acceptodds
Under review as a conference paper at ICLR 2027

Alaya-EVOKE: From Linear-Scaling Supervision to Endless World

21 views

Abstract

Interactive world models must simultaneously support persistent memory, responsive user interaction, and long-horizon generation, yet these requirements place conflicting demands on the underlying model. Maintaining history in the denoiser context or key-value cache incurs growing cost over time, forcing a trade-off between session length and retained memory, while low-latency interaction typically relies on few-step generation whose capabilities are ultimately bounded by its teacher. Alaya-EVOKE (Evoke) addresses both limitations by externalizing persistent world state and redesigning the teacher for long-horizon interactive generation. Scene geometry is maintained in an external, camera-indexed world state bank, from which only information relevant to the current view is retrieved, keeping the denoiser context bounded as the session grows. Rather than treating the teacher as a fixed high-quality generator, the teacher is explicitly designed for long-horizon supervision. Its sparse attention scheme combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, yielding linear growth in activation memory and computation while enabling supervision over long temporal horizons. Such supervision exposes content drift that remains locally plausible within short windows, while per-chunk conditioning enables prompt changes and event control throughout the sequence. A 30-second long-horizon distribution-matching objective, applied under self-forced rollouts, transfers both capabilities to a three-step student that uses no classifier-free guidance (CFG), improving resistance to long-term content drift while preserving responsive conditioning. With bounded context and recurrent external memory, Evoke supports open-ended, continuously evolving generation; on a single H200 at , each  s chunk is generated in  s. With three sampling steps, Evoke ranks first on the WBench navigation split and achieves competitive results on VBench-2.0, VBench-Long, and VBench-I2V.

Primary area:
generative models
Keywords:
World Models, Long-Horizon Video Generation, Interactive Generation

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 3 reviews, above 85% of the others.
Reviewer_xrzzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper externalizes the geometric memory of camera indices, and constructs a long-sequence teacher by combining sparse attention and global linear attention. It performs three steps of supervised training via joint distribution matching in approximately 30 seconds for a CFG-free student. The system supports continuous generation with a fixed memory budget, aiming to balance spatial revisit, long-term stability and segmented text control.

Critical strengths
  • The external geometry library handles spatial recall, teachers provide long-term supervision, and students maintain bounded context; fixed storage and retrieval budgets impose clear computational constraints for continuous operation.
  • In addition to the 2 hour qualitative example, statistical analysis of 8 65.5mins sequences is reported to conduct actual measurements of degradation issues during continuous operation.
Critical weaknesses
  • Section 4.3 acknowledges that the content descriptor and event realization rate do not significantly distinguish short and long teachers, with clear benefits concentrated on photometric stability. Existing evidence is insufficient to attribute content drift suppression and dynamic interaction capability to long supervision windows, and the effects of teacher quality, training volume and window length need to be disentangled.
Questions for authors
  • Varying only the supervision window (with fixed teacher, data, and budget), can achieve consistent gains in content metrics like returning individual identification and spatial layout?
  • How does VBench perform when using multiple random seeds without prompt tuning for evaluation samples? Do all methods in WBench adopt consistent prompt processing and result selection rules?
  • How to allow users update existing geometry while retaining the memory of unmodified areas? Please provide the success rate and consistency results of revisiting after editing.
3 comments
AuthorsOpenReview ↗

Supervision-Length Ablations, Revisit Memory, Prompt Processing, and Event Control

Comment

Thank you for the careful review and recognition of our geometric memory, long-horizon supervision, and bounded computation. Your feedback helps clarify supervision, prompt processing, and event control.

1 · W1 · Benefits of Long-Horizon Supervision

Thank you for your feedback; it will help clarify benefit attribution. The redesigned teacher responds to timed prompts under both long and short supervision. Students learn this ability through distillation, so similar event-realization rates match our design. Longer supervision targets accumulated color and exposure drift, including excessive brightness, darkness, and overexposure. Length extrapolation enables joint scoring of 20 student-generated chunks (approximately 30 seconds); Appendix Fig. A5 shows extrapolation and instruction responses.

To isolate supervision length, we use the same fixed teacher checkpoint and Stage 2 student initialization, with training data and other settings fixed. We vary only 2, 8, and 16 chunks and compare at training step 500, holding teacher training state and training-data volume constant.

