acceptodds
Under review as a conference paper at ICLR 2027

Step Forcing: Training-free Step-Matched KV Caching for Long Video Generation

Abstract

Autoregressive video diffusion models adapt bidirectional video diffusion models into causal architectures, enabling long video generation through autoregressive rollout. However, these models often suffer from model drift over time, leading to progressive degradation in visual quality. Existing studies generally attribute such model drift to error accumulation during autoregressive rollout. Although existing methods alleviate error accumulation, they overlook a fundamental mismatch in the chunk-wise diffusion process, where the noisy state of the current chunk attends to the same KV caches extracted from previously generated clean chunks. To address this mismatch, we present Step Forcing, a training-free approach designed to mitigate model drift during autoregressive long video generation. We find that reusing fixed clean KV caches throughout the denoising trajectory introduces a step-wise representational mismatch and that noisy and clean KV caches serve complementary roles in preserving visual quality and temporal consistency. Building on these insights, Step Forcing reuses noisy KV caches from corresponding denoising steps while retaining a subset of clean anchor tokens. Our proposed Step Forcing effectively mitigates model drift by using step-matched KV tokens, enabling baseline models to generate long videos with high visual quality and temporal consistency. Results on benchmark quantitatively demonstrate that Step Forcing improves the performance of baseline models.

Primary area:
generative models
Keywords:
Long video generation, Training free

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_DZ1rOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Step Forcing, a training-free method for mitigating model drift in autoregressive video diffusion. The method stores KV caches at different denoising steps, reuses step-matched noisy KV caches from previous chunks, and retains a subset of clean KV tokens as anchors for temporal consistency. Experiments on several autoregressive video generation baselines show noticeable improvements on CausVid and Causal Forcing, although the effectiveness is less clear on stronger training-free baselines.

Critical strengths

The paper identifies a potentially important mismatch between the noisy state of the current chunk and the clean historical KV cache reused throughout the denoising trajectory. Based on this observation, it proposes to store KV caches at different denoising steps, reuse step-matched noisy KV caches, and retain clean anchor tokens to preserve temporal consistency.

Critical weaknesses
  1. The comparison between Step Forcing and other training-free methods is not sufficiently fair or clearly defined. The official Deep Forcing implementation uses the Self-Forcing DMD checkpoint and functions as a training-free inference and KV-cache policy. Therefore, the appropriate comparison should apply Deep Forcing and Step Forcing separately to the same frozen Self-Forcing checkpoint, using the same prompts, seeds, sampling schedule, generation horizon, and memory budget. However, the paper presents Deep Forcing as a base framework and reports “Deep Forcing + Step Forcing” without clearly explaining this relationship or isolating the standalone effect of the two inference policies. The paper should also compare Step Forcing with closely related training-free methods such as TokenTrim or Relax Forcing. The small additional gain on Deep Forcing, only 0.09 in the VBench-Long Total score, further makes the claimed general effectiveness unclear.

  2. The effectiveness of the clean-anchor mixing strategy is still insufficiently established. Table 3 includes the r_max = r_min = 0 setting, which appears to remove clean-anchor replacement and therefore provides a noisy-only ablation. However, the paper does not provide a matched comparison among full clean KV, noisy-only Step KV, and the proposed noisy-clean mixed KV under the same configuration. It also does not validate the proposed clean-token selection rule, which selects tokens using the norm of clean keys and applies the same positions to both keys and values. The compatibility of clean and noisy representations, as well as the effect of possible anchor-selection errors, therefore remains unclear.

Questions for authors
  1. The official Deep Forcing implementation appears to use the Self-Forcing DMD checkpoint as its frozen base model. Could the authors clarify whether the Deep Forcing and “Deep Forcing + Ours” results in Table 2 are obtained on the same Self-Forcing checkpoint? For a fair comparison, could the authors compare Deep Forcing applied to the Self-Forcing checkpoint against Step Forcing applied to the same checkpoint, under identical prompts, seeds, sampling schedules, generation horizons, and memory budgets? If applicable, please also report the result of applying both policies together.

  2. Could the authors compare Step Forcing with other closely related training-free inference methods, such as TokenTrim or Relax Forcing, using the same frozen base model and evaluation protocol?

  3. Table 3 includes the r_max = r_min = 0 setting. Could authors provide a matched comparison among the original full-clean KV cache, noisy-only Step KV, and the proposed noisy-clean mixed KV strategy?

  4. Could the authors clarify why clean-token importance is estimated using the norm of the clean keys, and why the same selected positions are used to replace both keys and values? How sensitive is the method to random anchor selection or other importance criteria, such as attention-based selection?

