GB for RIFLEx/UltraViCo. It also works with a 4-step backbone.

Critical weaknesses
  1. A.4 doesn't match App. C, and this undercuts the main ablation. In A.4, all chunks share one global buffer X^0∣t\hat{X}_{0|t} and XsX_s is re-sliced from it. So neighboring chunks have the same noisy latent on the overlap. App. C's argument assumes the opposite (xtk≠xtk+1x_t^{k} \neq x_t^{k+1}, and 1/21/\sqrt{2} noise shrinkage). C's last sentence even admits that with a shared latent, fusing \hat{x}_{0|t} and fusing xsx_s are the same thing. So where does the Table 3 gap between Ours and xtx_t matching (0.63590.6359 vs. 0.58620.5862) actually come from?

  2. The novelty is overstated, and close prior work is missing. With the linear ramp λj\lambda_j, Tweedie matching is just a crossfade of \hat{x}_{0|t} across overlapping windows. The "measurement" is the neighbor's estimate, which itself gets updated, so this is really temporal co-denoising rather than an inverse problem. With ηt=1\eta_t = 1, the early phase reduces to "predict x^0∣t\hat{x}_{0|t}, re-noise with fresh ϵ\epsilon". This is very close to SynCoS, Gen-L-Video and SyncTweedies. SynCoS and Gen-L-Video aren't cited, and SyncTweedies gets one line in App. C. I'd like to see a comparison with at least SynCoS or FIFO-Diffusion.

  3. The VBench "Overall" score is mostly Dynamic Degree. Drop that one dimension and average the other six at 60s: Ours ≈0.826\approx 0.826, vs. Deep-Forcing 0.8420.842 and LongLive 0.8530.853. In Table 2, FlowLong also trails the strong AR baselines on Visual Stability and Content Retention. The extra motion is a real win, but "outperforming" in the abstract should be "on par, with much more motion".

Questions for authors
  1. How was xtx_t matching implemented, given the shared buffer in A.4? Can you confirm that in the ODE-only phase (ηt=0\eta_t = 0) it's identical to Ours, as the last sentence of App. C implies?
  2. How does FlowLong differ from SynCoS and Gen-L-Video, both conceptually and in experiments?
  3. What (F,O,S)(F, O, S) and KK did you use for Wan2.1?
  4. For the user study: how many participants, how many prompts, and how many judgments per pair?
  5. How sensitive is the method to t∗t^* and OO?
  6. What happens at 2–5 minutes? The constraint is purely local, and Content Retention is already tied for lowest at 60s, so I'd expect identity drift.
3 comments
AuthorsOpenReview ↗

Response by Authors (1/2)

W1 & Q1, Q3

W1. "A.4 doesn't match App. C, and this undercuts the main ablation. …"

Q1. "How was x_t matching implemented, given the shared buffer in A.4? …"

Q3. "What … and … did you use for Wan2.1?"

Thank you for pointing out this inconsistency. Appendix A.4 was incorrectly written, and we have revised it to be consistent with Appendix C. Each chunk keeps its own noisy state, initialized from independent noise. In this setting, fusing clean estimates before the update and fusing noisy latents after the update are not equivalent, and the fusion-space comparison in Table 3 is made under this setting. The x_t matching baseline is given as Algorithm 2 in Appendix G; since the chunks start from independent noise, it is not identical to ours in the ODE-only phase either. For Wan 2.1, we use t* = 0.555 with 30 sampling steps and flow shift 5; each chunk has F = 21 latent frames with an overlap of O = 11 and a stride of S = 10, and λ_j increases linearly over the overlap (Appendix A.1).

W2 & Q2

W2. "The novelty is overstated, and close prior work is missing. …"

Q2. "How does FlowLong differ from SynCoS and Gen-L-Video …?"

(1) Inverse problem vs. co-denoising. At each sampling step, the neighboring estimate is held fixed, and the weighted interpolation in Eq. (16) is the closed-form solution of the guidance loss in Eq. (12) for that measurement. The loss has the form 12∥y−Ax∥2\frac{1}{2}\lVert y-Ax\rVert^2, where A selects frames and y is a measurement, and y can be any observation. We additionally apply the same formulation with real measurements that stay fixed throughout sampling: long-video extension, where a real video prefix is the measurement (Fig. 14), and long-video frame interpolation, where observed clips guide the missing intervals (Fig. 13); both are in Appendix E. This extension was not addressed using the existing methods at inference time. The corresponding videos for both tasks are available on the revised anonymous project page. In these tasks, y does not change during sampling, so they are inverse problems in the standard sense. Co-denoising methods average the predictions of overlapping windows and have no measurement term.