Given the full evaluation's scale and limited time and resources, we sample 200 VBench-Long prompts across categories and expand them under fixed guidelines. Generation and scoring settings are identical. Table 1 averages 3 fixed random seeds; semantic metrics use the original prompts.

Table 1. VBench-Long (step 500).

WindowIQAQSubjBGMSDDSceneOverallAVG
2 chunks52.5654.9297.1196.4798.8578.9124.5624.6666.00
8 chunks53.0155.3997.0896.6898.8178.8127.5924.6966.51
16 chunks53.5255.4697.2896.6298.8178.9128.1524.6866.68

IQ/AQ: imaging/aesthetic quality; Subj/BG/Scene/Overall: consistency; MS: motion smoothness; DD: dynamic degree. All ↑.

Relative to 2 chunks, 16 chunks improve imaging quality, aesthetic quality, background consistency, and scene consistency; 8 chunks improve the latter three. Larger windows thus benefit several quality and consistency metrics.

Using 40 rewritten prompts spanning subjects, actions, and scenes, we generate approximately 30-second videos (20 inference chunks) with all three step-500 models under identical settings. Only the final 5 seconds are scored for imaging and aesthetic quality.

Table 2. Final 5-second quality (~30-second videos, step 500).

Metric2 chunks8 chunks16 chunks
IQ / AQ ↑ (×100)47.86/54.4748.11/55.8848.40/55.44

Both longer windows improve mean final-segment quality over 2 chunks, supporting long-horizon supervision for suppressing accumulated color and exposure deviations and improving visual stability.

We also compare event-triggering success rates:

Table 3. Event success by supervision window.

Metricw/o DMD2 chunks8 chunks16 chunks
Success ↑44% (22/50)62% (31/50)64% (32/50)62% (31/50)

Success rates are similar across windows and exceed 44% (22/50) without Stage 3, supporting transfer of event responses through distillation.

2 · Q1 · Supervision, Memory, and Camera Control

Long-horizon supervision suppresses accumulated color drift and maintains visual quality; retrieval and geometric reprojection provide cross-view scene memory and revisit consistency. Camera pose selects historical observations from the world state bank and defines their reprojection into the target view. This jointly supports scene recall and camera control (Fig. 4).

We compare revisit content similarity at step 500 to separate generation-quality and scene-memory effects.

Table 4. Revisit similarity (step 500; G: global, L: local; all ↑).

Window12s G12s L24s G24s L48s G48s LAVG
2 chunks0.83460.57050.78340.49740.71710.42270.6376
8 chunks0.83460.56730.78020.49480.72580.42340.6377
16 chunks0.83730.55980.77960.48580.71170.41420.6314

Evaluation covers 12 scenes × 3 seeds. Global/local similarity uses cosine similarity of CLS/aligned patch features. AVG equally averages the six unrounded means.

Revisit similarity does not improve with longer supervision; global scores remain comparable. This agrees with Sec. 3.2 and Fig. 4: scene memory relies primarily on world-state retrieval and geometric reprojection.

3 · Q2 · Prompt Processing and Length Ablations

Thank you for your feedback; it will help clarify the final manuscript. Video models typically learn from detailed descriptions, while evaluation prompts may be brief. Expanding prompts while preserving intent follows existing practice: CogVideoX aligns inference text with training captions; VPO studies the detailed-caption/short-input gap; Wan2.1 provides prompt extension.[1][2][3]

GPT expands prompts under predefined guidelines: preserve subject, action, and scene intent, add details, and meet the required length. Outputs with incorrect lengths or altered intent are rewritten under the same rules.

Table 5 compares three versions of the same 200 prompts using the paper's model checkpoint, with duration, resolution, and sampling fixed. Semantic metrics use the original prompts throughout.

Table 5. Prompt length (mean words).

PromptIQAQSubjBGMSDDSceneOverallAVG
Paper66.6258.8498.0897.9798.9674.3354.2226.4771.94
Original60.7656.6098.1797.2398.7350.4528.5925.2364.47
A (~30)61.5956.9697.5397.0598.7680.5042.1825.2269.97
B (~120)67.4059.1698.0297.9398.9873.2356.5627.0172.29

Both expansions improve mean imaging quality, aesthetic quality, and dynamic degree. Overall scores stabilize beyond a certain prompt length; overly short prompts may deviate from the training-text distribution and yield larger score shifts.

References: [1] CogVideoX (2024), Appendix E; [2] VPO (ICCV 2025), Prompt optimization; [3] Wan2.1, Prompt extension.

