acceptodds
Under review as a conference paper at ICLR 2027

Token Forcing: Geometry-Guided Search and Memory Selection for Autoregressive Video Generation

Abstract

Autoregressive video diffusion models generate long videos block by block, but over minute-long rollouts their scenes drift as the camera moves. Each block is committed once generated, so geometric errors accumulate, and what remains controllable is which content later blocks rely on. Existing approaches use the generator's own signals, require retraining, or score whole candidates with external rewards. Yet a generated block is rarely uniformly right or wrong, so the question is not which candidate to keep, but which parts of a single generation later blocks should build on. We introduce Token Forcing, a training-free method that answers this question token by token. During the early denoising stages of each block, a frozen recurrent 3D foundation model, whose state summarizes the committed history, keeps the tokens it judges consistent with the scene so far. The generator re-encodes these tokens at the matching noise level and injects their key–value features into the next block at the same denoising stage, where they act as soft geometric anchors. Because the 3D model only decides which tokens to use, neither model is trained. To measure drift over long camera trajectories, we introduce SceneLongBench, 100 static indoor and outdoor scenes for 60-second text-to-video and text-and-image-to-video generation. Relative to Causal Forcing, Token Forcing preserves video quality on motion-rich prompts while reducing reprojection error in all four SceneLongBench settings and epipolar error by up to 21.6%.

Primary area:
generative models
Keywords:
Autoregressive video generation, 3D foundation models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_LsnCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Token Forcing is a training free approach to reduce geometric drift in 3D video that is generated in long auto regressive operation. A recurrent 3D model (frozen) scores intermediate outputs by analysign reconstruction confidence / state attention. Some of these regios are re-encoded by the video generation model and injected additional KV context. Overall, the main contribution is then the geometry-guided token selection along the noise-matched KV injection. The paper introduces SceneLongBench which is a useful setting for long minute-long scenes with moving cameras.

Critical strengths

The paper proposed a concrete implementation to solve a real problem without training or adapters. The idea of passing token indices between models is also interesting and practical (allowing models to communicate without additional training).

Results are somewhat thorough with various 3D metrics that are relevant to the proposed setting.

The paper is good at acknowledging trade offs, with ablations over confounds, and clear about the inferrnece overhead that the method encompasses.

Critical weaknesses

The interpretation of the scores is not very clearly discussed. The model’s certainty over token reconstruction is not a direct measurement of 3D disagreement with previous views. The authors could have greatly motivated the paper if they had shown that the selected tokens indeed have lower reconstruction/cross-view error.

The evaluation needs some form of camera motion adherence metric, since lower cross-view error from selected tokens might simply be tokens with low camera motion. A degradation on VBench scores might be related to that (Lower dynamic degree).

Questions for authors

As per the weakness, in particular do selected tokens have measurably lower cross-view error? What happens with absent, reset, or shuffled geometry history?

3 comments
AuthorsOpenReview ↗

Response to Reviewer LsnC

We thank the reviewer for the meaningful and constructive feedback. We address each concern below.

W1 / Q1. Do selected tokens have measurably lower cross-view error?

At each of the first two denoising stages (D0 and D1), we select tokens from the clean image decoded from the generator’s intermediate prediction. We compared these selected tokens with an equal-sized set of lowest-scoring tokens (Bottom-k) from the same prediction. We evaluated ten 60-second videos and included only tokens with valid historical matches. For each token, we used VGGT to project its estimated 3D point into the matched historical views and measured its 2D reprojection error.

Token groupD0 reprojection error ↓D1 reprojection error ↓
Selected4.95 ± 2.003.78 ± 1.55
Bottom-k5.35 ± 2.154.08 ± 1.80

Selected tokens have lower mean reprojection error than Bottom-k at both stages, by 7.5% at D0 and 7.4% at D1. This confirms that selected tokens do have measurably lower cross-view error, so our confidence score is positively associated with geometric reliability rather than being an unrelated certainty signal.

W1 / Q2. Effect of absent, reset, and shuffled geometry history

We followed the setup of the ablation study in Figure 6 and used both confidence and attention for token selection in all settings. In the Absent setting, we omitted the input from previous video blocks. In the Reset setting, we reset the TTT3R state to zero before each token selection step. Absent and Reset produce identical inputs to the token selector, so we report a single result for both. In the Shuffled setting, we shuffled the historical frames before passing them to TTT3R. In the Default setting, we processed the frames in chronological order and retained the TTT3R state across video blocks.