(2) Differences from SynCoS, Gen-L-Video, and SyncTweedies. Gen-L-Video fuses the noisy latents of overlapping clips, and SyncTweedies averages Tweedie estimates; both use deterministic sampling throughout, and SyncTweedies is demonstrated on spatial synchronization (panoramas, textures, 3D). SynCoS adds a score-distillation optimization of the latent in the early steps, which requires iterative gradient updates at each of those steps. FlowLong differs in three ways. First, FlowLong is formulated as an inverse problem with a measurement y, so the same sampler extends to real observations: long video interpolation from observed clips (Fig. 13) and long video extension from a given prefix (Fig. 14). The methods above synchronize the predictions of overlapping windows and have no measurement term, so they do not extend to these applications. Second, the update of FlowLong is a closed-form solution of the guidance loss, with no optimization and no additional network evaluations. Third, we show that deterministic synchronization is not sufficient for temporal chunks: a deterministic update retains only 4–7% of each fusion correction, whereas a stochastic update retains 26–33% (Appendix D), which is why FlowLong injects noise in the early phase.

(3) Comparison with FIFO-Diffusion. Following the reviewer’s suggestion, we evaluated FIFO-Diffusion on Wan 2.1-1.3B and added it to Table 1. FIFO-Diffusion reaches a Dynamic Degree comparable to ours (0.8113 / 0.8252 vs. 0.7800 / 0.8200 at 30s / 60s), but with lower aesthetic quality (0.4948 / 0.4887 vs. 0.5777 / 0.5738) and imaging quality (0.5403 / 0.5361 vs. 0.6368 / 0.6391), and FlowLong achieves the higher Overall score at both lengths (0.8233 vs. 0.8032 and 0.8251 vs. 0.8014). SynCoS has no public implementation to our knowledge, so we discuss it in the related work. We have added SynCoS and Gen-L-Video to the related work and expanded the discussion of SyncTweedies there.

Table R1. FIFO-Diffusion vs. FlowLong on Wan 2.1-1.3B (VBench).

Aesthetic ↑Background ↑Dynamic ↑Imaging ↑Motion Smooth. ↑Subject ↑Temporal Flicker. ↑Overall ↑
FIFO-Diffusion (30s)0.49480.93980.81130.54030.96990.91580.95030.8032
FlowLong (30s)0.57770.93050.78000.63680.98770.87510.97530.8233
FIFO-Diffusion (60s)0.48870.93690.82520.53610.96850.90710.94770.8014
FlowLong (60s)0.57380.92130.82000.63910.98690.86050.97380.8251
AuthorsOpenReview ↗

Response by Authors (2/2)

W3

W3. "The VBench "Overall" score is mostly Dynamic Degree. …"

(1) VBench without Dynamic Degree. The Overall gain comes mainly from Dynamic Degree, and this is by design. The main weakness of autoregressive long-video models is that their motion fades into near-static content as the video gets longer, and FlowLong is designed to solve exactly this. Dynamic Degree is therefore the metric that measures our target problem, not an auxiliary dimension. Following the reviewer’s suggestion, we report the mean of the other six dimensions without Dynamic Degree.

Table R2. VBench mean without Dynamic Degree.

CausVidSelf-ForcingDeep-Forcing∞-RoPELongLiveFlowLong (Ours)
6-dim mean(30s)0.82960.83080.83990.84340.85450.8305
6-dim mean(60s)0.82010.80600.84210.84030.85280.8259

(2) Long-range drift (Table 2). FlowLong stays within the range of the AR baselines (30s: 0.8305 vs. 0.8296–0.8545; 60s: 0.8259 vs. 0.8060–0.8528). FlowLong roughly doubles Dynamic Degree (0.82 vs. 0.41–0.53 at 60s) while keeping the other dimensions at the same level. We also note that this comparison favors static methods: four of the remaining six dimensions (Subject Consistency, Background Consistency, Motion Smoothness, Temporal Flickering) measure frame-to-frame consistency, which near-static videos maximize. VBench [1,2] notes that static scenes can “cheat” these temporal consistency scores, which is why Dynamic Degree is a separate axis, and Self-Forcing++ [3] attributes the high temporal scores of AR baselines to motion stagnation. Table 2. On long-range drift, FlowLong has the lowest Semantic Drift (0.0003 vs. 0.0009–0.0093), and its Visual Stability (97.2) and Content Retention (0.927) are within a small margin of the strongest AR baselines (97.6–99.2 and 0.927–0.938), which are explicitly designed against drift and produce about half the motion.