4 · Q3 · Historical Geometry and Event Control

Historical geometry maintains world state and revisit memory; users trigger events through prompts without editing that geometry. Segment-specific prompts control auroras, meteor showers, crystal electrical storms, and stargate activation.

Using released Evoke weights, we test 50 prompts across these four event types. Event descriptions appear in chunks 5–7; the original prompt returns at chunk 8. Geometry retrieval, camera trajectories, and memory updates remain unchanged. Visual inspection of sampled frames counts an event as successful if it clearly appears during the specified interval and was absent beforehand.

Table 6. Event triggering.

EventAuroraMeteor showerCrystal stormStargateTotal
Success ↑69.23% (9/13)84.62% (11/13)83.33% (10/12)100.00% (12/12)84.00% (42/50)

42 of 50 prompts trigger the target event (84%). These events can thus be triggered without explicitly editing historical geometry.

Thank you for providing these additional results, which address most of my previous concerns. While I still have some reservations regarding the novelty of the proposed method, leading me to lower the soundness score to 1. But I recognize the extensive experiments provided by the authors. Accordingly, I will maintain my original overall rating.

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

The paper introduces Alaya-EVOKE, an interactive world model combining an external geometric world-state bank, a teacher designed for long-horizon supervision, and a three-step distilled generator. The system supports camera control and timed text changes while keeping per-step computation bounded as sessions grow.

Critical strengths
  1. The separation of persistent spatial memory from the denoiser is well motivated, and the paper explains how fixed storage and retrieval budgets enable sustained generation.
  2. The complete system achieves strong reported WBench results and demonstrates impressive two-hour rollouts.
  3. Matched teacher comparisons provide useful evidence that long supervision improves photometric stability. The authors also candidly report limitations in reconstruction fidelity, content consistency, and editing previously anchored geometry.
Critical weaknesses
  1. The conceptual novelty lies primarily in the integration and training design, since geometric memory, sparse attention, and distribution-matching distillation build on established techniques.
  2. Fig. 6 shows clear improvements in brightness and other photometric properties, but it is less clear whether long supervision helps preserve scene content over time, since the content descriptor shows no significant difference between the students.
  3. Generating 1.5 seconds of video takes 2.11 seconds before additional overhead (< 1x), so the current system does not establish real-time end-to-end interaction.
  4. The authors claim full training setup improves photometric stability, but it is harder to tell which changes drive the improvement. e.g., Fig. 6 does not separately isolate the contributions of sparse attention, global linear-attn, and long supervision. Bringing the teacher ablations and scoring-window comparisons from Appendix C into the main paper would make this clearer.
Questions for authors
  1. Which gains arise from the teacher’s attention architecture, training horizon, and student supervision window individually?
  2. Can stronger identity-/layout-sensitive evaluations demonstrate reduced content drift?
  3. What is the full e2e latency, including geometry retrieval, rendering, depth estimation, and decoding? What is the overhead of the World Bank?
3 comments
AuthorsOpenReview ↗

Methodological Contributions, Long-Horizon Supervision and Geometric Memory, and Deployment Overhead

Comment

Thank you for the careful review and suggestions on our contributions, supervision, and deployment. Your feedback helps clarify module roles and evaluation protocols. We combine W2/W4 and W3/Q3 below.

1 · W1 · Conceptual Contributions of the Method

Thank you for your feedback; it will help clarify our contributions in the final manuscript. We build a teacher with length extrapolation and per-chunk event insertion, demonstrate that longer V2V supervision suppresses accumulated color and exposure drift, and integrate camera control with scene memory into a system for hour-scale continuous generation.

Sparse attention and per-chunk text conditioning enable sequences beyond the original training-video duration and responses to mid-sequence instructions (Appendix Fig. A5).

Historical retrieval and geometric reprojection follow prior work. Our contributions are the teacher redesign, long-horizon supervision, and their integration with camera control and external memory. We will clarify technical attribution in the revision.

2 · W2 / W4 · Supervision, Memory, and Camera Control

Long-horizon supervision suppresses accumulated color drift and maintains visual quality; retrieval and geometric reprojection provide cross-view scene memory and revisit consistency. Camera pose selects historical observations from the world state bank and defines their reprojection into the target view. This jointly supports scene recall and camera control (Fig. 4).