Geometry historyPSNR ↑SSIM ↑LPIPS ↓EPI ↓RPX ↓RPT ↓
Absent/reset17.830.5710.3617.610.3795.44
Shuffled17.870.5730.3597.570.3785.37
Default (ours)17.940.5760.3567.520.3755.22

The default setting performed best across all six metrics.

W2. Camera motion and dynamic degree

For all methods in Table 1 of the main paper, here we report VGGT-estimated camera rotation speed and angular span alongside VBench Dynamic Degree (DD). Angular span is the maximum angle between viewing directions within an evaluation window.

MethodRotation speed (°/s)Angular span (°)DD (%)
Deep Forcing0.45 ± 0.261.36 ± 0.8131.00
Causal Forcing11.14 ± 2.5264.86 ± 15.1099.67
Relax Forcing8.36 ± 9.9427.39 ± 28.0792.00
Token Forcing (ours)10.73 ± 2.6763.06 ± 15.99100.00

Our proposed method, Token Forcing, preserves camera motion and achieves the highest DD.

Reviewer_LsnCOpenReview ↗

My concerns have been clarified

The authors addressed my concerns by adding more experimental results that are in line with the paper conclusions. The reprojection results satisfied my initial concerns on the geometric validity of the used scores, and that lowest-scoring results are less reliable, even if improvements are not large. Similarly, I appreciated the authors efforts in running the absent/shuffled/default on geometry history, it seems like the effect is very small and maybe directionally indicating that the historical state is somewhat useful. I am also satisfied with seeing that rotation speed is similar to that of causa forcing and I appreciate the addition of this experiment. I will mantain my weak acceptance score and encourage the authors to include the discussions in the main paper, specifically related to camera motion speed which I initially expected to be lower than competitors.

Thank you for your thoughtful and positive feedback. We are glad the additional experiments addressed your concerns. We will include these results in the revised paper and discuss the comparison of camera motion speeds in the main text.

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

This paper proposes Token Forcing, a training-free framework for improving 3D geometric consistency in long-horizon autoregressive video generation.

The authors observe that autoregressive video generators often suffer from accumulated geometric drift, including structural deformation, layout inconsistencies, and contradictions across camera viewpoints. Existing context-management methods typically rely on temporal positions, attention scores, or feature similarity without explicitly considering geometric reliability.

To address this limitation, Token Forcing introduces two components:

History-conditioned Token Selection, which uses a frozen recurrent 3D foundation model to identify geometrically reliable tokens based on reconstruction confidence and attention to the accumulated scene state.

Stage-aligned KV Injection, which re-noises the selected content and extracts generator-native KV features at the corresponding denoising stage, injecting them into subsequent video blocks as soft geometric anchors.

The paper further introduces SceneLongBench, a benchmark comprising 100 static indoor and outdoor scenes for evaluating geometric consistency in 60-second T2V and TI2V generation.

Experiments demonstrate improvements in several cross-view consistency metrics over Causal Forcing, while largely preserving video quality on motion-rich prompts.

Critical strengths

S1. Important and practically relevant problem.

Maintaining 3D geometric consistency over long autoregressive rollouts is an important challenge. The paper focuses on structural consistency across camera viewpoints, complementing existing work that primarily evaluates visual quality and temporal smoothness.

S2. Interesting geometry-guided token selection strategy.

The proposed method goes beyond conventional training-free context management based on recency, attention, or key similarity. Using a pretrained geometry foundation model to identify reliable tokens at a fine-grained level is an intuitive and potentially useful direction.

S3. No additional training is required.

Both the video generator and the recurrent geometry model remain frozen, and the approach does not require a learned feature-space adapter. This makes the method relatively straightforward to integrate into existing pretrained generators from a training perspective.

S4. Multiple geometric metrics and useful ablation studies.

The evaluation includes reconstruction-based and correspondence-based metrics, while the ablations investigate random versus geometry-guided token selection, confidence and attention signals, and different injection configurations.

These experiments provide encouraging evidence that geometry-aware token selection can improve cross-view consistency.

Critical weaknesses

W1. The geometric evaluation may not fully establish genuine 3D consistency.

