acceptodds
Under review as a conference paper at ICLR 2027

Pixelle-Video: A Cross-Shot and Cross-Modal Coordination Framework for End-to-End Video Production

Abstract

Recent video generation models can synthesize visually compelling short clips, yet producing a complete video requires coordinating persistent information across shots, modalities, and heterogeneous production stages. We introduce Pixelle-Video, a fully automated engine for end-to-end video production. A Stateful Storyboard persists throughout production, accumulating multimodal artifacts and execution metadata as a shared evolving state across stages. Building on this, Pixelle-Video introduces two coordination mechanisms: (1) Cross-Modal Temporal Coordination, which uses the duration of synthesized narration as a shared temporal clock for visual processing and clip composition; and (2) Persistent Visual Memory, which carries character, scene, and visual references across shots and combines multimodal inspection with targeted regeneration to maintain visual consistency. Evaluations on VBench, ViStoryBench and DirectorBench show that Pixelle-Video achieves the best performance across all nine dimensions among representative automated video-production systems, spanning semantic consistency, cross-modal synchronization, and overall video quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
language agent, video generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WbW6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents Pixelle-Video, an automated video-production framework that combines a stateful storyboard, narration-based temporal coordination, and persistent visual memory. Experiments on three benchmarks demonstrate improvements in visual consistency, cross-modal synchronization, and video quality.

Critical strengths

The proposed video generation agent has practical value and demonstrates strong performance.

Critical weaknesses
  1. The literature review needs improvement. It mainly lists state-of-the-art approaches without clearly explaining how Pixelle-Video differs from them or what constitutes its novelty. The framework appears to be an engineering integration of LLMs, VLMs, and visual generation tools, making it difficult to identify the key technical contributions. The authors should clarify the specific limitations of existing video-production systems that motivate their designs and explain how each component addresses these limitations, rather than relying on broad statements about the challenges of video generation.

  2. Important implementation details are missing. For example, the temporal coordination mechanism does not clearly explain how narration duration guides video generation or how actions within each shot are aligned with the events described in the narration. Matching durations alone does not guarantee event-level synchronization. Moreover, the cited MM-StoryAgent already uses narration duration to organize multimodal composition, so the authors should clarify what additional coordination their mechanism provides and how it is implemented. Additionally, the used AI tools ands all the parameters should stated for reproducibility.

  3. The experimental coverage and statistical analysis are limited. Evaluation on only 35 prompts makes it difficult to assess the robustness of the improvements or their generalization to longer videos.

Questions for authors
  1. How is narration duration incorporated into video generation, and how are duration mismatches handled? How does this mechanism differ substantively from adjusting video duration to match the narration after generation?

  2. In the supplementary material, Pixelle-Video produces portrait-format videos with fixed black backgrounds, titles, and subtitles, while the compared methods primarily produce landscape-format videos. Were these final videos used for automatic evaluation? Were the added regions removed and preprocessing standardized across methods? If not, does the ranking remain unchanged when evaluation is restricted to the actual generated visual content under a common preprocessing protocol?

  3. When visual memory is removed, are inspection and regeneration also disabled? What timing strategy replaces temporal coordination in the corresponding ablation?

  4. Since the framework relies on multiple AI models and tools, the authors should compare its computational efficiency with existing systems, including end-to-end generation time, peak GPU memory usage, and computational cost where measurable.

3 comments
AuthorsOpenReview ↗

Response to Reviewer WbW6 (Part 1 of 2)

We sincerely appreciate your constructive feedback. Below we address your concerns point by point, and we have accordingly updated the manuscript.

W1. Motivation and Relation to Prior Work.

Thank you for the detailed suggestion. We have revised Section 2 to connect the limitations of prior pipelines to our design and clarify the differences.

Our work targets two coordination limitations in existing video-production pipelines:

  • Individually coherent shots can exhibit subject and setting drift. Our Persistent Visual Memory addresses this through reusable references, global shot binding, and VLM-guided failure diagnosis and targeted revision.
  • Matching total video and narration durations does not ensure shot-level correspondence. MoneyPrinterTurbo loops footage to cover narration, while OpenMontage fits narration to a planned visual timeline. Cross-Modal Temporal Coordination instead uses each shot's measured narration duration to directly set the requested video generation length.

Both mechanisms coordinate visual and temporal information through the shared Stateful Storyboard throughout production.

W2 & Q1. Implementation of Cross-Modal Temporal Coordination.

