acceptodds
Under review as a conference paper at ICLR 2027

Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion

Abstract

Autoregressive (AR) video diffusion models in the Self Forcing family generate a video chunk by chunk, each conditioned on the ones before it, so they can keep rolling out with no preset length. In practice they are trained on short, fixed clips (, 5s). Once a rollout passes this training horizon, the cache fills with self-generated chunks the model never saw in training, and quality degrades: colors oversaturate, identities drift, structures collapse, and motion fades. We frame this failure as a , a domain shift between the within-horizon caches seen in training and those produced by open-ended testing, and address it with , a training-free inference-time cache policy. Within the base model's fixed cache budget, Rolling Sink holds the cache close to its stable within-horizon behavior: it restores low-drift content, sliding temporal indices, and evolving semantics, and keeps a small recent context. Attention sink and temporal re-indexing are existing mechanisms. Our contribution is the mismatch formulation, the joint bounded-cache policy, and the Rolling Semantics schedule. On two 5s-trained bases, Self Forcing and Causal Forcing, Rolling Sink attains the best averaged rank on at both 1- and 5-minute rollouts under the same inference setting, at about 0.56% throughput change, and holds GPU memory constant. It also sustains a 30-minute qualitative rollout, its training horizon. We validate at 1- and 5-minute rollouts and do not claim guaranteed quality at arbitrary duration.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-Ended Autoregressive Video Diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_v5wPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper relies on key observations: Models like Self Forcing are trained only on short clips, so during long rollouts their cache becomes increasingly out-of-distribution and causes color drift, structure collapse, identity changes, and fading motion. Rolling Sink addresses this by managing the cache in three ways: it reuses low-drift chunks from the initial training horizon as sink chunks, re-indexes their temporal positions so they appear recent, and uses a forward-reverse Rolling Semantics schedule so the sink content changes over time instead of remaining static. One recent chunk is always retained to preserve current local motion.

Critical strengths

I liked the honesty in the paper, which clearly separates the main contributions.

Experiments are well designed, followed the general protocol in the long video generation field. Also experiments clearly demonstrate the effect of the proposed method in preserving the details. Competitive results in 1-minute and 5-minute generations.

Critical weaknesses

There is a fundamental ambiguity in the Rolling Semantics, and I kindly ask authors to clarify the situation clearly:

Let's consider the following cache transition.

{x1,x2,x3,x4,x5,xrecentx_1, x_2, x_3, x_4, x_5, x_{\text{recent}}} →\rightarrow {x2,x3,x4,x5,x5ˉ,xrecentx_2, x_3, x_4, x_5, \bar{x_5}, x_{\text{recent}}}

Assume:

x5=x_5 = {z13,z14,z15z_{13}, z_{14}, z_{15}}

x5ˉ=\bar{x_5} = {z15,z14,z13z_{15}, z_{14}, z_{13}}

Then transitioned cache becomes:

{x2,x3,x4,x5,x5ˉ,xrecentx_2, x_3, x_4, x_5, \bar{x_5}, x_{\text{recent}}} = {x2,x3,z10,z11,z12,z13,z14,z15,z15,z14,z13,xrecentx_2, x_3, z_{10}, z_{11}, z_{12}, z_{13}, z_{14}, z_{15}, z_{15}, z_{14}, z_{13}, x_{\text{recent}}}

I can understand the motivation behind this sink schedule. The sink becomes dynamic rather than static + there is no large jump when the traversal reaches the end of the sink bank, since the latent frames are reversed and the trajectory smoothly changes direction.

However, I kindly argue that the proposed reversed sink schedule is not working as intended in the pixel space.

The reason is that each latent frame in the Wan VAE corresponds to a 4 pixel-space frames. Importantly, the pixel frames represented by each individual latent are still ordered forward in time. Therefore, although {z15,z14,z13z_{15}, z_{14}, z_{13}} looks like a temporally reversed sequence with no jump in latent space, it does not necessarily correspond to a truly reversed video in pixel space.