Sparse attention controls long-sequence cost, the global linear branch adds cross-chunk information, and per-chunk text conditioning enables timed event prompts. This allows the teacher to jointly score 20 student-generated chunks (approximately 30 seconds), making long-horizon supervision practical. Appendix Fig. A5 shows length extrapolation and instruction responses; Fig. 6 measures accumulated appearance deviations through brightness, saturation, and contrast.

We initialize from the same Stage 2 student checkpoint, hold training data and other settings fixed, and compare 2, 8, and 16 chunks of supervision at training step 500.

Given the full evaluation's scale and limited time and resources, we sample 200 VBench-Long prompts across categories and expand them under fixed guidelines. Generation and scoring settings are identical. Table 1 averages 3 fixed random seeds; semantic metrics use the original prompts.

Table 1. VBench-Long (step 500).

WindowIQAQSubjBGMSDDSceneOverallAVG
2 chunks52.5654.9297.1196.4798.8578.9124.5624.6666.00
8 chunks53.0155.3997.0896.6898.8178.8127.5924.6966.51
16 chunks53.5255.4697.2896.6298.8178.9128.1524.6866.68

IQ/AQ: imaging/aesthetic quality; Subj/BG/Scene/Overall: consistency; MS: motion smoothness; DD: dynamic degree. All ↑.

Relative to 2 chunks, 16 chunks improve imaging quality, aesthetic quality, background consistency, and scene consistency; 8 chunks improve the latter three. Larger windows thus benefit several quality and consistency metrics.

We compare revisit content similarity at step 500 to separate generation-quality and scene-memory effects.

Table 2. Revisit similarity (step 500; G: global, L: local; all ↑).

Window12s G12s L24s G24s L48s G48s LAVG
2 chunks0.83460.57050.78340.49740.71710.42270.6376
8 chunks0.83460.56730.78020.49480.72580.42340.6377
16 chunks0.83730.55980.77960.48580.71170.41420.6314

Evaluation covers 12 scenes × 3 seeds. Global/local similarity uses cosine similarity of CLS/aligned patch features. AVG equally averages the six unrounded means.

Revisit similarity does not improve with longer supervision; global scores remain comparable. This agrees with Sec. 3.2 and Fig. 4: scene memory relies primarily on world-state retrieval and geometric reprojection.

3 · W3 / Q3 · Real-Time Interaction, Deployment, and End-to-End Overhead

Sec. 4.2 reports 2.11 seconds per 1.5-second chunk on one H200. We tested H200 4+1 and 8+1 deployments: diffusion uses 4 or 8 GPUs, plus one for geometry and conditioning encoding.

Both use identical models, inputs, and settings: 640×384, 24 fps, three-step generation, BF16 DiT / FP32 VAE, and 1.5-second continuations. We measure 20 consecutive windows after warm-up.

Table 3. Steady-state performance, including geometry and output.

H200 setupMean (s) ↓P95 (s) ↓Client fps ↑
4+1, single node1.2391.25329.04
8+1, two nodes1.1881.22330.27

H200 4+1 produces 1.5 seconds of video in 1.239 seconds (29.04 fps), supporting sustained real-time generation. Timing includes geometry readiness, diffusion, decoding, and output. Client rates count complete JPEG reception, excluding network transmission and browser display. First-frame waits including warm-up are 20.55/20.14 seconds after model readiness. The 8+1 deployment uses two nodes and socket communication.

Table 4. Mean stage time (ms).

StageH200 · 4+1H200 · 8+1
DiT diffusion716.8577.1
VAE decoding507.6505.7
CPU image output33.733.9
Depth + preprocessing + bank update568.1575.8
History sampling + selection + coordinates22.130.7
Geometry reprojection + rendering282.9286.1
Condition VAE encoding283.4286.3
Geometry readiness wait0.191.3

Auxiliary-GPU geometry processing overlaps with diffusion and decoding; stage times are therefore not additive.

Single-GPU real-time generation remains challenging; current real-time performance uses multi-GPU parallelism and pipeline overlap. We will optimize memory access, conditioning encoding, and VAE decoding to reduce reprojection and encoding/decoding overhead and deployment cost.

4 · Q1 · Teacher and Student Factors

Attention architecture and per-chunk text conditioning support length extrapolation and event insertion; teacher training horizon affects learned temporal dependencies; student supervision windows determine the self-generated history scored jointly.