  5. Could the authors report the peak GPU memory and inference latency of Step Forcing compared with Deep Forcing and other training-free methods under the same model, generation horizon?

3 comments
AuthorsOpenReview ↗

Response to Reviewer DZ1r (1/2)

We appreciate your constructive comments and suggestions. We respond to your concerns below to further improve the quality of our paper. In addition, we have revised the manuscript and highlighted the changes in blue for ease of reference.

W1.1: The comparison between Step Forcing and other training-free methods.

All four 30-second configurations in Table 2 use the same frozen Self Forcing DMD checkpoint. Deep Forcing and Step Forcing are both applied at inference. Specifically, “Self Forcing + Ours” applies Step Forcing alone, “Deep Forcing” applies Deep Forcing alone, and “Deep Forcing + Ours” applies both inference policies to the same checkpoint.

Inference policyTotal ↑Quality ↑Semantic ↑
Original Self Forcing80.8481.9476.42
Step Forcing alone81.9483.0477.51
Deep Forcing alone83.0384.3177.90
Deep Forcing + Step Forcing83.1284.4078.03

Applied individually, Step Forcing and Deep Forcing increase the Total score over the original Self Forcing baseline by 1.10 and 2.19 points, respectively. Applying both policies yields a Total score of 83.12.

W1.2&Q2: Explanation of performance gains and comparison with Relax Forcing.

Deep Forcing already supports long-horizon video generation. At 30 seconds, integrating Step Forcing modestly increases the VBench-Long Quality, Semantic, and Total scores by 0.09, 0.13, and 0.09 points, respectively.

To address this concern, beyond the qualitative 120-second results, we further evaluated Deep Forcing with and without Step Forcing on 120-second video generation. As shown in the table below, integrating Step Forcing improves all three aggregate scores: Quality, Semantic, and Total. Together, these results show improved aggregate performance at long-horizen generation, such as 120 seconds.

ModelTotalQuality ScoreSemantic Score
Relax Forcing77.4677.4277.63
Deep Forcing82.5683.5378.68
Step Forcing82.8984.1078.03

TokenTrim’s implementation is currently unavailable, so we include its published 30-second VBench-Long results in the table below. Both TokenTrim and Step Forcing achieve higher Total scores than Self Forcing. Step Forcing also achieves higher Semantic and Total scores than those reported for TokenTrim.

ModelTotalQuality ScoreSemantic Score
Step Forcing80.8481.9476.42
+Ours82.8984.1078.03
+TokenTrim81.8489.7973.89

W2.1&Q3: Explanation of the effects of clean and noisy KV representations and anchor token selection.

To understand the effects of clean and noisy KV representations and anchor token selection, we provide a comparison among Full Clean KV, Noisy-only Step KV, and the proposed Noisy-Clean Mixed KV under identical experimental settings (rows 1–3 of the table below). As shown, Noisy-only Step KV outperforms Full Clean KV in Quality, Semantic, and Total scores. Although the proposed Noisy-Clean Mixed KV achieves a slightly lower Semantic score than Noisy-only Step KV, it delivers higher Quality and Total scores, demonstrating a better trade-off between Quality and Semantic performance.

Furthermore, to better understand the effect of anchor selection and justify our design choice, we conduct additional ablations using (rmax⁡,rmin⁡)=(0.8,0.4)(r_{\max}, r_{\min}) = (0.8, 0.4) from Table 3 as the reference configuration, corresponding to clean-anchor ratios of ([0.8, 0.5597, 0.4]) from the first to the last denoising step. Reversing the schedule to (0.4, 0.5597, 0.8) improves the Semantic score but substantially reduces the Quality score, resulting in a lower Total score. Similarly, using a fixed ratio of (0.5865667), which matches the average ratio of the original schedule, yields lower Quality, Semantic, and Total scores. These results indicate that allocating more clean anchors to earlier denoising steps achieves a better balance between Quality and Semantic performance.

MethodTotal↑Quality↑Semantic↑
Full clean KV79.0580.3473.88
Noisy-only Step KV82.2583.3277.97
The proposed noisy-clean mixed KV82.9384.4376.96
r=(0.8, 0.5597, 0.4)82.9384.4376.69
r=(0.4, 0.597, 0.8)82.5883.7178.05
r=(0.5865667,0.5865667,0.5865667)82.4083.7876.88
AuthorsOpenReview ↗

Response to Reviewer DZ1r (2/2)

W2.2&Q4: Explanation of the Clean-Token Selection

Video tokens often share visual content across space and time, which can lead to correlations among their key representations. Inspired by prior work on low-rank approximations of attention representations (Saxena et al., 2024), we assume that the clean-key matrix is approximately low-rank. We adopt a row-sampling strategy inspired by (Frieze et al., 2004), using the squared ℓ2\ell_2 norm of each clean key as a query-independent proxy for token importance. The ratio ∣ki∣22/∣Kclean∣F2|k_i|2^2/|K{\mathrm{clean}}|_F^2 measures the fraction of the clean-key matrix’s energy contributed by token (i). Eq.(5) uses this ratio to give higher-energy clean keys a greater probability of being selected when the anchor budget is limited.