Example:

Suppose {z13,z14,z15z_{13}, z_{14}, z_{15}} latent frames corresponds to following pixel frames.

z13=z_{13} = { p49,p50,p51,p52p_{49}, p_{50}, p_{51}, p_{52}}

z14=z_{14} = { p53,p54,p55,p56p_{53}, p_{54}, p_{55}, p_{56}}

z15=z_{15} = { p57,p58,p59,p60p_{57}, p_{58}, p_{59}, p_{60}}

Then {z15,z14,z13z_{15}, z_{14}, z_{13}} in the cache corresponds to following structure in pixel space

{ p57,p58,p59,p60,p53,p54,p55,p56,p49,p50,p51,p52p_{57}, p_{58}, p_{59}, p_{60}, p_{53}, p_{54}, p_{55}, p_{56}, p_{49}, p_{50}, p_{51}, p_{52}}

which eventually creates temporal jumps entire paper attempted to avoid.

Questions for authors

The most significant issue, which is against of the main claim of the paper, is shared in Critical Weaknesses section. Unfortunately, the paper is very limited regarding the ablations of the Rolling order schedule, so we can't quite see the direct advantage of reversed rolling sink.

Apart from that, can the authors explain what causes the performance differences across different sink orders in the cache? From an information-theoretic perspective, the cache contains the same amount of information, and only the temporal ordering or recency of that information changes.

3 comments
AuthorsOpenReview ↗

Response to Reviewer v5wP

The proposed reversed sink schedule is not working as intended in the pixel space.

  1. Thank you for pointing out the distinction between latent and pixel reversal. This prompted us to propose a broader version of Rolling Sink with three levels: Chunk-Level, Latent-Level (submitted version), and Pixel-Level. Please check App. J. We agree that this does not imply a reversed RGB sequence under temporal VAE compression.

  2. We now distinguish:

  • Chunk-Level Rolling Sink: reverses chunk order only, keeping latent order within each chunk forward.
  • Latent-level Rolling Sink: also reverses latent order within each reversed chunk, keeping the RGB frames of each latent forward.
  • Pixel-level Rolling Sink: reverses the RGB frames and re-encodes the latents for each reversed chunk. The re-encoding is needed only once, adding only 2.48% / 0.50% runtime at 1 / 5 minutes.
  1. As shown in Tabs. 1-2 in our General Response, on Self Forcing, Pixel-level Rolling Semantics improves average rank from 3.38 to 2.44 at 1 minute and 3.63 to 2.50 at 5 minutes. The submitted latent-level schedule also remains ahead of other ablated variants (except pixel-level) in both settings.

What causes the performance differences across different sink orders in the cache?

  1. The key is the train–test gap in cache behavior. During training, cached chunks and the current chunk occupy consecutive temporal positions that advance as a sliding window. Keeping early sink chunks at their original indices breaks this pattern as AR generation proceeds. Temporal re-indexing restores this positional relationship, bringing the test cache closer to its training behavior and improving performance. The Temporal Re-index ablation is in the submission's App. C, Tabs. S2-S3, and Fig. 5. The ablations in Tabs. 1-2 of the General Response keep temporal re-indexing fixed while comparing forward-only, random, fixed/random, and forward–reverse policies.

  2. Changing sink order changes the content–position assignments. RoPE makes attention depend on these relative positions: placing chunk A closer to the current chunk than chunk B gives different conditioning from placing B closer to the current chunk than A. Advancing every index by the same offset preserves relative positions.

AuthorsOpenReview ↗

Authors' Kind Reminder to Reviewer v5wP

Dear Reviewer v5wP,

Thank you for taking the time to review our paper and provide your valuable feedback. We have submitted detailed responses to your comments and suggestions, addressing the key points raised. As the discussion is scheduled to end on 2026-11-18, we would greatly appreciate it if you could review our responses and provide any further clarifications at your earliest convenience. Your feedback is extremely important to us.

Thank you again for your time and effort.

Best,

Authors of Submission 837

AuthorsOpenReview ↗

Authors' Second Kind Reminder to Reviewer v5wP (Only 1 Day Left)

Dear Reviewer v5wP,

We sincerely thank you for your time and effort for reviewing our submission, and writing your responses.

However, as the discussion deadline approaches in less than 1 day, we would greatly appreciate it if you could carefully review our latest responses and provide further clarifications at your earliest convenience. We highly value any opportunities to address any potential remaining issues, which we believe will be helpful for a higher rating. Your timely feedback is extremely important to us!

Thanks so much again for your time and effort, we deeply appearciate it!