Teacher architecture and text conditioning. Sparse attention and the global linear branch support length extrapolation; per-chunk text conditioning enables timed event prompts. Supervision extends beyond the training-video duration; Appendix Fig. A5 shows four-minute event responses.

Teacher training horizon. This affects learned temporal dependencies and long-context scoring. Fig. 6 shows improved photometric stability for students distilled from long- versus short-horizon teachers.

Student supervision window. This extends DMD scoring over self-generated trajectories to address accumulated color and exposure deviations. Tables 1, 2, and 5 vary only 2/8/16 chunks with teacher, initialization, data, and updates fixed, reporting quality, revisit similarity, and final-segment quality.

Long-horizon supervision improves generation quality and photometric stability; scene memory relies primarily on retrieval and geometric reprojection.

5 · Q2 · Final-Segment Quality

Using 40 rewritten prompts spanning subjects, actions, and scenes, we generate approximately 30-second videos (20 inference chunks) with all three step-500 models under identical settings. Only the final 5 seconds are scored for imaging and aesthetic quality.

Table 5. Final 5-second quality (~30-second videos, step 500).

Metric2 chunks8 chunks16 chunks
IQ / AQ ↑ (×100)47.86/54.4748.11/55.8848.40/55.44

Both longer windows improve mean final-segment quality over 2 chunks, supporting long-horizon supervision for suppressing accumulated color and exposure deviations and improving visual stability.

Reviewer_MceuOpenReview ↗

response to the rebuttal

Thanks to the authors for the additional experiments. This has addressed most of my concern. I still strongly recommend authors update the manuscript with the clarification on the performance gains and the e2e breakdown, which may strengthen your work. While novelty is still limited, I will raise my score to 3.

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

Alaya-EVOKE uses a bounded geometric state bank to provide reprojected conditioning for the current generation, injecting this information at the low-resolution stage. Each chunk uses three denoising evaluations without classifier-free guidance (CFG). The long-horizon teacher combines bounded sparse context, global linear attention, and segmented text conditioning. The student is then trained through distribution matching over complete self-generated trajectories, with auxiliary camera supervision to preserve controllability.

The evaluation includes WBench, several VBench settings, photometric statistics for long sessions, revisit experiments with different memory-retention lengths, and control of dynamic events. The strongest evidence is that retained geometry improves short-term revisits and that the long-horizon teacher recipe improves photometric stability.

Critical strengths
  1. A sustainable system design with a measurable memory mechanism A fixed retention window, a fixed generation budget per chunk, and geometric reprojection provide a concrete engineering approach. Controlled experiments support the effectiveness of the memory, making the evidence more convincing than a long-video demonstration alone.

  2. Quantitative evidence for both stability gains and failure modes The long-horizon teacher recipe improves the student’s photometric stability.

Critical weaknesses
  1. The scope and source of the long-horizon supervision benefit remain unclear The clearest benefit to the student is photometric stability. The results do not show the same improvement in long-term content organization or the execution of dynamic conditions. The current comparison supports the overall effect of different teacher recipes, but the matching details do not sufficiently separate the supervision-window length from the teacher’s training state. It is therefore difficult to determine how much is gained by extending the supervision horizon itself. This limits how confidently the methodological contribution can be transferred to other systems.

  2. The benchmark rankings do not provide independent evidence of generalization Some of the most prominent generation-quality results come from prompt tuning on known evaluation cases. Fixed random seeds make these results closer to development performance on a predetermined set of cases. The rankings therefore do not directly establish a reliable advantage on new prompts and seeds, and in particular do not resolve the gap in long-horizon semantic evaluation.

Questions for authors
  1. With the same teacher checkpoint and otherwise matched training settings, does extending the distillation supervision window improve the student’s long-range scene identity or layout consistency?

  2. Does the reported quality advantage over the main baselines persist on held-out prompts and fresh seeds, with prompt processing fixed before evaluation?

3 comments
AuthorsOpenReview ↗

Supervision-Length Ablations, Prompt Processing, and Mechanism Clarification

Comment

Thank you for the careful review and recognition of our system design, geometric memory, and long-term stability. Your feedback helps clarify long-horizon supervision and our prompt-processing and evaluation protocols.

1 · W1 · Benefits of Long-Horizon Supervision

Thank you for your suggestion; it will help clarify the final manuscript. We use the same fixed teacher checkpoint, student initialization, training data, and other settings, varying only the supervision window across 2, 8, and 16 chunks and comparing students at the same training step.