Narration duration directly determines the requested clip length before video generation. Pixelle-Video first synthesizes each shot's narration and stores its measured duration in the Stateful Storyboard. The video generator receives the shot content, the visual anchor from Persistent Visual Memory, and this duration as the requested clip length. Any residual mismatch is handled per shot by trimming excess video or extending the final frame.

Unlike MM-StoryAgent, which uses narration duration to set image-clip playback length during composition, Pixelle-Video incorporates it into the video generation request itself. Thus, narration duration guides generation rather than serving solely as a target for post-generation adjustment.

We have clarified the model configurations and parameters in Section 4.1.

W3-1. Statistical Robustness.

Robustness. We repeated end-to-end generation three times for each method on the same 35 prompts and report mean ± standard deviation across runs. Pixelle-Video achieves the highest mean scores across all nine metrics, with limited variation across runs, supporting its robustness.

MethodMotion Smooth.Background Consist.Subject Consist.Style & Scene Self-Consist.Individual ActionCharacter InteractionScene ScoreText–Video Consist.Video–Audio Sync.
MoneyPrinterTurbo0.9816 ± 0.0040.7511 ± 0.0200.5377 ± 0.0160.4312 ± 0.0351.5415 ± 0.1382.2402 ± 0.0112.6236 ± 0.1740.3968 ± 0.0440.5645 ± 0.012
OpenMontage0.9894 ± 0.0010.8449 ± 0.0130.7135 ± 0.0190.5466 ± 0.0252.7924 ± 0.1323.0326 ± 0.2283.2414 ± 0.1170.7521 ± 0.0320.5410 ± 0.008
ShortGPT0.9837 ± 0.0010.7698 ± 0.0170.5581 ± 0.0320.4919 ± 0.0132.1176 ± 0.1372.4672 ± 0.0932.8459 ± 0.1200.6315 ± 0.0700.5656 ± 0.009
Anim-Director0.9810 ± 0.0010.8411 ± 0.0180.7569 ± 0.0420.7669 ± 0.0072.7066 ± 0.0612.7763 ± 0.0293.1039 ± 0.0650.7497 ± 0.0740.4727 ± 0.068
ScriptAgent0.9863 ± 0.0010.8301 ± 0.0070.6913 ± 0.0120.7803 ± 0.0192.4273 ± 0.0982.5458 ± 0.0743.0535 ± 0.0850.6636 ± 0.0320.5502 ± 0.013
Pixelle-Video0.9902 ± 0.0000.9036 ± 0.0180.8314 ± 0.0080.8589 ± 0.0363.3808 ± 0.1013.3967 ± 0.1503.5303 ± 0.1280.8366 ± 0.0230.5840 ± 0.012

Evaluations on more tasks. We further expanded evaluation to all 75 DirectorBench instructions, using the same experimental settings and nine metrics. Pixelle-Video ranks first across all nine metrics, consistent with our main findings.

MethodMotion Smooth.Background Consist.Subject Consist.Style & Scene Self-Consist.Individual ActionCharacter InteractionScene ScoreText–Video Consist.Video–Audio Sync.
MoneyPrinterTurbo0.98110.75150.55460.42251.45602.24202.56470.39520.5510
OpenMontage0.98900.84120.72450.55102.84173.15733.23470.71720.5390
ShortGPT0.98300.76770.56290.47712.03862.39952.77750.70880.5600
Anim-Director0.98090.85080.77420.76892.64012.75863.03170.81300.4352
ScriptAgent0.98530.82220.67830.76112.53632.63113.04340.70080.5405
Pixelle-Video0.99000.89430.82360.88483.44393.33643.39090.82740.5714
AuthorsOpenReview ↗

Response to Reviewer WbW6 (Part 2 of 2)

W3-2. Generalization to Longer Videos.

We further extended the target duration from 1 to 3 minutes. Background and Subject Consistency decline by 15.3% and 42.4% for ShortGPT, and 10.6% and 26.0% for OpenMontage, respectively. In contrast, Pixelle-Video maintains high scores on both metrics.