Best,

Authors of Submission 837

Reviewer_WfHrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses quality drift in autoregressive video diffusion when generation exceeds the short (5s) training horizon, framing it as a KV-cache distribution shift. It introduces Rolling Sink, a training-free cache policy that preserves within-horizon behavior via attention sinks, RoPE re-indexing, and a forward-reverse “Rolling Semantics” schedule over early clean chunks. On two 5s-trained bases, it attains the best averaged VBench-Long rank at 1- and 5-minute rollouts with negligible overhead (~0.56% throughput, constant memory), outperforming even LongLive trained on 1-minute clips.

Critical strengths
  1. The paper is easy ton read and follow.

  2. Results are impressive with high potential.

  3. The method is training-free, requires no architecture changes, keeps constant GPU memory and ~unchanged throughput.

Critical weaknesses
  1. Narrow algorithmic novelty: Attention sink and RoPE re-indexing are existing mechanisms; the sole new technical element is the forward-reverse scheduling over the chunk bank. The contribution is primarily a framing and combination of known parts. Furtheremore, Ablation shows attention sink provides the largest gain, with Temporal Re-index and Rolling Semantics adding smaller increments. This shows that the novel component acutalyy contributes here the lease.

  2. Missing baselines: Several relevant methods are not compared: FLEX (listed in Table 1 yet absent from experiments), Self-Forcing++ (4+ min videos from a 5s base), and FAST-AR (training-free KV-cache policy via temporal compression).

  3. The model forever re-reads the same K within-horizon chunks, including backward passes (a partial time-reversal). No new long-range content is created; I find this issue as a major concern. The whole point of AR video diffusion models is to be able to generate long and consistent videos. The paper seems to replace one problem (drift) with another (quick forgetting).

  4. Limited evaluation protocol: Only 10 prompts per VBench-Long dimension; Authors should use the same setup of current approaches.

Questions for authors

My major concern is the limited novely, that the authors themself point in the paper. A concrete answer would potentially change my assessment.

6 comments
AuthorsOpenReview ↗

Response to Reviewer WfHr (1/2)

W1

Narrow algorithmic novelty: The sole new technical element is Rolling Semantics. Furtheremore, ablation shows attention sink provides the largest gain, with Temporal Re-index and Rolling Semantics adding smaller increments. The novel component contributes the least.

  1. We explicitly credit prior work for attention sinks[1] and RoPE re-indexing[2]. Our new component is Rolling Semantics: a sliding window over a forward–reverse scheduling of the finite early bank.

  2. The marginal gain is not small at all sink ratios. In Tab. S3 of our App. C and Fig. 5 of Sec. 4, adding Rolling Semantics raises the 5-minute average score from 56.31 to 63.96 at 1/6, 60.17 to 67.61 at 2/6, 64.47 to 68.78 at 3/6, 67.25 to 69.69 at 4/6, and 69.32 to 70.03 at 5/6. It consistently improves the score at all sink ratios, with much larger gains at lower sink ratios.

  3. Additional ablations also demonstrate the value of Rolling Semantics. As shown in Tabs. 1 & 2 in our General Response's Main Concern #1, we added new ablation experiments on: all-random selected sink chunks, 2 fixed + 3 random sink chunks, and 3 fixed + 2 random sink chunks. Forward-reverse based Rolling Semantics retains the best average rank across both 1- and 5-minute AR video generation. These results further demonstrate the effectiveness of forward-reverse scheduling.

W2

Missing baselines: Several relevant methods are not compared: FLEX (listed in submission's Table 1 yet absent from experiments), Self-Forcing++ (4+ min videos from a 5s base), and FAST-AR (training-free KV-cache policy via temporal compression).

  1. We additionally compare Rolling Sink against: Relax Forcing[4], ID-Forcing[5], TetherCache[6], and DensityKV[7]. We respectfully omitted your mentioned FLEX[8], Self-Forcing++[9], and FAST-AR[10] because they're not open-sourced (at least until 2026-11-10).

  2. As shown in Tabs. 3-6 in our General Response's Main Concern #3, Rolling Sink still retains the best averaged rank across 2 base models and 2 durations (1min and 5min).

W3

No new long-range content is created. The paper seems to replace one problem (drift) with another (quick forgetting).

A. Why we only use early bank

  1. With a fixed and short training horizon, the early chunks generated within the training horizon can be responsibly considered as "low-drift". Continually replacing them with later predictions may accumulate drifted content in the AR cache. We preserve the early bank as a stable reference while incorporating new content through the updated recent chunk.