We select anchor positions based on the clean keys based on the Eq.(5) and replace both the key and the value at each selected position with their clean counterparts. This preserves positional embedding alignment between keys and values, ensuring that the attention weight computed from a clean key is applied to the clean value from the same token position.

Furthermore, we use clean-key norms because they allow us to select anchors as soon as a chunk finishes denoising. An attention-based score depends on queries from the next chunk, which become available only during its denoising. Norm-based selection therefore lets us prepare the mixed KV cache in advance and release the temporary clean and noisy KV buffers before generating the next chunk.

Q1: Experimental Setup for the Comparison.

All four configurations in Table 2 use the same frozen Self Forcing DMD checkpoint. For the 30-second comparison, we use the GPT-4o-extended prompts released by Self Forcing (https://github.com/guandeh17/Self-Forcing/blob/main/prompts/vbench/all_dimension_extended.txt) and follow the VBench-Long evaluation protocol (https://github.com/Vchitect/VBench/blob/master/prompts/README.md).

For each configuration, we generate five videos per prompt for Appearance Style (90 prompts), Color (85), Human Action (100), Multiple Objects (82), Object Class (79), Overall Consistency (93), Scene (86), Spatial Relationship (84), Subject Consistency (72), and Temporal Style (100). For Temporal Flickering, we generate 25 videos per prompt for 75 prompts. This yields 6,230 videos across 946 prompts per configuration.

We derive each seed deterministically using the following function. Each baseline and its Step Forcing counterpart use identical values for all four inputs, yielding the same prompt–seed pairs.

def stable_seed(base_seed, folder, prompt, sample_index):
    payload = f"{base_seed}\0{folder}\0{prompt}\0{sample_index}".encode("utf-8")
    return int.from_bytes(hashlib.sha256(payload).digest()[:4], "little") & 0x7FFFFFFF

The original Self Forcing configuration achieves a Total score of 80.84. Applying Step Forcing alone yields 81.94, and applying Deep Forcing alone yields 83.03. Applying both policies together yields 83.12, an observed gain of 0.09 points over Deep Forcing alone.

Q4.2: How sensitive is the method to random anchor selection?

We compare our anchor token selection with random selection. Replacing our anchor token selection with random selection reduces the Total score by 0.29 points. The drop is larger in Quality than in Semantic, indicating modest sensitivity to the selection rule in this setting.

MethodTotal ↑Quality ↑Semantic ↑
Anchor Token Selection (Ours)82.7984.3376.62
Random Anchor Selection82.5083.9976.53

Q5: Peak GPU memory usage and inference latency.

We report peak GPU memory and inference latency for Self Forcing, Relax Forcing, Deep Forcing, and Step Forcing under the same model and generation horizon.

MethodPeak GPU memory usage (GB)Inference latency (s)
Self Forcing32.94756.347
Relax Forcing61.99844.710
Deep Forcing36.49452.060
Step Forcing42.93652.543

Step Forcing has comparable latency to Deep Forcing (52.543 vs. 52.060 s) and uses 6.442 GB more peak GPU memory.

Dear Reviewer DZ1r,

Thank you so much for your valuable feedback and the insightful discussions. They have truly helped improve our work!

In our responses, we compare Step Forcing with other training-free methods, explain the effects of clean and noisy KV representations and the role of Anchor Token Selection, provide details on the experimental setup, evaluate the method’s sensitivity to random anchor selection, and report peak GPU memory usage and inference latency.

As the discussion phase comes to a close on November 18, we would like to check in and see if we’ve addressed all of your questions. If you're satisfied with our responses, we would be grateful if you’d consider updating your score.

Thank you again for all your time and support!