A major concern is that the primary geometric evaluation relies on learned reconstruction and correspondence estimation, rather than independent ground-truth 3D geometry.

Specifically, the proposed token selection uses the TTT3R geometry model, while reconstruction-based evaluation relies on VGGT. Although these are different models, both rely on learned geometric priors and may exhibit correlated systematic biases.

Furthermore, lower reprojection or epipolar errors do not necessarily imply that the generated video follows the intended camera trajectory or reconstructs the correct underlying scene geometry.

For example, a model that produces smaller camera movements or more conservative viewpoint changes may achieve lower geometric errors without genuinely improving its ability to model complex 3D scenes.

The current evaluation does not sufficiently disentangle geometric consistency from camera-motion fidelity.

Additional evaluations of camera trajectory accuracy, independent geometric correctness, and human-perceived spatial consistency would substantially strengthen the central claims.

W2. Significant inference overhead despite the training-free formulation.

The training-free nature of Token Forcing is presented as an important practical advantage. However, Appendix E explicitly states that the proposed method approximately doubles video generation time under the reported experimental configuration.

This overhead arises from decoding intermediate predictions, executing the recurrent 3D reconstruction model, and performing additional generator passes to construct the injected KV features.

Although avoiding retraining is valuable, inference efficiency is itself a critical consideration in autoregressive and streaming video generation.

The paper does not provide a detailed runtime breakdown or comparisons under matched inference-time computational budgets.

For example, it remains unclear whether spending the same additional computation on more denoising steps, expanded temporal context, or alternative test-time refinement strategies could achieve comparable geometric improvements.

A systematic evaluation of the geometric-quality versus computational-cost trade-off is necessary to establish the practical benefit of Token Forcing.

W3. Limited generalization beyond static scenes.

SceneLongBench contains 100 static indoor and outdoor scenes in which only the camera moves. While this controlled setup is appropriate for evaluating cross-view reconstruction consistency, it covers only a restricted subset of long-video generation scenarios.

Real-world long videos frequently contain moving objects, dynamic occlusions, non-rigid deformations, object interactions, and evolving scene states.

In such cases, reconstruction confidence may reflect the difficulty of modeling dynamic content rather than the reliability of its consistency with historical geometry.

Consequently, the proposed token selection mechanism might preferentially preserve static background regions while underrepresenting moving objects or newly appearing scene elements.

Although the paper includes experiments on motion-rich prompts, these evaluations focus primarily on VBench scores and do not directly measure geometric consistency in dynamic scenes.

Therefore, the current results convincingly support the approach for static-scene camera-motion generation, but provide limited evidence for its general applicability to dynamic long-form video generation.

W4. Insufficient validation of token reliability and incomplete isolation of injection design choices.

The proposed method assumes that high reconstruction confidence and strong state attention identify tokens that are geometrically consistent with previously generated content.

However, high reconstruction confidence does not necessarily imply correctness or historical consistency, particularly when a geometry model confidently predicts an incorrect structure.

Similarly, the correspondence between state attention and actual geometric consistency is not directly established.

The ablations in Figure 6 show that confidence-only selection already achieves most of the benefits. Adding state attention provides relatively modest further improvements, such as reducing EPI from 7.57 to 7.52 and RPT from 5.35 to 5.22.

This raises questions about the independent contribution of the attention-based selection signal.

Furthermore, comparisons of injection configurations do not fully isolate noise-level alignment from changes in injection location, stage coverage, or the amount of additional context. As acknowledged in Section 4.3, the two-stage injection setting also uses a larger total token budget than either single-stage variant.

More carefully controlled ablations, including matched token budgets, deliberately mismatched noise levels, and direct analysis of the geometric errors of selected tokens, would provide stronger evidence for the proposed design.

W5. Limited statistical analysis and lack of human perceptual evaluation.

Although Token Forcing improves several geometric metrics, the improvements are not uniform across all settings.

For instance, under outdoor TI2V generation, EPI increases from 4.95 to 5.03, while the average VBench score decreases from 86.21 to 85.41.

Furthermore, the gains on motion-rich prompts are relatively small. At 60 seconds, the average VBench score increases only from 85.72 to 85.84 over Causal Forcing.