B. Why the full AR cache keeps evolving

  1. The temporal indices keep advancing as a sliding window on an infinite unidirectional axis.

  2. Backward traversal applies only to sink chunks. The recent chunk stays in forward order, and each new output chunk is generated from the current context.

C. Our results show sustained consistency together with continued visual progression in minute-long AR video generation.

  1. Tabs. 3-6 of our General Response demonstrate high consistency with active and smooth motions. Rolling Sink has the highest subject and background consistency (SubCon and BgCon) on both bases at both 1 and 5 minutes. Its scores on HumAct, MoSmo, and DyDg also show that motion remains active and smooth. Please see this anonymous link for 20+ 5-minute examples: https://iclr-27-rebuttal-anonymous.github.io/

  2. Consistency persists throughout the rollout (Fig. 4 in submission). Overall Consistency remains stable across the 5-minute trajectory on both bases, while both base models decline sharply. Imaging Quality and Aesthetic Quality also remain stable.

  3. Content progression is tested directly (App. D-E in submission). Fig. S2 shows sustained distance from the opening frame without the sharp dips (i.e., periodic replay) seen for LongLive[3]; Fig. S3 shows changing poses and viewpoints with the subject and background scene preserved. Across five 5-minute videos, Tab. S4 reports nearly zero distance dips for Rolling Sink under both DINOv2 and CLIP, versus +0.05 / +0.02 for an explicit bank-loop control. These results demonstrate Rolling Sink's visual progression rather than periodic replaying.

AuthorsOpenReview ↗

Response to Reviewer WfHr (2/2)

W4

Limited evaluation protocol: Only 10 prompts per VBench-Long dimension.

  1. All quantitative experiments (both newly added and existing) within each base model and duration share exactly the same evaluation protocol.

  2. We also evaluated Rolling Sink on 10 new, zero-overlapped prompts per dimension, obtaining 69.80 ± 0.05 Avg Score (mean ± standard deviation) on Self Forcing at 1 minute AR video generation.

  3. Please also see General Response's Main Concern #2.

  • [1] Xiao, Guangxuan, et al. "Efficient streaming language models with attention sinks."
  • [2] Yi, Jung, et al. "Deep forcing: Training-free long video generation with deep sink and participative compression."
  • [3] Yang, Shuai, et al. "Longlive: Real-time interactive long video generation."
  • [4] Zhao, Zengqun, et al. "Relax forcing: Relaxed kv-memory for consistent long video generation."
  • [5] Shin, Jeongwoo, et al. "In-Distribution Forcing for Long Video Generation at Test Time."
  • [6] Meng, Yu, et al. "TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted Alignment."
  • [7] Zhao, Wenqu, et al. "DensityKV: Density-guided KV cache compression for long video generation."
  • [8] Li, Jia, et al. "Train short, inference long: Training-free horizon extension for autoregressive video generation."
  • [9] Cui, Jiaxing, et al. "Self-forcing++: Towards minute-scale high-quality video generation."
  • [10] Samuel, Dvir, et al. "Fast autoregressive video diffusion and world models with temporal cache compression and sparse attention."

Thank you for the authors’ detailed response and additional experiments. The added baselines address my concern regarding missing comparisons, and the new ablations provide support for the forward–reverse schedule. However, my concerns about limited novelty and the method’s inability to retain content beyond the early bank and a single recent chunk remain; the lower Dynamic Degree relative to several baselines reinforces the latter. I also share Reviewer v5wP’s concern, which the authors acknowledged, that the submitted latent-level schedule does not reverse content in pixel space. I therefore maintain my score.

Reviewer_6NHQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Rolling Sink, a training-free cache policy for extending autoregressive video diffusion beyond its training horizon. It combines early attention sinks and temporal RoPE re-indexing with Rolling Semantics, which slides a reference window through a fixed bank of early chunks in a forward–reverse schedule while retaining recent context. Experiments on Self Forcing and Causal Forcing report improved long-horizon video quality with bounded memory and negligible inference overhead.

Critical strengths
  1. Simple and efficient design. The paper addresses quality degradation when autoregressive video diffusion models generate videos beyond their short training horizon. It reuses a fixed bank of early, low-drift chunks through a forward–reverse rolling schedule, while retaining recent context to support local continuity. This provides a lightweight, training-free approach to long-horizon generation with bounded cache memory and negligible inference overhead.