MethodOutput DurationMotion Smooth.Background Consist.Subject Consist.Style & Scene Self-Consist.Individual ActionCharacter InteractionScene Score
Pixelle-Video1 min0.99060.89270.83890.81823.43403.56793.5571
Pixelle-Video3 min0.99080.90420.83230.82043.31113.46673.4857
MoneyPrinterTurbo1 min0.97800.73080.52250.46971.70122.25002.4871
MoneyPrinterTurbo3 min0.98420.69830.39850.39691.41482.34292.5429
OpenMontage1 min0.98860.85980.72420.56862.64242.76903.1276
OpenMontage3 min0.99060.76890.53590.47542.61312.79763.0476
ShortGPT1 min0.98390.78740.58720.49622.03882.42832.7754
ShortGPT3 min0.98830.66720.33830.50331.91922.17772.5725
Anim-Director1 min0.98190.85170.78760.77272.75822.76073.1571
Anim-Director3 min0.98260.77440.60610.71892.62692.82543.1071
ScriptAgent1 min0.98590.83460.70040.78122.39902.50642.9743
ScriptAgent3 min0.98710.81070.63140.72812.38312.58732.6627

Q2. Fairness of Automatic Evaluation Across Video Layouts.

Thank you for raising this concern. The black backgrounds, titles, and subtitles in our supplementary videos are added solely for presentation. For automatic evaluation, we evaluate only the generated visual content across all methods, so presentation layouts do not affect the reported scores or rankings.

Q3. Settings of the Ablation Variants.

The original w/o Persistent Visual Memory disables VLM inspection and regeneration, removes cross-shot references, and switches from I2V to T2V. To isolate these effects, we add two variants: w/o VLM Inspector and Regen., which retains visual references, and w/o Cross-shot Ref., which retains I2V generation with independently generated first-frame anchors.

The full configuration achieves the best results across all nine metrics. Removing inspection and regeneration reduces Character Interaction, while removing cross-shot references decreases Subject Consistency and Style & Scene Self-Consistency. These results demonstrate the complementary benefits of diagnostic refinement and persistent visual references.

MethodMotion Smooth.Background Consist.Subject Consist.Style & Scene Self-Consist.Individual ActionCharacter InteractionScene ScoreText–Video Consist.Video–Audio Sync.
Pixelle-Video0.99060.89270.83890.81823.43403.56793.55710.81970.5939
w/o VLM Inspector and Regen.0.99040.88710.83260.81382.95112.93923.11370.79650.5898
w/o Cross-shot Ref.0.98950.84570.61900.55113.36153.23933.42160.80480.5893
w/o Persistent Visual Memory0.98950.83370.57240.41732.90923.01073.10000.78040.5816

For w/o Cross-Modal Temporal Coordination, we adopt MoneyPrinterTurbo's composition-time matching: generate 5-second clips without narration-derived timing, concatenate visuals and narrations separately and loop or truncate visuals to match the narration.

Q4. Computational Cost.

For efficiency evaluation, we targeted one-minute outputs using Wan3.0 at 720p (USD 0.090/s) and Wan2.7-image (USD 0.030/image), including token costs for DeepSeek-v4-Flash and Qwen3.7-Plus. Pixelle-Video costs USD 0.0964 per video second, averaging 25.6 model calls per video.

Some baselines cost less per output second than the video generation model's rate because MoneyPrinterTurbo, ShortGPT, and ScriptAgent extend clips by holding their final frames to match narration durations without generating additional footage. These simpler workflows achieve weaker visual consistency and content alignment (Table 1 and case studies). Pixelle-Video invests additional computation in persistent visual references and VLM-guided refinement, averaging 0.63 regenerations per image (5.39 per video), to achieve stronger performance.

MethodAPI Cost/s (USD)Model Calls/VideoRuntime (min)/Video
Pixelle-Video0.096425.619.51
MoneyPrinterTurbo0.039410.914.32
OpenMontage0.09005.019.21
ShortGPT0.085414.419.51
Anim-Director0.097729.925.67
ScriptAgent0.089213.129.23
Reviewer_WbW6OpenReview ↗

Reply to authors

Thanks for your response. I have no questions.

Reviewer_qUT7OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Pixelle-Video is an automated agentic pipeline that turns a topic, a script, or user assets into a finished multi-shot video with narration. A “Stateful Storyboard” stores per-shot text, audio, timing, visual anchors, and outputs. Persistent Visual Memory builds multi-view reference sheets for characters and panoramic references for settings, binds them to shots, and uses VLM inspection with up to three regeneration rounds. Cross-Modal Temporal Coordination uses the measured TTS duration of each shot as its timing for visual generation and composition. On 35 DirectorBench prompts, with all systems using the same backbone stack (qwen3.7-plus, wan2.7-image, wan3.0-video, Edge TTS), it reports the best scores on nine VBench, ViStoryBench, and DirectorBench metrics, a 70.3% to 81.2% pairwise human preference, and component ablations.