The paper does not report confidence intervals, standard deviations, or statistical significance tests for these results. The component ablations are also conducted on only 20 randomly sampled scenes with a fixed seed.

Without additional statistical analysis, it is difficult to determine how consistently the smaller improvements would generalize across different prompts and sampling seeds.

Moreover, the paper lacks systematic human evaluation to establish whether the improvements in reconstruction-based metrics translate into perceptually meaningful gains in geometric consistency.

A pairwise human GSB evaluation assessing structural consistency, natural motion, and overall video quality would strengthen the empirical conclusions.

Questions for authors

Q1. Geometric Evaluation and Camera Motion.

Since the primary geometric metrics rely on learned reconstruction models, could the authors provide additional evidence that the improvements correspond to genuine 3D consistency and accurate camera motion? In particular, how can the evaluation rule out the possibility that lower geometric errors result from reduced camera motion or more conservative viewpoint changes?

Q2. Inference Efficiency and Computational Fairness.

Appendix E states that Token Forcing approximately doubles generation time. Could the authors provide a detailed runtime breakdown and comparisons under matched inference-time computational budgets? For example, would additional denoising steps or simpler test-time refinement methods achieve comparable geometric improvements?

Q3. Generalization to Dynamic Scenes.

How does Token Forcing perform in scenes containing moving objects, dynamic occlusions, non-rigid motion, or newly appearing entities? Could the authors provide geometric consistency evaluations on dynamic scenes rather than relying primarily on aggregate VBench metrics?

Q4. Reliability of Token Selection and Stage-aligned Injection.

Does high reconstruction confidence consistently correspond to geometrically reliable tokens? Could the authors quantitatively evaluate the geometric correctness of selected tokens and compare different injection stages and noise-level alignment strategies under matched token budgets?

Q5. Statistical and Human Evaluation.

Could the authors report uncertainty estimates or statistical significance tests for the main quantitative results? Additionally, could they conduct a pairwise human preference evaluation, allowing ties, to assess whether the reported improvements translate into perceptually meaningful gains in structural consistency and overall video quality?

2 comments
AuthorsOpenReview ↗

Response to Reviewer UTGc (Part 1)

We thank the reviewer for the constructive feedback and thoughtful questions. We address each point below.

W1 / Q1. Geometric evaluation and camera motion

3D consistency evaluation. In text-to-video and text-image-to-video generation, no ground-truth 3D geometry or reference camera trajectories are available for the generated videos. To test whether the observed gains in 3D consistency also hold with a different estimator, we repeat the evaluation with Depth Anything 3 (DA3) [1], another strong 3D geometry foundation model. We use the same videos and evaluation protocol as in the experiments reported in Table 1 of the main paper.

MethodPSNR ↑SSIM ↑LPIPS ↓EPI ↓RPX ↓RPT ↓
Deep Forcing16.540.520.4010.170.4210.61
Causal Forcing17.020.570.358.770.378.82
Relax Forcing16.480.500.5118.300.5319.85
Token Forcing (ours)17.200.590.347.420.358.05

With DA3, Token Forcing still achieves the best score on all six metrics.

Camera motion and viewpoint coverage. Here we measured rotation speed and angular span with VGGT for all methods in Table 1. Angular span is the maximum angle between viewing directions within a window. VBench Dynamic Degree (DD) measures overall video motion.

MethodRotation speed (°/s)Angular span (°)DD (%)
Deep Forcing0.45 ± 0.261.36 ± 0.8131.00
Causal Forcing11.14 ± 2.5264.86 ± 15.1099.67
Relax Forcing8.36 ± 9.9427.39 ± 28.0792.00
Token Forcing (ours)10.73 ± 2.6763.06 ± 15.99100.00

Token Forcing exhibits rich camera motion and achieves the highest DD score (100%).

W2 / Q2. Inference overhead and runtime-matched comparisons

We followed the setup in Sec. 4.3 and used the same 20 prompts to generate 60-second videos. On average, Token Forcing took 2.15× as long as the 4-step Causal Forcing baseline. Of the added time, an estimated 51% went to TTT3R reconstruction, 17% to intermediate VAE decoding, and 32% to KV extraction and other overhead.