(3) Abstract. We have also revised the abstract. Instead of “outperforming both training-free and autoregressive baselines”, it now states that FlowLong achieves “the best Overall VBench score with substantially higher motion, and remains comparable on the other dimensions and on long-range drift metrics.”


Q4. The study has 63 participants and 10 prompts. Each participant made 10 pairwise comparisons, with each of the five baselines appearing twice, and judged every pair on three axes. This gives 126 judgments per baseline and 630 judgments in total for each axis.

Q5. t* trades cross-chunk consistency for detail, and Table 3 shows its two extremes: without a stochastic phase (full ODE) the Quality is 0.6075, and with stochasticity at every step (full SDE) it drops to 0.5298, against 0.6359 for our schedule. Appendix D shows that a stochastic update retains more of the fusion correction than a deterministic one at every measured noise level, so the benefit does not depend on a particular t.

Q6. Generating videos longer than two minutes is a very challenging task for bidirectional models alone. We consider it future work, as noted in the Conclusion.


[1] Huang et al., VBench: Comprehensive Benchmark Suite for Video Generative Models, CVPR 2024.

[2] Huang et al., VBench++: Comprehensive and Versatile Benchmark Suite for Video Generative Models, 2024.

[3] Cui et al., Self-Forcing++: Towards Minute-Scale High-Quality Video Generation, 2025.

We appreciate your time and constructive feedback throughout the discussion. As the discussion period is ending soon, we would like to kindly follow up on our response. We hope it has addressed your remaining concern about Table 1. If there is anything else we can clarify, we would be happy to answer before the discussion closes.

Reviewer_fdG7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

From my understanding, the paper proposes FlowLong which is an inference-time method for generating videos several times longer than a pretrained flow-based video model’s native window. The long video is split into K overlapping chunks. Each chunk starts from independent noise and is denoised by the unmodified model, and at every step the chunks are coupled through their overlaps.

Tweedie matching is framed as one DDS-style gradient step on a quadratic loss between chunk k’s overlap frames and chunk k+1’s denoised estimate. Stochastic early-phase sampling re-noises with fresh Gaussian noise (FlowDPS-style, η=1) for t ≥ t* and with the deterministic Euler step afterwards, so that chunks don’t fall back onto their independent ODE paths.

Experiments use Wan2.1-1.3B and LTX-2 on 100 MovieGen prompts at 30s and 60s. Comparisons are to RIFLEx, UltraViCo and five distilled AR models on seven VBench metircs.

Critical strengths

Simple, training-free, backbone-agnostic, with bounded memory. The method needs no distillation and no architecture hooks. It runs on Wan2.1, LTX-2 (including its audio stream) and a VIST3A pipeline essentially unchanged. Memory is fixed by the window size (21 GB vs. 45–51 GB for RIFLEx/UltraViCo in Table 5). It also avoids the self-conditioning feedback that makes distilled AR models lose motion over long horizons; the motion gap in Fig. 7–9 and the Dynamic Degree column are real effects. For practitioners who want long outputs from a new base model without retraining, this is a useful recipe, and it’s the main thing in the paper’s favor.

Critical weaknesses
  1. The quantitative superiority comes almost entirely from dynamic degree, which is a coarse, flow-thresholded metric that rewards instability as readily as meaningful motion, and there is no human study to separate the two. A high Dynamic Degree is expected here by construction. Each chunk is sampled from independent noise and coupled only to its neighbors, so content can wander across chunks, and RAFT-based dynamic degree counts that wandering as motion. The paper rightly notes that static videos can game VBench consistency metrics; the reverse is also true.
  2. RIFLEx and UltraViCo are run at roughly 6× extrapolation, likely beyond the factors they were designed for. Meanwhile, the AR baselines run 4 steps against FlowLong’s 30, and FlowLong takes about 13 min per 30s video vs. 77–113s.
  3. The core technique is already established, and the closest prior work is missing or under-credited. The paper claims to be “the first to formulate long video generation as an inverse problem”. But Eq. 16 is temporal co-denoising of overlapping clips with fusion in Tweedie (x̂₀) space. That is essentially Gen-L-Video’s idea, which treats a long video as overlapping short clips denoised in parallel. The x̂₀-space choice is precisely SyncTweedies’ finding that averaging the outputs of Tweedie’s formula synchronizes best across applications. FIFO-Diffusion and FreeNoise are other standard training-free long-video baselines. FIFO is cited but not compared, and FreeNoise is missing entirely.
Questions for authors
  1. Please report VBench Overall without Dynamic Degree, and add a cross-chunk identity metric, such as DINO or face-ID similarity between the first and last chunk or between non-adjacent chunks. That would separate “rich motion” from “content wandering between independently seeded chunks”.
  2. Can you compare directly against Gen-L-Video-style fusion and SyncTweedies-style x̂₀ averaging (no stochastic phase), FreeNoise, FIFO-Diffusion and SynCoS on Wan2.1-1.3B, using the same prompts and lengths? Can you also compare against StochSync-style maximal stochasticity in place of your binary schedule? What does FlowLong do that these don’t?
7 comments
AuthorsOpenReview ↗