Critical strengths
  1. Replacing every baseline’s LLM, image, video, and TTS modules with one shared stack is a good control. It focuses the comparison on pipeline design rather than model choice.
  2. The system is practical and complete, and the human preference results are consistent across all five baselines.
Critical weaknesses
  1. The scientific contribution is thin. Using narration duration as the clock is close to what MoneyPrinterTurbo already does at composition time, and reference sheets with VLM-checked regeneration are common in agentic video pipelines. Section 3 describes the method mostly through abstract operators (Φ_extract, Φ_bind, Φ_shot) and gives little detail on the actual mechanisms.
  2. The evidence is limited. There are only 35 prompts, with no variance or significance testing. The headline sync gain is small (0.5755 to 0.5939). Up to three regeneration rounds give Pixelle-Video extra compute that the baselines do not get, and cost and runtime are not reported.
Questions for authors
  1. What are the average number of generation calls, regeneration rounds, and wall-clock cost per video compared with the baselines? How do the baselines do with an equal budget, for example best-of-N with the same VLM inspector?
  2. Concretely, how does Cross-Modal Temporal Coordination differ from duration matching at composition time? Does the clock change the video generation request itself (length, motion), or only trimming and padding?
  3. How many annotators and videos were in the human study, and what was the inter-annotator agreement?
4 comments
AuthorsOpenReview ↗

Response to Reviewer qUT7 (Part 1 of 2)

We sincerely appreciate your constructive feedback. Below we respond to your concerns point by point, and we have accordingly revised the manuscript.

W1. Main Contributions and Differences from Prior Pipelines.

Difference from MoneyPrinterTurbo. MoneyPrinterTurbo simply uses the total narration duration to assemble, repeat, and truncate clips during composition. In contrast, Pixelle-Video uses each shot's measured narration duration to guide video generation, followed by per-shot duration reconciliation. Pixelle-Video achieves higher Text–Video Consistency (0.8197 vs. 0.3536) and Video–Audio Synchronization (0.5939 vs. 0.5694) than MoneyPrinterTurbo.

Difference from Prior Reference-Based Pipelines. We agree that pipelines such as Anim-Director use reference images and VLM checks. However, unlike Anim-Director's VLM-based selection of the best image from multiple candidates, Pixelle-Video combines VLM inspection, failure diagnosis, targeted prompt revision, and targeted regeneration. It also globally binds visual references to relevant shots and maintains visual states and narration-driven timing constraints through the Stateful Storyboard.

We have expanded Section 3 with descriptions of these mechanisms to complement the abstract operators.

W2-1. Generation Stability and Evaluation Scale.

We repeated end-to-end generation three times for each method on the same 35 prompts and report mean ± standard deviation across runs. Pixelle-Video achieves the highest mean scores across all nine metrics, with limited variation across runs, supporting its robustness. Meanwhile, Pixelle-Video achieves higher Text–Video Consistency than the strongest baseline (0.8366 vs. 0.7521), while maintaining the highest Video–Audio Synchronization.

MethodMotion Smooth.Background Consist.Subject Consist.Style & Scene Self-Consist.Individual ActionCharacter InteractionScene ScoreText–Video Consist.Video–Audio Sync.
MoneyPrinterTurbo0.9816 ± 0.0040.7511 ± 0.0200.5377 ± 0.0160.4312 ± 0.0351.5415 ± 0.1382.2402 ± 0.0112.6236 ± 0.1740.3968 ± 0.0440.5645 ± 0.012
OpenMontage0.9894 ± 0.0010.8449 ± 0.0130.7135 ± 0.0190.5466 ± 0.0252.7924 ± 0.1323.0326 ± 0.2283.2414 ± 0.1170.7521 ± 0.0320.5410 ± 0.008
ShortGPT0.9837 ± 0.0010.7698 ± 0.0170.5581 ± 0.0320.4919 ± 0.0132.1176 ± 0.1372.4672 ± 0.0932.8459 ± 0.1200.6315 ± 0.0700.5656 ± 0.009
Anim-Director0.9810 ± 0.0010.8411 ± 0.0180.7569 ± 0.0420.7669 ± 0.0072.7066 ± 0.0612.7763 ± 0.0293.1039 ± 0.0650.7497 ± 0.0740.4727 ± 0.068
ScriptAgent0.9863 ± 0.0010.8301 ± 0.0070.6913 ± 0.0120.7803 ± 0.0192.4273 ± 0.0982.5458 ± 0.0743.0535 ± 0.0850.6636 ± 0.0320.5502 ± 0.013
Pixelle-Video0.9902 ± 0.0000.9036 ± 0.0180.8314 ± 0.0080.8589 ± 0.0363.3808 ± 0.1013.3967 ± 0.1503.5303 ± 0.1280.8366 ± 0.0230.5840 ± 0.012

We further expanded evaluation to all 75 DirectorBench instructions, using the same experimental settings and nine metrics. Pixelle-Video ranks first across all nine metrics, consistent with our main findings.