Best regards, Authors of submission 2775

Reviewer_HX16OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Step Forcing, a training-free, plug-and-play method to mitigate drift in chunk-wise autoregressive video diffusion models. In standard models, every denoising step of the current chunk attends to the same clean KV cache of previous chunks. The authors argue that this mismatches the step-specific representations and contributes to drift. Instead, Step Forcing reuses the noisy KV caches of previous chunks from the corresponding denoising step (Step KV). It also replaces a subset of these tokens with clean anchor tokens to provide the model with reliable content information. The anchors are sampled by key norm, and their number decreases with SNR along the denoising trajectory. To offset the memory cost of storing per-step caches, the noisy caches are compressed.

Critical strengths

S1. The paper is easy to follow, and the motivation is clear.

S2. The method is training-free and plug-and-play. It is applied to four different baselines, both training-based and training-free.

Critical weaknesses

W1. Incomplete component-wise validation. Further ablations are needed to better validate the contributions of individual components. While the paper studies different clean-anchor ratios, comparisons with alternative anchor selection strategies, such as random versus SNR-dependent selection, would help clarify the design choices. In addition, the semantic score appears to be higher when no anchors are used, which warrants further investigation and explanation. Overall, the reported improvements seem to be driven largely by step KV, while the contribution of clean-anchor mixing remains unclear.

W2. Limited novelty over HiAR. Although HiAR already explores conditioning on context at corresponding noise levels, the extent to which Step Forcing advances beyond this idea remains unclear. A clearer comparison with HiAR is needed to establish the novelty of the proposed method, particularly regarding the role of clean-anchor mixing and its contribution beyond step-matched KV caching. In particular, HiAR explores different timestep configurations, and the rationale for the timestep choices adopted in this paper should be explained. A direct comparison with HiAR under comparable settings would also help clarify the advantages of the proposed approach and substantiate its claimed contributions.

Questions for authors

Q1. In Fig. 2, could the authors evaluate the second chunk using ground-truth clean historical context in a controlled or teacher-forced setting? Disabling historical KV reuse removes both the potentially mismatched representations and useful content information. Using accurate historical context would help distinguish the proposed step-wise mismatch hypothesis from error propagation due to imperfect historical predictions.

Q2. Could the authors provide more targeted component ablations, including norm-based versus random anchor selection, an SNR-dependent versus fixed anchor schedule under a comparable token budget, and the effect of clean-anchor mixing across different baselines? In particular, how much of the improvement over the original models comes from Step KV alone?

Q3. Why does Step Forcing use input-level KV representations at tjt_j, rather than output-level representations at tj+1t_{j+1}? HiAR explicitly compares these choices and reports the results. Could the authors evaluate both settings under the Step Forcing framework?

Q4. How does Step Forcing compare with HiAR's matched-noise conditioning, including its input-level variant and training-free inference setting, under comparable models, prompts, and evaluation metrics? Such a comparison would clarify whether the gains arise from step-matched conditioning itself or from the proposed clean-anchor mechanism.

3 comments
AuthorsOpenReview ↗

Response to Reviewer HX16 (1/2)

We appreciate your constructive comments and suggestions. We respond to your concerns below to further improve the quality of our paper. In addition, we have revised the manuscript and highlighted the changes in blue for ease of reference.

W1&Q2.1: Explanation of Step KV and Anchor Token Selection.

Table 3 evaluates different clean-anchor ratios. To further clarify our design choice, we conduct additional ablations using (rmax⁡,rmin⁡)=(0.8,0.4)(r_{\max}, r_{\min}) = (0.8, 0.4) reported in Table 3 as the reference configuration, corresponding to clean-anchor ratios of (0.8, 0.5597, 0.4) from the first to the last denoising step. Reversing this schedule to (0.4, 0.5597, 0.8) improves the Semantic score but substantially reduces the Quality score, resulting in a lower Total score. Similarly, using a fixed ratio of (0.5865667), which matches the average ratio of the original schedule, yields lower Quality, Semantic, and Total scores. These results suggest that allocating more clean anchors to earlier denoising steps achieves a better balance between Quality and Semantic performance.

rTotal↑Quality↑Semantic↑
0.8, 0.5597, 0.4 (rmax=0.8, rmin=0.4)82.9384.4376.96
0.4, 0.597, 0.882.5883.7178.05
0.586566782.4083.7876.88

Furthermore, the higher Semantic score observed in the last row of Table 3, where no clean anchors are used, is driven by improvements in individual semantic metrics, such as Color and Object Class, at the expense of the Quality score. A similar trend is observed in Table 1.

Importantly, the improvements are not primarily driven by Step KV alone, but by the complementary contributions of Step KV and Anchor Token Selection. Although Step KV alone achieves a relatively high Semantic score, it degrades Quality metrics, including Subject Consistency and Motion Smoothness, as also illustrated in Figure 3 and Table 1. Combining Step KV with Anchor Token Selection provides a better trade-off between Quality and Semantic performance, ultimately achieving a higher Total score.