MethodTime per video (s) ↓EPI ↓RPX ↓RPT ↓VBench Avg. ↑
4-step Causal Forcing342.98.840.3825.6686.43
26-step Causal Forcing744.015.640.5358.3284.32
4-step Causal Forcing, 8× history window700.89.950.4936.6978.45
Token Forcing737.27.520.3755.2286.58

At comparable runtimes, Token Forcing achieved lower geometric errors than both the 26-step baseline and the baseline with an 8× history window.

W3 / Q3. Generalization to dynamic scenes

We evaluated 60-second videos on the 100 motion-rich prompts from Causal Forcing (Fig. 4). To answer your question, we further evaluate consistency using two complementary metrics:

  • EDINOE_{\rm DINO} measures feature consistency across frames after compensating for motion. It uses optical flow to locate corresponding points in adjacent frames and compares their DINO features. Both visible moving objects and the static background contribute to the score.

  • EgeoE_{\rm geo} measures the geometric stability of the static background across viewpoints. In regions classified as static and visible, it measures disagreement between estimated optical flow and the rigid flow predicted from depth and camera motion. It also checks whether the reprojected depth agrees with the depth in the next frame.

MethodEDINOE_{\rm DINO} ↓EgeoE_{\rm geo} ↓
Deep Forcing0.1440.310
Causal Forcing0.1440.125
Relax Forcing0.1700.379
Token Forcing (ours)0.1370.124

Token Forcing achieves the lowest EDINOE_{\rm DINO} and EgeoE_{\rm geo} on motion-rich prompts, demonstrating better feature consistency for moving content and better background geometric stability in dynamic scenes.

AuthorsOpenReview ↗

Response to Reviewer UTGc (Part 2)

W4 / Q4. Reliability of token selection and stage-aligned injection

We address the two parts separately.

(i) Geometric correctness of selected tokens. Tokens are selected on the clean image decoded from the intermediate prediction x^0\hat{x}_0 at D0D_0 and D1D_1. To test whether these selections are geometrically reliable, we compared the selected tokens with an equal-sized set of lowest-scoring tokens (Bottom-k) from the same prediction. For ten 60-second videos, we used VGGT to project each token’s estimated 3D point into matched historical views and measured its 2D reprojection error. We used only tokens with valid historical matches. Selected tokens have lower mean reprojection error than Bottom-k at both stages, by 7.5% at D0D_0 and 7.4% at D1D_1.

Token groupD0 reprojection error ↓D1 reprojection error ↓
Selected4.95 ± 2.003.78 ± 1.55
Bottom-k5.35 ± 2.154.08 ± 1.80

(ii) Noise-level alignment and injection mode under matched token budgets. We fixed Append and varied the source encoding noise level across Clean, Step k+1, and Step k (ours). We then fixed Step k and compared injection modes Replace, Mask-unselected, and Append (ours). All settings use matched token budgets. Matching the receiver’s noise level (Step k) clearly outperforms Clean and Step k+1. With Step k fixed, Append gives the lowest EPI and highest VBench, with RPT comparable to Step k+1, while Replace and Mask-unselected are worse on RPT and VBench. Masking unselected tokens is therefore unnecessary, and appending leaves them in context without harming geometry.

Source encodingInjection modeEPI ↓RPX ↓RPT ↓VBench ↑
CleanAppend12.220.52111.8185.25
Step k+1Append8.040.3835.1986.10
Step kReplace7.860.4015.8186.51
Step kMask-unselected7.830.4125.5086.21
Step kAppend (ours)7.520.3755.2286.58

W5 / Q5. Statistical analysis and human evaluation

We thank the reviewer for the suggestion and agree that these experiments would strengthen the paper. The experiments are underway. We will share the results during the discussion period if available, or include them in the final version otherwise.

Reviewer_r5EwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Token Forcing, a training-free method that mitigates scene drift in autoregressive video diffusion by pairing a frozen generator (Causal Forcing) with a frozen recurrent 3D foundation model (TTT3R), which selects geometrically reliable tokens to condition subsequent blocks. The authors also introduce SceneLongBench, a benchmark of 100 static indoor and outdoor scenes for 60-second T2V and TI2V generation. Compared with Causal Forcing, Token Forcing reduces reprojection error in all four SceneLongBench settings.

Critical strengths
  1. The paper is well-organized and easy to read.