Given the full evaluation's scale and limited time and resources, we sample 200 VBench-Long prompts across categories and rewrite them under a predefined procedure. Starting from Stage 2, we compare all windows at training step 500 with identical generation and scoring settings. Table 1 averages the same 3 generation seeds, randomly selected and fixed before evaluation.

Table 1. VBench-Long (step 500).

WindowIQAQSubjBGMSDDSceneOverallAVG
2 chunks52.5654.9297.1196.4798.8578.9124.5624.6666.00
8 chunks53.0155.3997.0896.6898.8178.8127.5924.6966.51
16 chunks53.5255.4697.2896.6298.8178.9128.1524.6866.68

IQ/AQ: imaging/aesthetic quality; Subj/BG/Scene/Overall: consistency; MS: motion smoothness; DD: dynamic degree. All ↑.

Relative to 2 chunks, 16 chunks improve imaging quality, aesthetic quality, background consistency, and scene consistency; 8 chunks improve the latter three. Larger windows thus benefit several quality and consistency metrics.

2 · W2 / Q2 · Prompt Processing and Length Ablations

Thank you for your feedback; it will help clarify the final manuscript. Video models typically learn from detailed descriptions, while evaluation prompts may be brief. Expanding prompts while preserving intent follows existing practice: CogVideoX aligns inference text with training captions; VPO studies the detailed-caption/short-input gap; Wan2.1 provides prompt extension.[1][2][3]

GPT expands prompts under predefined guidelines: preserve subject, action, and scene intent, add details, and meet the required length. Outputs with incorrect lengths or altered intent are rewritten under the same rules.

Table 2 compares three versions of the same 200 prompts using the paper's model checkpoint, with duration, resolution, and sampling fixed. Semantic metrics use the original prompts throughout.

Table 2. Prompt length (mean words).

PromptIQAQSubjBGMSDDSceneOverallAVG
Paper66.6258.8498.0897.9798.9674.3354.2226.4771.94
Original60.7656.6098.1797.2398.7350.4528.5925.2364.47
A (~30)61.5956.9697.5397.0598.7680.5042.1825.2269.97
B (~120)67.4059.1698.0297.9398.9873.2356.5627.0172.29

Both expansions improve mean imaging quality, aesthetic quality, and dynamic degree. Overall scores stabilize beyond a certain prompt length; overly short prompts may deviate from the training-text distribution and yield larger score shifts.

References: [1] CogVideoX (2024), Appendix E; [2] VPO (ICCV 2025), Prompt optimization; [3] Wan2.1, Prompt extension.

3 · Q1 · Supervision, Memory, and Camera Control

Long-horizon supervision suppresses accumulated color drift and maintains visual quality; retrieval and geometric reprojection provide cross-view scene memory and revisit consistency. Camera pose selects historical observations from the world state bank and defines their reprojection into the target view. This jointly supports scene recall and camera control (Fig. 4).

We compare revisit content similarity at step 500 to separate generation-quality and scene-memory effects.

Table 3. Revisit similarity (step 500; G: global, L: local; all ↑).

Window12s G12s L24s G24s L48s G48s LAVG
2 chunks0.83460.57050.78340.49740.71710.42270.6376
8 chunks0.83460.56730.78020.49480.72580.42340.6377
16 chunks0.83730.55980.77960.48580.71170.41420.6314

Evaluation covers 12 scenes × 3 seeds. Global/local similarity uses cosine similarity of CLS/aligned patch features. AVG equally averages the six unrounded means.

Revisit similarity does not improve with longer supervision; global scores remain comparable. This agrees with Sec. 3.2 and Fig. 4: scene memory relies primarily on world-state retrieval and geometric reprojection.

Using 40 rewritten prompts spanning subjects, actions, and scenes, we generate approximately 30-second videos with all three step-500 models under identical inference settings. Only the final 5 seconds are scored for imaging and aesthetic quality.

Table 4. Final 5-second quality (~30-second videos, step 500).

Metric2 chunks8 chunks16 chunks
IQ / AQ ↑ (×100)47.86/54.4748.11/55.8848.40/55.44

Both longer windows improve mean final-segment quality over 2 chunks, supporting long-horizon supervision for suppressing accumulated color and exposure deviations and improving visual stability.

Thank you for the author’s response. I still believe that providing leaderboard scores based on a fixed seed would help rule out any suspicion of "gaming" the leaderboard. I maintain my original rating of "weak accept."

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.