W2.1: Novelty and Differences from HiAR

We agree that HiAR already reports the drift–continuity trade-off when hierarchical denoising is applied only at inference. However, HiAR draws its motivation from the original bidirectional Wan2.1 model, which jointly denoises all frames at a shared noise level. In Wan2.1, each denoising step uses only the keys and values computed at that step. How reusing the same clean historical KV cache at every denoising step affects visual quality and temporal continuity in chunk-wise autoregressive video models remains an empirical question. To examine this directly, we selectively disable clean KV reuse at individual denoising steps in a pretrained autoregressive video model (Figure 2). Disabling reuse at selected steps reduces visual drift, indicating that drift is sensitive to when clean historical KV is reused.

Furthermore, HiAR addresses the drift–continuity trade-off by retraining under its hierarchical denoising pipeline. Motivated by our observations, Step Forcing modifies how historical KV is reused at inference within the existing chunk-wise generation pipeline. Specifically, it retains selected clean KV tokens as anchors for stable historical context and uses step-matched noisy KV for the remaining tokens.

W2.2&Q4: A direct comparison with HiAR.

As suggested, we compare Step Forcing with HiAR at the same 20-second generation horizon. HiAR uses Wan2.1-1.3B as its backbone and follows the Self Forcing training setup. Therefore, we compare HiAR’s published 20-second results with our evaluation of Step Forcing applied to the pretrained Self Forcing checkpoint.

MethodTraining-freeTotal ↑Quality ↑Semantic ↑
HiAR×82.184.672.3
Ours√82.8083.7179.16

The quantitative evaluation shows that our method achieves higher Total and Semantic scores than HiAR, with a comparable Quality score. These results show that our method achieves a better balance between quality and semantic performance, resulting in a higher overall score than HiAR.

AuthorsOpenReview ↗

Response to Reviewer HX16 (2/2)

Q1: In Fig. 2, could the authors evaluate the second chunk using ground-truth clean historical context in a controlled or teacher-forced setting?

To address this question, we evaluate chunk 2 using clean historical KV extracted from a real video matching the prompt “A brown bear walks in a grassy field.” Since the generated first chunk has no corresponding ground-truth frames, we use the real frames as a prompt-matched reference. We construct its clean KV using the same procedure as in vanilla autoregressive generation and substitute it for the generated history’s clean KV at the first denoising step of chunk 2. As shown in Appendix B (Fig. 6d), visual drift persists with real-video-derived historical KV, and its effects remain visible in the final output. Compared with vanilla generation, the real-video condition yields higher FVD at both steps and lower MS-SWD at the final step. These findings further suggest that reusing clean historical KV caches in autoregressive generation contributes to model drift, even when the historical context is derived from real video frames.

First-step FVD ↓First-step MS-SWD ↓Final-step FVD ↓Final-step MS-SWD ↓
Generated clean KV (vanilla, Fig. 2a&6a)2002.964.45620.652.61
Historical KV removed (Fig. 2a&6b)409.232.58583.932.06
Real-video-derived clean KV (Fig. 6d)2207.495.031011.912.09

Q2.2: The original model uses only Step KV.

The original CausVid model achieves a Total score of 79.05 (Table 2). Using step-matched KV without clean anchors (rmax⁡=rmin⁡=0)(r_{\max}=r_{\min}=0) increases the Total score to 82.25 (last row of Table 3).

Q3: Why Step Forcing uses input-level rather than output-level KV representations.

We analyze the standard denoising process of a pretrained autoregressive video model. At each denoising step tjt_j, the model computes the current chunk’s KV from its input latent at tjt_j and reads historical KV cached from the final clean latents of previously generated chunks. To examine how this historical conditioning affects drift, we selectively disable clean KV reuse at individual denoising steps Figure 2. Disabling reuse at selected steps reduces visual drift, indicating that drift is sensitive to how clean historical KV is used during denoising. This observation motivates Step Forcing, our training-free method for long-video generation. At each step tjt_j, we use historical KV computed at the same input noise level as the current chunk and retain selected clean KV tokens as anchors.

Dear Reviewer HX16,

Thank you so much for your valuable feedback and the insightful discussions. They have truly helped improve our work!

In our responses, we elaborate on Step KV and Anchor Token Selection, clarify our method’s novelty and how it differs from HiAR, report an evaluation of the second chunk conditioned on clean historical KV extracted from real video frames, and explain why Step Forcing operates at the input level.