  2. The paper mitigates scene drift in autoregressive video diffusion with a novel training-free approach: a frozen 3D foundation model guides a frozen generator purely through token selection. This decoupling, in principle, lets different generators and 3D models be freely combined and reused.

Critical weaknesses
  1. Unjustified KV injection design. Selected tokens are re-noised to the receiver's stage and appended to its context, but neither choice is justified. During training, the base model never sees cross-block context at its own noise level, so the rationale for same-level encoding is unclear. Moreover, appending leaves unselected, possibly unreliable tokens in context. Since stage-aligned injection is a core contribution, it remains unclear where the gains come from.

  2. Limited evidence of generality. The method is presented as training-free and adapter-free, yet all experiments use a single generator (Causal Forcing on Wan2.1-T2V-1.3B). This leaves its applicability to other autoregressive video diffusion models unverified.

Questions for authors
  1. Could the authors justify the choice of same-level encoding and the appending design, and ablate both the encoding noise level and the injection mode?

  2. Could the authors evaluate the method on other autoregressive video diffusion models to further demonstrate its generality?

  3. Minor errors (the misspelling "proceed" for "proceeds" in Fig. 1 caption; the citation "(Liu et al., 2025b;?)" is broken).

3 comments
AuthorsOpenReview ↗

Response to Reviewer r5Ew

We are grateful to the reviewer for acknowledging the novelty of our method. We respond to each concern in turn below.

W1 / Q1. Encoding noise level and KV injection mode

To further explain our design choices, we ran ablations following Sec. 4.3. Across all settings, we kept the receiver at denoising step k and used the same token budget. With Append fixed, we compared three noise levels for source encoding (Clean, Step k+1, and Step k). With both blocks at step k, we compared three KV injection modes (Replace, Mask-unselected, and Append).

Source encodingInjection modePSNR ↑SSIM ↑LPIPS ↓EPI ↓RPX ↓RPT ↓VBench Avg. ↑
CleanAppend15.860.490.49212.220.52111.8185.25
Step k+1Append17.420.560.3608.040.3835.1986.10
Step kReplace17.030.550.3827.860.4015.8186.51
Step kMask-unselected16.880.540.3817.830.4125.5086.21
Step k (ours)Append (ours)17.940.580.3567.520.3755.2286.58

Our design performs best on six of the seven metrics.

Why encode at the receiver’s noise level? During training, tokens within a block attend to one another at a shared noise level. This motivated us to test the same alignment across blocks at inference. FIFO-Diffusion [1] and History-Guided Video Diffusion [2] also adjust noise levels at inference. In our ablation, matching the receiver’s noise level performed best overall, although next-stage encoding gave a slightly lower RPT.

Why append the selected KV features? Each region has a fixed selection budget, so unselected tokens can still provide useful scene context. Appending preserves that context while giving the receiver additional access to reliable positions. Deep Forcing [3] likewise augments a frozen generator’s KV context at inference. Append outperformed both Replace and Mask-unselected across all metrics. Under the same append procedure, geometry-based selection reduced epipolar error, while random selection left it unchanged (Table 3 in the main paper).

W2 / Q2. Generalization to other autoregressive generators

Here we evaluated Token Forcing on SkyReels-V2-DF-1.3B-540P for 60-second text-to-video generation using the same 20 prompts as in Sec. 4.3. Token Forcing improved all six multi-view consistency metrics, reducing RPT by 21.8% (3.48 to 2.72) while maintaining VBench quality (72.07 to 72.09).

MethodPSNR ↑SSIM ↑LPIPS ↓EPI ↓RPX ↓RPT ↓VBench Avg. ↑
SkyReels Baseline22.990.780.223.290.233.4872.07
+ Token Forcing23.040.790.213.200.222.7272.09

Q3. Minor spelling errors.

We thank the reviewer for pointing this out. We have corrected it in the revision.

[1] Jihwan Kim et al. FIFO-Diffusion: Generating Infinite Videos from Text without Training. NeurIPS, 2024.

[2] Kiwhan Song et al. History-Guided Video Diffusion. ICML, 2025.

[3] Jung Yi et al. Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression. ICML, 2026

Reviewer_r5EwOpenReview ↗

Concerns Addressed – Score Raised

Thanks the authors for their detailed response. The authors ran thorough comparison experiments on the encoding noise level, and they also tested the method on additional models. These results answer my questions, so I will raise my score to Weak Accept.