Response by Authors (1/4)

W1 & Q1

W1. "The quantitative superiority comes almost entirely from dynamic degree, …"

Q1. "Please report VBench Overall without Dynamic Degree, and add a cross-chunk identity metric …"

To address the reviewer's concern that Dynamic Degree alone does not separate meaningful motion from instability, we provide four pieces of evidence.

(1) VBench without Dynamic Degree. The Overall gain comes mainly from Dynamic Degree, and this is by design. The main weakness of autoregressive long-video models is that their motion fades into near-static content as the video gets longer, and FlowLong is designed to solve exactly this. Dynamic Degree is therefore the metric that measures our target problem, not an auxiliary dimension. Following the reviewer’s suggestion, we report the mean of the other six dimensions without Dynamic Degree. It stays within the range of the AR baselines (30s: 0.8305 vs. 0.8296–0.8545; 60s: 0.8259 vs. 0.8060–0.8528), while Dynamic Degree is 1.6–2.0× higher (0.82 vs. 0.41–0.53 at 60s).

Table R1. VBench mean without Dynamic Degree.

CausVidSelf-ForcingDeep-Forcing∞-RoPELongLiveFlowLong (Ours)
6-dim mean (30s)0.82960.83080.83990.84340.85450.8305
6-dim mean (60s)0.82010.80600.84210.84030.85280.8259

(2) Long-range drift (Table 2). If content wandered across chunks, it would show up as drift. FlowLong has the lowest Semantic Drift (0.0003 vs. 0.0009–0.0093) and keeps Visual Stability (97.2) and Content Retention (0.927) on par with the strongest AR baselines.

(3) Human study (Appendix B.2). We ran a 2AFC study against all five AR baselines. Raters preferred FlowLong for motion diversity (92.4%) and also for semantic consistency (84.3%), and no baseline exceeds 31% on any axis. Raters would not judge wandering content as more semantically consistent.

(4) Prompt-object retention between the first and last chunk (Grounding DINO). Whole-frame feature similarity between distant chunks drops with any change of pose or camera framing, so it cannot separate rich motion from identity loss. We therefore measure identity at the object level with Grounding DINO, an open-set detector that returns the confidence that a named object is present. On each 60 s video, Subject Presence is the fraction of frames where the main subject is detected, and Object Retention is the detection confidence of the prompt objects in the last 5 s divided by that in the first 5 s. Motion is the top-5% optical-flow magnitude between neighboring frames, independent of the thresholded Dynamic Degree. Details are in Appendix B.3.

Table R2. Prompt-object retention on 60s videos (Grounding DINO).

CausVidSelf-ForcingDeep-Forcing∞-RoPELongLiveFlowLong (Ours)
Motion (optical flow) ↑2.822.792.922.862.413.95
Subject Presence ↑0.9070.9320.9320.9360.9360.948
Object Retention ↑0.9390.7980.9690.9901.0100.998

FlowLong produces 1.4–1.6× larger motion than the AR baselines, while keeping the prompt objects present most often and retaining them as well as the most static baselines. This shows that the higher Dynamic Degree comes from meaningful motion rather than content wandering.

AuthorsOpenReview ↗

Response by Authors (2/4)

W2

W2. "RIFLEx and UltraViCo are run at roughly 6× extrapolation, …"

(1) Extrapolation factor of RIFLEx and UltraViCo. Our goal is to generate videos of 30s, 60s, and longer without any training. To reach these lengths, RIFLEx and UltraViCo must extrapolate beyond the factors they were proposed for (about 2× and up to 4×). Their failures in this regime come from their design itself, because modifying positional frequencies or attention does not scale to lengths several times beyond the native window, such as the 30s setting in Table 1. FlowLong keeps the backbone at its native window length, so the target length is not bounded by an extrapolation factor.