As the discussion phase comes to a close on November 18, we would like to check in and see if we’ve addressed all of your questions. If you're satisfied with our responses, we would be grateful if you’d consider updating your score.

Thank you again for all your time and support!

Best regards, Authors of submission 2775

Reviewer_D2TiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper modifies historical KV representations in autoregressive video diffusion without retraining. Each denoising step reads corresponding historical step caches with selected clean anchors; a noise-dependent schedule controls anchor counts, and token selection reduces storage. This changes the effective denoiser while preserving chunk-wise generation. Evaluation covers four pipelines, with quantitative results for 5-second and 30-second generation and qualitative examples up to 120 seconds.

Critical strengths
  1. The approach preserves chunk-wise generation and reuses representations produced during denoising. Clean-anchor mixing offers a practical intervention without retraining, with a reported storage/quality trade-off (Sections 3.3–3.4; Table 5).

  2. CausVid's total scores improve from 78.43 to 83.63 for 5-second generation and from 79.05 to 82.79 for 30-second generation. Integration into four pipelines supports applicability beyond one model (Table 2).

Critical weaknesses
  1. The causal diagnosis and distinctive consistency benefit are insufficiently isolated. With fixed history and identical causal conditioning, exact KV reuse is equivalent to recomputation. Figure 2 instead removes historical KV, changing conditioning and attention normalization; it does not isolate step mismatch from reduced history dependence or error propagation. Equations (6)–(7) modify the effective denoiser, as the Introduction acknowledges. This is a legitimate empirical intervention, but its gains do not establish the claimed mismatch mechanism. The causal language should be qualified unless supported by targeted controls, such as matched-budget aligned versus misaligned step caches with fixed history. The cited HiAR already reports the inference-only drift/continuity trade-off. Table 3 lacks direct consistency measures; Table 2 lowers Self Forcing's subject/background consistency from 98.58/97.06 to 97.42/96.29. Deep Forcing gains only 0.09 overall without uncertainty, and 120-second evidence remains qualitative. Practical gains do not establish robust long-horizon consistency gains.

  2. The compressed method used in the main experiments is underspecified. Section 3.4 selects noisy-key tokens using Equation (5)'s strategy, while clean anchors are also sampled. It reuses s for the post-compression length, but does not connect that convention to the anchor budget in Equation (8), or specify the joint candidate domain, order, and how selected anchors map onto retained positions. Algorithm 1 omits beta and uses the current chunk's clean KV before completion. Main settings use beta=0.5/0.6/0.7, so this affects the reported method, not an optional optimization. Table 5 defines ΔMem only as a memory increase (%); that caption does not state whether the quantity is peak GPU memory during full generation. I do not infer that its values are false.

Questions for authors
  1. Implementation clarification: Please provide pseudocode matching the main experiments: when are step caches selected, clean KV computed, anchors sampled/mixed, and buffers released? Please confirm whether Equation (8) uses the post-compression length, and specify the candidate domain, position mapping, clean-refresh context, and a correction of Algorithm 1.

  2. Memory clarification: Table 5 reports ΔMem as a memory increase (%) without defining its scope. Please specify what is measured, including whether it is peak allocated GPU memory during full generation or another quantity, and state cache lifetimes, additional forwards, and whether each host's original window/eviction policy is retained.

  3. Mechanism clarification: How does Figure 2 distinguish step mismatch from weakened historical conditioning or suppressed error propagation? Please identify supporting evidence or qualify the causal claim. Matched-budget step-alignment and compression-only controls are suggestions for additional evidence, not clarification-only requests.

  4. Evaluation clarification: What prompt counts, seeds, pairing, and aggregation produced Table 2 and Figure 2? Can existing records explain the five repeated CausVid component scores across durations and indicate uncertainty for Deep Forcing's gain?

4 comments
AuthorsOpenReview ↗

Response to Reviewer D2Ti (1/3)

We appreciate your constructive comments and suggestions. We respond to your concerns below to further improve the quality of our paper. In addition, we have revised the manuscript and highlighted the changes in blue for ease of reference.

W1.1&Q3: Clarification of Diagnosing Drift in Section 3.2 and Figure 2.

Section 3.2 and Figure 2 are intended to demonstrate that KV cache reuse in autoregressive generation contributes to model drift. Specifically, Figure 2 shows that removing historical clean KV caches at certain denoising steps alleviates drift, providing empirical evidence that clean KV cache reuse contributes to this issue. To clarify this point and avoid overstating the underlying mechanism, we have revised Section 3.2 and the conclusion of Figure 2 as follows:

In conclusion, reusing clean KV caches in autoregressive generation contributes to model drift, highlighting the need to provide KV representations that better match the requirements of each denoising step in the current chunk.