Thank you very much for reviewing the additional experiments and raising your score. We truly appreciate your feedback.

Reviewer_GS4QOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Token Forcing aims to reduce geometric drift in long autoregressive videos. A frozen recurrent 3D model scores regions in intermediate predictions. The scores combine reconstruction confidence and attention to the scene state. The video generator re-encodes the source prediction at a matching noise level. KV features from selected positions are then added to the next block. Neither model’s weights are updated.The paper introduces SceneLongBench, with 100 static indoor and outdoor scenes for 60-second T2V and TI2V evaluation. RPX and RPT improve across all four settings. Ablations test the selection rule, selection signals, and injection settings. Tests on motion-rich prompts show similar average video quality to the base model. Generation time roughly doubles.

Critical strengths

1.A useful approach to local content selection. A generated block can contain both reliable and inconsistent regions. The method uses external geometry signals to choose which regions receive extra emphasis in later generation. The 3D model supplies positions, while the generator supplies its own features. This allows the two frozen models to work together without a learned feature adapter.
2.The comparisons support the selection rule. RPX and RPT improve across the evaluated settings. The random-selection control uses the same validity filter, regional budgets, and injection process. The proposed ranking performs better than random selection. This supports the value of choosing which content to reuse.

Critical weaknesses

1.The contribution of persistent history remains unclear. The confidence-only variant still uses historical state. It does not separate the effect of long history from local reconstruction confidence. The experiments show that the ranking helps, but provide less evidence about how much persistent history contributes.
2.The long-range evaluation needs clearer reporting. The paper does not clearly specify the time gaps between evaluated frames or the coverage of valid matches. Differences in camera movement are also not quantified. Smaller viewpoint changes could make geometric consistency easier to achieve.

Questions for authors

1.What evidence supports the contribution of persistent history beyond local reconstruction confidence?
2.What time gaps do the geometric metrics cover? How are frames and valid matches selected? How do actual camera motion and viewpoint coverage compare across methods?

3 comments
AuthorsOpenReview ↗

Response to Reviewer GS4Q

We thank the reviewer for providing constructive comments and for recognizing our work. We address each concern below.

W1 / Q1. Does history help when tokens are selected by confidence alone?

Yes. Following the ablation setup in Figure 6, we use reconstruction confidence as the sole criterion for token selection in both settings. We vary only how we handle the TTT3R state:

  • Reset state: We reset the TTT3R state before each selection step.
  • Persistent historical state: We retain the TTT3R state accumulated over previous video blocks.

Retaining the state reduces EPI from 7.72 to 7.57, RPX from 0.381 to 0.378, and RPT from 5.62 to 5.35. These results show that persistent history improves geometric consistency even when token selection relies solely on reconstruction confidence.

W2 / Q2. Evaluation details: time gaps and how the tokens are selected.

Following the evaluation framework of VIGOR [1], we evaluate each 60-second video using 12 overlapping 6-second windows. We uniformly sample 10 frames per window.

To select valid matches, we exclude sky regions and points with invalid depth. We retain only correspondences whose projections and tracks fall within the image bounds and whose confidence is at least 0.05 in both views.

W2 / Q2. Camera motion and viewpoint coverage.

For each method in Table 1 of the main paper, we report camera rotation speed and angular span, both estimated using VGGT. We also report the VBench Dynamic Degree (DD) score.

MethodRotation speed (°/s)Angular span (°)DD (%)
Deep Forcing0.45 ± 0.261.36 ± 0.8131.00
Causal Forcing11.14 ± 2.5264.86 ± 15.1099.67
Relax Forcing8.36 ± 9.9427.39 ± 28.0792.00
Token Forcing (ours)10.73 ± 2.6763.06 ± 15.99100.00

Token Forcing demonstrates rich camera motion and achieves the highest DD score (100%).

[1] T. Yin et al. VIGOR: Video Geometry-Oriented Reward for Temporal Generative Alignment. ECCV, 2026.

Thank you for the detailed rebuttal and for providing the additional clarifications and paper details. They have addressed my questions well. After carefully reviewing your response, I will maintain my original score.

Thank you for reading our rebuttal carefully. We are glad that the additional results and clarifications addressed your questions. We have incorporated them into the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.