(2) Number of sampling steps. The AR baselines are few-step generators obtained by distillation, so running them with more steps is not meaningful. For a step-matched comparison, we report FlowLong∗ in Appendix B.4, which applies FlowLong to a 4-step distilled backbone. FlowLong∗ keeps the quality (Overall 0.8244 vs. 0.8233 with 30 steps) and runs in 168 s, the same order as the AR baselines (77–113 s).

(3) Training cost. The AR baselines need a separate distillation for each backbone, which trains a student model from a bidirectional teacher on many GPUs. CausVid reports about 2 days on 64 H100 GPUs for its whole training, and LongLive reports 32 GPU-days (about 12 hours on 64 H100 GPUs) to fine-tune its 1.3B model with Wan2.1-T2V-14B as the teacher, following a Self-Forcing-style ODE initialization and DMD stage. FlowLong uses public models without any training, which is why it applies directly to Wan2.1, LTX-2, and text-to-3DGS. When a faster distilled backbone is available, FlowLong uses it as is, as shown by FlowLong∗.

AuthorsOpenReview ↗

Response by Authors (3/4)

W3 & Q2

W3. "The core technique is already established, and the closest prior work is missing or under-credited. …"

Q2. "Can you compare directly against Gen-L-Video-style fusion and SyncTweedies-style x̂₀ averaging …?"

Differences from SyncTweedies and contributions of the inverse-problem formulation: Compared with existing chunk synchronization methods (SyncTweedies, Gen-L-Video, MultiDiffusion, etc.), FlowLong is grounded in an inverse-problem formulation that (1) accommodates diverse real-video measurements, (2) enables flexible optimization designs, and (3) highlights the limitations of deterministic synchronization for long-video generation and the importance of early sampling stochasticity. This formulation also provides an optimization-based interpretation of chunk fusion. We elaborate below.

(1) Diverse downstream tasks with generalizable measurements. FlowLong formulates the guidance loss in Eq. (12) using the inverse-problem template 12∥y−Ax∥2\frac{1}{2}\lVert y-Ax\rVert^2, where selects observed frames, is the current chunk’s Tweedie estimate, and is a measurement. Although our main experiments use neighboring chunks’ denoised estimates as, this choice is not required. We additionally apply FlowLong on autoregressive long-video extension, using a real video prefix as the measurement (Fig. 14, Appendix E), and long-video frame interpolation, using multiple observed frames or clips to guide generation of the missing intervals (Fig. 13, Appendix E). The corresponding videos for both tasks have been added to our revised anonymous project page. These results demonstrate that the same formulation supports both chunk synchronization and conditioning on real observations, extending beyond the generation settings considered in the main SyncTweedies experiments.

(2) Customizable optimization design. Although Eq. (12) uses an L2 loss between denoised estimates, the framework accommodates additional regularizers and alternative losses. For video extension and interpolation, we find that anchoring generation to the low-frequency components of the observations substantially improves temporal consistency:

$ \ell_k(x;t)=\frac12\left( |y-Ax|^2+ |P_L(y)-P_L(x)|^2 \right), $

where extracts compatible low-frequency representations of the observation and generated chunk. Thus this additional regularizer pulls the generated chunk toward the observed low-frequency content (Eq. (26)). Therefore, weighted interpolation is one instance of our latent optimization framework, which supports broader optimization design choices for improved performance and task versatility, potentially including reward-based updates.

(3) Importance of Stochastic Early-Phase Sampling. We provide additional video examples demonstrating the importance of early sampling stochasticity (Fig. 12 (Appendix D)). Whereas SyncTweedies and MultiDiffusion primarily demonstrate spatial synchronization, temporal synchronization must additionally reconcile motion across video chunks. We show that stochasticity plays a key role in temporal synchronization, injecting noise early to break ODE trajectory inertia and facilitate cross-chunk mixing. We verify this directly in Appendix D: a deterministic update retains only 4–7% of each fusion correction, whereas a stochastic update retains 26–33%. Requested comparisons. The variants the reviewer asks for are the rows of Table 3, run on Wan 2.1-1.3B with the same prompts: SyncTweedies-style x̂₀ averaging without a stochastic phase is “Full ODE sampling”, Gen-L-Video-style fusion of noisy latents is “x_t matching”, and StochSync-style maximal stochasticity is “Full SDE sampling”.