In the subsequent sections, we further explore KV representations that better align with the requirements of each denoising step.

W1.2: The cited HiAR already reports the inference-only drift/continuity trade-off.

We agree that HiAR already reports the drift–continuity trade-off when hierarchical denoising is applied only at inference. However, HiAR draws its motivation from the original bidirectional Wan2.1 model, which jointly denoises all frames at a shared noise level. In Wan2.1, each denoising step uses only the keys and values computed at that step. How reusing the same clean historical KV cache at every denoising step affects visual quality and temporal continuity in chunk-wise autoregressive video models remains an empirical question. To examine this directly, we selectively disable clean KV reuse at individual denoising steps in a pretrained autoregressive video model (Figure 2). Disabling reuse at selected steps reduces visual drift, indicating that drift is sensitive to when clean historical KV is reused.

Furthermore, HiAR addresses the drift–continuity trade-off by retraining under its hierarchical denoising pipeline. Motivated by our observations, Step Forcing modifies how historical KV is reused at inference within the existing chunk-wise generation pipeline. Specifically, it retains selected clean KV tokens as anchors for stable historical context and uses step-matched noisy KV for the remaining tokens.

W1.3: Table 3 lacks direct consistency measures.

To provide direct measures of consistency, we report the per-dimension VBench-Long results for all configurations in Table 3 in the tables below.

Quality Evaluation on VBench-Long

(rmax⁡r_{\max}, rmin⁡r_{\min})Subject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessAesthetic QualityImaging QualityDynamic DegreeQuality Score
(0.9, 0.3)98.0897.0099.2697.8464.5967.2761.2684.26
(0.8, 0.4)98.1197.0499.2897.8464.7267.3562.7584.43
(0.8, 0.2)98.0997.0499.3397.8364.5967.2862.8584.41
(0.6, 0.2)98.2197.1799.2297.7965.1067.1455.7683.94
(0.0, 0.0)98.1096.3899.0497.8263.0067.0555.4583.32

Semantic Evaluation on VBench-Long

(rmax⁡r_{\max}, rmin⁡r_{\min})Object ClassMultiple ObjectsHuman ActionColorSpatial RelationshipSceneAppearance StyleTemporal StyleOverall ConsistencySemantic Score
(0.9, 0.3)90.4571.8993.6082.2171.8051.8622.0424.2525.2276.23
(0.8, 0.4)92.3173.5494.4683.4372.1051.7421.9724.4625.4176.96
(0.8, 0.2)91.5372.6694.8683.7072.3951.7922.0224.4625.3676.90
(0.6, 0.2)92.7277.1492.6780.6973.3452.0622.0524.0925.1076.91
(0.0, 0.0)91.5278.5993.6085.0673.5851.8121.6224.8226.3877.97

Total Evaluation on VBench-Long

(rmax⁡r_{\max}, rmin⁡r_{\min})Total ScoreQuality ScoreSemantic Score
(0.9, 0.3)82.6584.2676.23
(0.8, 0.4)82.9384.4376.96
(0.8, 0.2)82.9184.4176.90
(0.6, 0.2)82.5483.9476.91
(0.0, 0.0)82.2583.3277.97

Compared with the setting (rmax⁡,rmin⁡)=(0,0)(r_{\max}, r_{\min})=(0,0), the (0.8,0.4)(0.8,0.4) setting improves Background Consistency from 96.38 to 97.04 and the Temporal Flickering score from 99.04 to 99.28. Subject Consistency and Motion Smoothness increased marginally, from 98.10 to 98.11 and from 97.82 to 97.84, respectively. Overall Consistency decreases from 26.38 to 25.41. Across the evaluated settings, (0.8,0.4)(0.8,0.4) achieves a better balance between quality and semantic.

AuthorsOpenReview ↗

Response to Reviewer D2Ti (2/3)

W1.4&Q4.3: Explanation of Performance Gains for Self Forcing and Deep Forcing in Table 2

We agree that the 30-second Self Forcing results in Table 2 reveal trade-offs across different evaluation metrics. Adding Step Forcing lowers Subject Consistency and Background Consistency, while improving Aesthetic Quality and slightly increasing Motion Smoothness. The aggregate Quality, Semantic, and Total scores increase by 1.10, 1.09, and 1.10 points, respectively. The examples in Figure 6 illustrate improved visual quality and reduced drift in the displayed sequences.