  2. Evaluation across multiple settings. Experiments cover two base models and both one- and five-minute rollouts, with comparisons against relevant training-free baselines under matched inference settings. The method achieves the best averaged rank in all four settings, while component and sink-ratio ablations further characterize its effectiveness.

Critical weaknesses
  1. Limited justification of the rolling mechanism. Rolling Semantics is the main new component, as attention sinks and temporal re-indexing build on prior work. However, it is unclear whether its gains depend on the forward–reverse temporal structure or mainly reflect greater diversity in the selected references. The forward-only ablation offers some support, but comparisons with random selection from the same early bank or a mixture of fixed and random sinks would help establish the value of the specific schedule and its claimed role in restoring evolving semantics.

  2. Limited quantitative evaluation. The main comparison tables use only 10 prompts per VBench-Long dimension, providing limited coverage of long-horizon generation. The same subset is also used to select the default sink ratio, without validation on held-out prompts. With no reported sampling variability, it remains unclear whether the average-rank advantages over competing methods would hold on a broader evaluation set.

Questions for authors
  1. It would be helpful to test whether simpler sink selection strategies can achieve similar gains. For example, all five sink chunks could be randomly selected from the same initial bank, or two could remain fixed while the other three are randomly selected. Comparing these alternatives under the same cache budget, recent context, and temporal re-indexing would help clarify the benefit of the proposed rolling schedule.

  2. It would also be helpful to extend the main comparisons to a larger prompt set, providing stronger support for the reported performance advantages.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6NHQ

W1

Limited justification of the rolling mechanism. It's unclear whether its gains depend on the forward–reverse temporal structure or mainly reflect greater diversity in the selected references. Ablations with random selection from the same early bank or a mixture of fixed and random sinks would help.

  1. As shown in Tabs. 1 & 2 in our General Response's Main Concern #1, we added all-random selection, 2 fixed + 3 random sinks, and 3 fixed + 2 random sinks. All use the same 6-chunk early bank (5 sink chunks and 1 recent chunk), temporal re-indexing, and evaluation protocol.

  2. Latent-level Rolling Semantics (our submitted Rolling Sink) achieves average ranks of 3.38 / 3.63 at 1 / 5 minutes, versus 4.19 / 4.19 for random selection, 4.50 / 4.44 for the 2+3 mixture, and 4.44 / 4.19 for the 3+2 mixture.

  3. These matched comparisons demonstrate the forward–reverse schedule brings more performance gains over the ablated alternatives. Simply making the selected chunks more diverse doesn't bring as much gains as the forward–reverse schedule.

W2

Limited quantitative evaluation The same subset is also used to select the default sink ratio, without validation on held-out prompts. With no reported sampling variability, it remains unclear whether the average-rank advantages over competing methods would hold on a broader evaluation set.

  1. Held-out / sampling variability evaluation: We additionally evaluated Rolling Sink 2 times on 10 newly sampled (unused) prompts per dimension each time. On Self Forcing at 1 minute, the average score is 69.80 ± 0.05 (mean ± std). This demonstrates the high robustness of our evaluation protocol.

  2. Please also see General Response's Main Concern #2.

AuthorsOpenReview ↗

Authors' Kind Reminder to Reviewer 6NHQ

Dear Reviewer 6NHQ,

Thank you for taking the time to review our paper and provide your valuable feedback. We have submitted detailed responses to your comments and suggestions, addressing the key points raised. As the discussion is scheduled to end on 2026-11-18, we would greatly appreciate it if you could review our responses and provide any further clarifications at your earliest convenience. Your feedback is extremely important to us.

Thank you again for your time and effort.

Best,

Authors of Submission 837

AuthorsOpenReview ↗

Authors' Second Kind Reminder to Reviewer 6NHQ (Only 1 Day Left)

Dear Reviewer 6NHQ,

We sincerely thank you for your time and effort for reviewing our submission, and writing your responses.

However, as the discussion deadline approaches in less than 1 day, we would greatly appreciate it if you could carefully review our latest responses and provide further clarifications at your earliest convenience. We highly value any opportunities to address any potential remaining issues, which we believe will be helpful for a higher rating. Your timely feedback is extremely important to us!