MethodMotion Smooth.Background Consist.Subject Consist.Style & Scene Self-Consist.Individual ActionCharacter InteractionScene ScoreText–Video Consist.Video–Audio Sync.
MoneyPrinterTurbo0.98110.75150.55460.42251.45602.24202.56470.39520.5510
OpenMontage0.98900.84120.72450.55102.84173.15733.23470.71720.5390
ShortGPT0.98300.76770.56290.47712.03862.39952.77750.70880.5600
Anim-Director0.98090.85080.77420.76892.64012.75863.03170.81300.4352
ScriptAgent0.98530.82220.67830.76112.53632.63113.04340.70080.5405
Pixelle-Video0.99000.89430.82360.88483.44393.33643.39090.82740.5714

Thanks for your rebuttal. I raised my review rating by one level.

AuthorsOpenReview ↗

Response to Reviewer qUT7 (Part 2 of 2)

Q1 & W2-2. Computational Cost and Equal-Budget Comparison

For efficiency evaluation, we targeted one-minute outputs using Wan3.0 at 720p (\0.090/s)andWan2.7−image(0.090/s) and Wan2.7-image (\\0.030/image), including token costs for DeepSeek-v4-Flash and Qwen3.7-Plus. Pixelle-Video costs \$0.0964 per video second, averaging 25.6 model calls per video.

Some baselines cost less per output second than the video generation model's rate because MoneyPrinterTurbo, ShortGPT, and ScriptAgent extend clips by holding their final frames to match narration durations without generating additional footage. These simpler workflows achieve weaker visual consistency and content alignment (Table 1 and case studies). Pixelle-Video invests additional computation in persistent visual references and VLM-guided refinement, averaging 0.63 regenerations per image (5.39 per video), to achieve stronger performance.

MethodAPI Cost/s (USD)Model Calls/VideoRuntime (min)/Video
Pixelle-Video0.096425.619.51
MoneyPrinterTurbo0.039410.914.32
OpenMontage0.09005.019.21
ShortGPT0.085414.419.51
Anim-Director0.097729.925.67
ScriptAgent0.089213.129.23

Equal-budget best-of-N comparison. Anim-Director already employs best-of-N selection (selecting the best among four image candidates). Pixelle-Video achieves better performance (Table 1) at a slightly lower cost (\0.0964vs.0.0964 vs. \\0.0977/s). For other baselines, the same best-of-N strategy is not directly applicable, as their original workflows do not incorporate VLM-based verification of reference images.

Q2. How does Cross-Modal Temporal Coordination differ from duration matching at composition time? Does the clock change the video generation request itself?

Yes—the clock changes the video generation request itself.

  • MoneyPrinterTurbo loops footage to cover narration during composition, potentially repeating visuals unrelated to the current narration segment.
  • OpenMontage generates clips with planned durations and fits narration to the assembled timeline, which may misalign narration segments with shot boundaries.
  • Pixelle-Video coordinates narration and visuals before generation, using each shot's measured narration duration as the requested clip length to support shot-level alignment.

Q3. Human Evaluation Protocol.

As detailed in Appendix C, our human study involved 15 computer science Ph.D. annotators. The pairwise preference study included 93 video pairs, each independently assessed by three annotators, yielding 279 preference judgments. The perceptual scoring study evaluated 72 unique videos, each independently rated by three annotators across seven dimensions, yielding 216 video-level assessments. Table 5 reports inter-annotator correlations across these dimensions, with Pearson's r ranging from 0.742 to 0.889, Spearman's ρ from 0.744 to 0.874, and Kendall's τ from 0.681 to 0.815.

AuthorsOpenReview ↗

Thanks for your prompt feedback and support!

We sincerely appreciate your recognition of our rebuttal efforts and the increased score. Thank you once again for your time in reviewing our paper!