Table 3Consistency ↑Motion ↑Quality ↑
Full SDE sampling (maximal stochasticity)0.94270.94490.5298
Full ODE sampling (x̂₀ averaging, no stochastic phase)0.96040.96210.6075
x_t matching (noisy-latent fusion)0.95790.96900.5862
FlowLong (Ours)0.96150.96850.6359

Without the stochastic phase, the chunks stay sharp, but their content breaks between chunks (Fig. 12, project page). With stochasticity at every step, the content is consistent but high-frequency detail is lost, which lowers Quality from 0.6359 to 0.5298. The binary schedule of FlowLong keeps both.

(Continued in the next comment.)

Reviewer_a7BsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FlowLong extends pretrained flow-based video generators without additional training. It denoises overlapping short windows, treats neighboring clean predictions as measurements, and blends their shared frames through an update termed Tweedie matching. Early sampling refreshes Gaussian noise, whereas later sampling is deterministic. The implementation aggregates predictions into a global buffer before slicing the next windows. Experiments evaluate 30/60-second videos, audio-video generation, and text-to-3D Gaussian splatting. The main empirical result is stronger dynamic activity with competitive drift scores; supplementary HunyuanVideo-1.5 examples also support portability. The method nevertheless trades some appearance consistency and image quality for motion.

Critical strengths
  1. The inference-only design is simple and practically transferable. Equations (15)–(18) operate through overlap blending and noise management while each backbone processes its native window; the method does not require architecture changes or additional training. Demonstrations with Wan, LTX-2, supplementary HunyuanVideo-1.5 examples, and the 3D pipeline support portability across the tested settings. This reduces the cost of adapting a new short-video generator to longer outputs. The audio-video and 3D results are useful demonstrations of applicability, although they do not independently establish precise synchronization or geometric accuracy.

  2. The reported motion–appearance tradeoff and staged sampling are useful empirical observations. At 60 seconds, Dynamic Degree is 0.8200 versus 0.5253 for Deep-Forcing, with visual stability of 97.2 (Tables 1–2). Table 3 also tests the sampling extremes: the hybrid schedule improves Quality from 0.6075 under full ODE to 0.6359, while Consistency changes from 0.9604 to 0.9615. Figure 5 provides a concrete visual comparison of layout alignment and detail preservation. These results give the practical design substance beyond a new interpretation. The advantage has limits: 60-second subject consistency is 0.8605 versus 0.9279 for LongLive, and Dynamic Degree measures dynamic activity rather than action diversity or freedom from repetition. I regard the reported combination of activity and stability as the main positive result.

Critical weaknesses
  1. The contribution and manifold claim need clearer justification. Section 4.1 motivates clean-space matching through an inverse-problem formulation, but SyncTweedies already synchronizes denoised Tweedie estimates. The paper should identify the algorithmic increment over this framework, including the role of early noise refresh. Treating a neighboring prediction as a measurement is plausible; the issue is what this formulation contributes beyond existing synchronization. Equations (14)–(16) also do not establish that framewise interpolation preserves the clean-video manifold. The relevant assumptions from DDS need to be connected to the frame-masked update, or the geometric claim should be explicitly heuristic. These distinctions determine the assessment of novelty and theoretical contribution.

  2. The visual comparisons and ablations provide limited support for the stronger motion and mechanism claims. Figure 3's caption describes repetitive motion, drift, and exposure bias without identifying which subject, posture, background, or color differences readers should inspect. Its three frames without clear timestamps do not independently establish repetitive motion or attribute differences to exposure bias, particularly across both autoregressive and bidirectional baselines. Figure 1 illustrates FlowLong's capabilities without providing comparative quality evidence; reporting native and generated durations would clarify the extrapolation shown. Figure 5 and the denser sequences in Figures 7–9 offer more concrete guidance, so this criticism concerns the main comparison's evidentiary precision. Dynamic Degree supports activity but does not directly establish non-repetition.

Table 3 supports hybrid sampling, but does not isolate the proposed trajectory-inertia mechanism. Its fusion-space comparison needs consistent state/noise definitions: global re-slicing implies shared overlap states, while the explanation assumes independently noisy states (Appendices A.4 and C). With shared current states, common affine updates, and normalized linear weights, clean-prediction fusion and next-state fusion can be equivalent. The fusion order and noise-sharing rules are therefore essential before attributing the gap to fusion space. Missing sampling configurations and metric definitions further limit interpretation; the conflicting RIFLEx backbone labels also need reconciliation. These gaps leave the scope of the motion advantage and the cause of the ablation gains insufficiently established.

Questions for authors
  1. What is the algorithmic increment over existing clean-estimate synchronization, and which DDS assumptions apply to the frame-masked update?

  2. Please identify the existing video sequences and time intervals that best demonstrate non-repetitive motion and reduced drift in Figure 3, and add specific visual pointers to its caption/discussion.