Thanks so much again for your time and effort, we deeply appearciate it!

Best,

Authors of Submission 837

Comments on the paper

8 comments
AuthorsOpenReview ↗

General Response to All Reviewers and AC (8/8)

Table 6 Baseline comparison: Causal Forcing [5 minutes]

MethodCode ReleaseSubCon ↑BgCon ↑AesQl ↑ImgQl ↑ObjCls ↑MulObj ↑Col ↑SpaRel ↑Sce ↑TmpSty ↑OvrCon ↑HumAct ↑TmpFli ↑MoSmo ↑DyDg ↑AppSty ↑Avg Score ↑Avg Rank ↓
Causal Forcing--92.3594.94‾\underline{94.94}40.3055.1229.118.3375.0521.404.6611.3311.1038.3996.9796.7031.4520.8045.507.56
Deep Forcing2026-0194.7492.8158.3768.1485.5429.8479.1385.9316.1321.8420.3186.1394.9096.9589.5220.69‾\underline{20.69}65.066.44
MemRoPE2026-0394.1692.4959.6168.8591.8338.1071.8888.1515.0525.08‾\underline{25.08}22.2380.0093.5096.7593.9520.4665.766.00
TetherCache2026-0694.0393.5159.3369.1790.3034.3178.7385.4616.5921.3221.3376.7798.5497.74‾\underline{97.74}94.9320.2465.775.88
∞-RoPE2025-1195.1993.3262.8670.1595.0237.3681.5192.0326.1624.9823.21‾\underline{23.21}80.0095.3395.4199.60‾\underline{99.60}20.3968.284.19
DensityKV2026-0895.4393.6862.0270.57‾\underline{70.57}98.5740.1884.2094.4316.9223.5121.9182.5897.42‾\underline{97.42}97.3684.1519.8267.674.06
ID-Forcing2026-1096.02‾\underline{96.02}93.1461.1171.7293.0840.9776.0095.7217.0325.2922.2688.0697.0597.2399.2119.5468.343.75
Relax Forcing2026-0895.9793.3962.40‾\underline{62.40}70.2398.08‾\underline{98.08}43.09‾\underline{43.09}85.81‾\underline{85.81}98.3417.2024.9421.9986.45‾\underline{86.45}95.1595.72100.0019.9269.29‾\underline{69.29}3.63‾\underline{3.63}
Rolling Sink--97.0995.5761.9770.3291.7972.8088.2997.84‾\underline{97.84}17.56‾\underline{17.56}21.5523.2778.3996.2698.1684.6820.1269.733.44
  • [1] Zhao, Zengqun, et al. "Relax forcing: Relaxed kv-memory for consistent long video generation."
  • [2] Shin, Jeongwoo, et al. "In-Distribution Forcing for Long Video Generation at Test Time."
  • [3] Meng, Yu, et al. "TetherCache: Stabilizing Autoregressive Long-Form Video Generation with Gated Recall and Trusted Alignment."
  • [4] Zhao, Wenqu, et al. "DensityKV: Density-guided KV cache compression for long video generation."
  • [5] Li, Jia, et al. "Train short, inference long: Training-free horizon extension for autoregressive video generation."
  • [6] Cui, Jiaxing, et al. "Self-forcing++: Towards minute-scale high-quality video generation."
  • [7] Samuel, Dvir, et al. "Fast autoregressive video diffusion and world models with temporal cache compression and sparse attention."
AuthorsOpenReview ↗

General Response to All Reviewers and AC (7/8)

Table 5 Baseline comparison: Self Forcing [5 minutes]