Deep Forcing already supports long-horizon video generation. At 30 seconds, integrating Step Forcing modestly increases the VBench-Long Quality, Semantic, and Total scores by 0.09, 0.13, and 0.09 points, respectively. To address the concern that our 120-second evidence is qualitative, we further evaluated Deep Forcing with and without Step Forcing on 120-second generation. As shown in the table below, integrating Step Forcing improves all three aggregate scores: Quality, Semantic, and Total. Together, these results show improved aggregate performance at long-horizen generation, such as 120 seconds.

ModelTotalQuality ScoreSemantic Score
Deep Forcing82.5683.5378.68
Step Forcing82.9584.1078.34

W2.1&Q1: The compressed Step Forcing variant used in the main experiments is underspecified, particularly how determines the clean-anchor budget, how anchors are selected and mapped to retained KV positions, when clean KV becomes available in Algorithm 1, and how Table 5 measures ΔMem\Delta\mathrm{Mem}.

We apologize for the confusion and clarify below how KV compression and clean anchor selection work together. In the Methods section, we first describe the core algorithm, then explain in Section 3.4 how compression is incorporated into token selection. In the compressed variant used in our main experiments, the clean anchor budget stjs_{t_j} in Eq. (8) is defined relative to the post-compression cache length.

Once a chunk has been fully denoised, we compute its clean KV and construct the compressed caches for subsequent chunks. For each denoising step tjt_j, we first apply Eq. (5) to the noisy keys to select the retained token indices, then gather both noisy K and V at those indices. We then apply Eq. (5) to the clean keys only at the retained indices to select stjs_{t_j} clean anchors. At each selected position, we replace both the noisy key and value with their clean counterparts at the same original token index.

We have added the β\beta-dependent compression steps to Algorithm 1 in the updated pdf.

Q4.1: What prompt counts, seeds, pairing, and aggregation produced Table 2 and Figure 2?

For Table 2, we follow the official VBench and VBench-Long evaluation protocols (https://github.com/Vchitect/VBench/blob/master/prompts/README.md) and use the GPT-4o-extended prompts released by Self Forcing (https://github.com/guandeh17/Self-Forcing/blob/main/prompts/vbench/all_dimension_extended.txt). We generate five videos per prompt for Appearance Style (90 prompts), Color (85), Human Action (100), Multiple Objects (82), Object Class (79), Overall Consistency (93), Scene (86), Spatial Relationship (84), Subject Consistency (72), and Temporal Style (100). For Temporal Flickering, we generate 25 videos for each of 75 prompts. This yields 6,230 videos across 946 prompts for each evaluated configuration. To make sampling reproducible, we join base_seed, folder, prompt, and sample_index with null bytes (\0) and hash the result with SHA-256 as shown below.

def stable_seed(base_seed, folder, prompt, sample_index):
    payload = f"{base_seed}\0{folder}\0{prompt}\0{sample_index}".encode("utf-8")
    return int.from_bytes(hashlib.sha256(payload).digest()[:4], "little") & 0x7FFFFFFF

For each baseline and its Step Forcing counterpart, we use identical values for all four inputs to obtain the same prompt–seed pairs.

We compute the individual dimension scores using the official evaluator, then normalize and aggregate them into Quality and Semantic scores using the benchmark weights. The Total score is (0.8 \times \text{Quality} + 0.2 \times \text{Semantic}).

For Figure 2, we report statistical results based on 100 prompts randomly sampled from MovieGenVideoBench_extended.txt.

AuthorsOpenReview ↗

Response to Reviewer D2Ti (3/3)

Q4.2: Can existing records explain the five repeated CausVid component scores across durations?

Thank you for flagging the repeated values. We checked the original evaluation records and confirmed that the 5-second CausVid baseline row in Table 2 contained incorrect values. The corrected Quality, Semantic, and Total scores are 83.91, 79.30, and 82.99, respectively. With this correction, Step Forcing still improves the 5-second Total score from 82.99 to 83.63, a gain of 0.64 points. We have updated Table 2 and the corresponding discussion in the revised manuscript to reflect these corrections.

Dear Reviewer D2Ti,

Thank you so much for your valuable feedback and the insightful discussions. They have truly helped improve our work!

In our responses, we clarify how we diagnose drift in Section 3.2 and Figure 2, explain our method’s novelty and how it differs from HiAR, report consistency metrics in Table 3, describe the compression steps added to Algorithm 1, and provide additional experimental details.

As the discussion phase comes to a close on November 18, we would like to check in and see if we’ve addressed all of your questions. If you're satisfied with our responses, we would be grateful if you’d consider updating your score.

Thank you again for all your time and support!

Best regards, Authors of submission 2775

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.