  3. Please provide parallel single-step pseudocode for the main method and the Table 3 noisy-latent baseline, specifying overlap states, fusion order, initialization, and fresh-noise sampling/sharing. Which factors are controlled?

  4. Please provide the existing threshold t*, sampling schedules, four-step checkpoint, window configurations, and Table 3 metric definitions, and reconcile the RIFLEx backbone/checkpoint labels.

Optional evidence, if available: a matched synchronization baseline or a comparison controlling effective matching strength could help isolate the contributions. These are diagnostic suggestions, not mandatory new rebuttal experiments.

7 comments
AuthorsOpenReview ↗

Response by Authors (1/4)

W1-1 & Q1

W1-1. "SyncTweedies already synchronizes denoised Tweedie estimates. The paper should identify the algorithmic increment over this framework, including the role of early noise refresh."

Q1. "What is the algorithmic increment over existing clean-estimate synchronization …?"

Differences from SyncTweedies and contributions of the inverse-problem formulation: Compared with existing chunk synchronization methods (SyncTweedies, Gen-L-Video, MultiDiffusion, etc.), FlowLong is grounded in an inverse-problem formulation that (1) accommodates diverse real-video measurements, (2) enables flexible optimization designs, and (3) highlights the limitations of deterministic synchronization for long-video generation and the importance of early sampling stochasticity. This formulation also provides an optimization-based interpretation of chunk fusion. We elaborate below.

(1) Diverse downstream tasks with generalizable measurements. FlowLong formulates the guidance loss in Eq. (12) using the inverse-problem template 12∥y−Ax∥2\frac{1}{2}\lVert y - Ax \rVert^2, where selects observed frames, is the current chunk’s Tweedie estimate, and is a measurement. Although our main experiments use neighboring chunks’ denoised estimates as, this choice is not required. We additionally apply FlowLong to autoregressive long-video extension, using a real video prefix as the measurement (Fig. 14, Appendix E), and long-video frame interpolation, using multiple observed frames or clips to guide generation of the missing intervals (Fig. 13, Appendix E), which are not addressed using existing methods such as SyncTweedie. The corresponding videos for both tasks have been added to our revised anonymous project page. These results demonstrate that the same formulation supports both chunk synchronization and conditioning on real observations, extending beyond the generation settings considered in the main SyncTweedies experiments.

(2) Customizable optimization design. Although Eq. (12) uses an L2 loss between denoised estimates, the framework accommodates additional regularizers and alternative losses. For long-video inverse problems such as frame interpolation, a data-consistency term can complement Tweedie matching Eq. (12): the former enforces agreement with actual degraded observations, while the latter promotes smooth transitions between generated chunks. Also, for video extension and interpolation, we find that anchoring generation to the low-frequency components of the observations substantially improves temporal consistency:

$ \ell_k(x;t)=\frac12\left( |y-Ax|^2+ |P_L(y)-P_L(x)|^2 \right). $

This additional regularizer pulls the generated chunk toward the observed low-frequency content (Eq. (26) in Appendix E; results in Figs. 13 and 14). Thus, simple weighted interpolation is one instance of our latent optimization framework, which supports broader optimization and regularization designs for improved performance and task versatility, potentially including reward-based updates.

(3) Importance of Stochastic Early-Phase Sampling. We provide additional video examples demonstrating the importance of early sampling stochasticity (Fig. 12 (Appendix D), project page). Whereas SyncTweedies and MultiDiffusion primarily demonstrate spatial synchronization, temporal synchronization must additionally reconcile motion across video chunks. We revealed that stochasticity plays a key role in such temporal synchronization, where noise is injected during the initial stages to break ODE trajectory inertia and facilitate cross-chunk mixing. We verify this directly in Appendix D: a deterministic update retains only 4–7% of each fusion correction, whereas a stochastic update retains 26–33%.

AuthorsOpenReview ↗

Response by Authors (2/4)

W1-2 & Q1

W1-2. "Equations (14)–(16) also do not establish that framewise interpolation preserves the clean-video manifold."

Q1. "… which DDS assumptions apply to the frame-masked update?"