MethodCode ReleaseSubCon ↑BgCon ↑AesQl ↑ImgQl ↑ObjCls ↑MulObj ↑Col ↑SpaRel ↑Sce ↑TmpSty ↑OvrCon ↑HumAct ↑TmpFli ↑MoSmo ↑DyDg ↑AppSty ↑Avg Score ↑Avg Rank ↓
Self Forcing--94.2496.1042.8957.0133.3914.2761.0535.704.3011.3211.3927.1098.2098.6524.1920.5345.659.06
MemRoPE2026-0396.6994.9457.9471.76‾\underline{71.76}99.6462.6673.6195.4623.6621.3521.8280.6597.7998.4765.3219.4767.587.69
TetherCache2026-0695.4194.8357.6567.8698.1948.9381.9297.9026.7224.5622.8785.1298.8097.1775.5419.6468.327.06
∞-RoPE2025-1197.1494.4359.0071.8289.1549.0982.4195.2127.24‾\underline{27.24}21.9719.9990.32‾\underline{90.32}97.8897.7670.5619.72‾\underline{19.72}67.736.44
Deep Forcing2026-0196.7694.5960.5469.2999.9060.5286.65‾\underline{86.65}98.9733.3323.7222.1683.5597.8897.6778.23‾\underline{78.23}19.5770.21‾\underline{70.21}6.25
Relax Forcing2026-0897.3394.7860.7170.1399.9662.2081.5999.4726.8924.5323.1586.1197.6497.6773.3819.5669.695.63
DensityKV2026-0896.8395.3860.8970.8499.98‾\underline{99.98}62.1287.1698.9226.8324.4723.86‾\underline{23.86}80.6298.6198.2063.6019.1769.225.44
ID-Forcing2026-1096.6195.7559.4270.4799.7657.2878.8899.3526.9723.9722.7590.8498.8198.0378.3119.7169.815.44
Rolling Sink (chunk-level)--98.1697.0662.2271.2999.8073.19‾\underline{73.19}80.6599.9026.0924.4623.7786.6099.19‾\underline{99.19}98.4061.9418.6170.084.81
Rolling Sink--98.04‾\underline{98.04}96.2962.9669.87100.0072.8478.83100.0026.1625.3323.1087.1098.3298.59‾\underline{98.59}64.1118.9170.034.50‾\underline{4.50}
Rolling Sink (pixel-level)--98.0097.01‾\underline{97.01}62.87‾\underline{62.87}71.2299.9073.5980.6899.98‾\underline{99.98}26.6224.68‾\underline{24.68}24.7187.2699.2198.5165.8118.9870.563.50
AuthorsOpenReview ↗

General Response to All Reviewers and AC (6/8)

Table 4 Baseline comparison: Causal Forcing [1 minute]

MethodCode ReleaseSubCon ↑BgCon ↑AesQl ↑ImgQl ↑ObjCls ↑MulObj ↑Col ↑SpaRel ↑Sce ↑TmpSty ↑OvrCon ↑HumAct ↑TmpFli ↑MoSmo ↑DyDg ↑AppSty ↑Avg Score ↑Avg Rank ↓
Causal Forcing--95.6596.11‾\underline{96.11}52.7362.8669.0232.2374.5365.9113.0218.3917.7350.5796.3097.1754.6421.1357.377.69
Deep Forcing2026-0197.49‾\underline{97.49}95.1861.0568.9989.5035.9184.3392.0916.8322.1920.3975.1495.3897.4893.5720.3466.625.94
TetherCache2026-0696.2395.6759.0169.5494.9241.0385.2991.8421.1221.3222.0374.3999.3097.5786.8520.0767.265.63
MemRoPE2026-0397.0094.8060.8371.26‾\underline{71.26}94.8940.9279.1396.98‾\underline{96.98}16.0123.2421.1176.4793.5397.3988.2420.55‾\underline{20.55}67.025.31
ID-Forcing2026-1096.2195.2360.3870.4695.7645.1783.5092.8122.91‾\underline{22.91}23.0023.0772.7799.26‾\underline{99.26}97.58‾\underline{97.58}95.6819.7968.354.69
DensityKV2026-0896.5695.7059.9671.3595.6545.8285.7992.2620.1122.2522.6675.7498.7197.5688.5620.0268.044.50
∞-RoPE2025-1196.9995.4062.6870.1696.12‾\underline{96.12}39.8880.1495.7721.9023.16‾\underline{23.16}21.6182.29‾\underline{82.29}95.4495.4698.9320.0268.504.50
Relax Forcing2026-0896.2494.9860.7470.7696.1145.91‾\underline{45.91}89.0495.9922.9723.0923.19‾\underline{23.19}77.6198.9696.5995.77‾\underline{95.77}19.7469.23‾\underline{69.23}3.88‾\underline{3.88}
Rolling Sink--98.1696.3862.42‾\underline{62.42}70.4598.6674.9689.02‾\underline{89.02}98.5222.8621.8623.3084.2996.2498.0488.5720.0071.482.81

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.