We thank the reviewer for highlighting this distinction. Our connection to DDS lies in extending its clean-space optimization perspective from image inverse problems to video generation. Specifically, frame-masked matching can be viewed as a video counterpart of image inpainting, where consistency is imposed on selected frames of a complete spatiotemporal video chunk rather than on selected pixels. DDS provides a tangent-space argument under the assumption that the clean-data manifold's tangent space coincides with an appropriate Krylov subspace, within which conjugate-gradient updates are performed. In contrast, FlowLong does not explicitly impose such a Krylov-subspace restriction. Instead, its frame-masked correction takes the closed form of a weighted interpolation between neighboring Tweedie estimates. Importantly, this interpolation can equivalently be interpreted as the solution of a regularized quadratic objective that balances neighboring-chunk consistency with proximity to the current Tweedie estimate. This implicit proximal anchoring discourages large deviations from the pretrained model's clean-video prediction, while subsequent denoising continues to leverage the learned video prior. Our experiments demonstrate that the proposed update remains effective even under substantial temporal extrapolation beyond the backbone's native window. We agree that neither proximal anchoring nor empirical extrapolation performance formally establishes manifold preservation. We will therefore clarify that the manifold interpretation provides a geometric motivation for our approach rather than a theoretical preservation guarantee.

AuthorsOpenReview ↗

Response by Authors (3/4)

W2 & Q2

W2. "The visual comparisons and ablations provide limited support for the stronger motion and mechanism claims. …"

Q2. "Please identify the existing video sequences and time intervals … in Figure 3 …"

(1) Figure 3. We replaced Figure 3 with a denser comparison: seven frames per method, sampled every 5 s from 0 to 30 s, with the timestamps shown above each column. The main text (Sec. 5.1) now names what to inspect. The bidirectional baselines stop moving within the first seconds and saturate in color. Self-Forcing and Deep-Forcing simplify the scene and lose the festival crowd over time, and the remaining motion repeats a small set of poses. FlowLong keeps the subject and the scene through 30 s while the pose and camera framing keep changing. More prompts are shown in the appendix, and the full videos are in the supplementary material.

(2) Figure 1. We added the native and generated lengths to the caption of Figure 1: the videos are extended from the 5 s native window to up to 30 s, and the text-to-3DGS scenes contain 1.64× more Gaussians than the native-window baseline. Both are now stated explicitly in the caption.

(3) Dynamic Degree and non-repetition. Dynamic Degree measures the amount of motion, and we complement it with the human study in Appendix B.2. Raters compared the videos directly and preferred FlowLong for motion diversity in 92.4% of comparisons against the five AR baselines, with no baseline above 16.7%. Table R1. Motion repetition on 60s videos (DINOv2).

CausVidSelf-ForcingDeep-Forcing∞-RoPELongLiveFlowLong (Ours)
Change ↑0.1040.1900.1990.1880.2380.377
Non-revisit ↑0.9270.9420.8840.9720.8830.974
Distinct states / min ↑133125243388

We also measure repetition directly. We extract DINOv2 patch features from the frames of each 60s video and compare every frame with its past. Change is the feature distance between a frame and the frame 2 s earlier. Non-revisit is the distance to the closest frame more than 2 s in the past, divided by the distance to the frame 2 s earlier; a value near 1 means the video never returns to an earlier state, and a low value means it repeats. Distinct states counts how many frames per minute are farther than a fixed distance from all states seen so far.

FlowLong’s non-revisit score is on par with the best baseline (∞-RoPE) while changing twice as much, and it visits more than 2.5× as many distinct states per minute as any AR baseline. The larger motion of FlowLong is therefore new content over time, not a repetition of the same motion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by Authors

We sincerely thank the reviewers for their constructive and thorough comments. We are encouraged that the reviewers recognized that our method is simple, training-free, and backbone-agnostic (a7Bs, fdG7, Rjm3), that its memory does not grow with the video length (fdG7, Rjm3), and that it sustains motion over long horizons where autoregressive baselines lose it (a7Bs, fdG7, Rjm3).

In response to the comments, we have made the following major revisions, including additional experiments, highlighted in blue in the revised paper:

  • Applications of the inverse-problem formulation to real measurements: long video interpolation and video extension (Appendix E)
  • Validation of trajectory inertia and comparison of full ODE, full SDE, and our sampler (Appendix D)
  • Content consistency evaluation with Grounding DINO (Appendix B.3)
  • Comparison with FIFO-Diffusion (Table 1) and discussion of Gen-L-Video, SyncTweedies, SynCoS, FreeNoise, and StochSync (Section 2)
  • Sampling configuration, Table 3 metric definitions, and pseudocode of the fusion-space comparison (Appendices A.1 and G)
  • Clarifications on Figure 1, Figure 3, and the comparison with baselines (Abstract and Section 5.1)

The corresponding videos are available on the revised anonymous project page. Below, we address the comments of each reviewer in detail.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.