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

The paper proposes Pixelle-Video an agentic pipeline for longer form video generation. Mainly proposes to have Visual items generated prior to video generation and used accross all shots generated (called the Persistent Visual Memory module in the paper) and likewise to to similar for the audio called the (cross modal temporal coordination module in the paper), all governed by an agent and a storyboard (see figure 2)

Critical strengths

the paper is clear easy to follow builds on prior works

Critical weaknesses

figure 1 already has a problem the bowl of cat food dissapears from the 5s to 12s mark. was this the best case? or is there a better teaser?

the paper does not make it clear that it is an agentic video generation pipeline, where the agent is tasked with ensuring consistency accross the pipeline

the paper does not make it clear the flow of the pipeline, like the whole flow from user's input promt till the final composed clip (detailing what does what and who does what), fiigure 2 attemps to do so but barely reproduceable from that schema alone.

the paper does not include some recent works in the comparison like Google's CANVAS (https://research.google/blog/coherent-long-form-video-generation/). We will also like to see a comparison if possible with Huawei's Workswarm Video generation pipeline that looks similar (https://github.com/openJiuwen-ai/jiuwenswarm.git)

Not much on the theory side too

Questions for authors

Please see weaknesses above

1 comment
AuthorsOpenReview ↗

Response to Reviewer Vdqa

We are grateful for your constructive feedback and your positive recognition of our work. In below, we address your concerns point by point, and update the manuscript accordingly.

W1: The Disappearing Food Bowl in Figure 1

We would like to clarify that the frames at 5 s and 12 s show two distinct storyboard shots. In the first shot, the cat asks for food, and the bowl is a shot-specific prop. In the second, the video cuts to the cat chasing a red dot. The bowl's absence therefore reflects the story progression and change in action. Our Persistent Visual Memory preserves recurring identities and settings while allowing shot-specific content to change.

W2. Clarification of the Agentic Workflow

Thanks for your detailed suggestion. Pixelle-Video is an agentic video production framework organized around a shared Stateful Storyboard, which tracks the evolving production state across four stages:

  • StoryBoard Planning: An LLM transforms the user request into an ordered storyboard with shot-level narrations.
  • Persistent Visual Memory coordinates visuals across shots. The agent extracts recurring subjects and settings, constructs reusable visual references, and binds them to relevant shots. An image-generation model produces shot-level visual anchors, while a VLM inspects the generated references and anchors to guide prompt revision and targeted regeneration.
  • Cross-Modal Temporal Coordination coordinates text, audio and video within each shot. The system synthesizes shot-level narrations using TTS and records their measured durations in the Stateful Storyboard. The video generator then uses the corresponding visual anchors and takes the measured duration as the requested clip length.
  • Final Composition: The compositor assembles the shots in storyboard order into the final video.

We have revised Figure 2 and Section 3 to clarify the agents' responsibilities and the complete workflow.

W3. Comparison with Recent Works

We have added Google's CANVAS to Related Work in the revised manuscript. CANVAS focuses on visual continuity across storyboard frames, while Pixelle-Video jointly coordinates cross-shot visual consistency and cross-modal timing throughout end-to-end production.

Thank you also for suggesting WorkSwarm. We have reproduced Huawei's WorkSwarm Design workflow as an additional baseline and included the comparison in the updated Table 1 and 2. Under the same experiment settings, Pixelle-Video outperforms WorkSwarm across all 9 metrics, supporting the effectiveness of our coordination framework.

MethodMotion Smooth.Background Consist.Subject Consist.Style & Scene Self-Consist.Individual ActionCharacter InteractionScene ScoreText–Video Consist.Video–Audio Sync.
MoneyPrinterTurbo0.97800.73080.52250.46971.70122.25002.48710.35360.5694
OpenMontage0.98860.85980.72420.56862.64242.76903.12760.75930.5500
ShortGPT0.98390.78740.58720.49622.03882.42832.77540.57390.5755
Anim-Director0.98190.85170.78760.77272.75822.76073.15710.66820.5510
ScriptAgent0.98590.83460.70040.78122.39902.50642.97430.64430.5643
WorkSwarm0.98340.85710.71110.75252.57872.55563.02780.77620.5204
Pixelle-Video0.99060.89270.83890.81823.43403.56793.55710.81970.5939

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Summary of Revisions

We sincerely thank you for your time and constructive feedback. We appreciate the reviewers' recognition that:

  • Pixelle-Video is pratical and complete, demonstrating strong performance. [qUT7, WbW6]
  • The shared backbone stack provides a good control for comparing pipeline designs. [qUT7]
  • The paper is clear and easy to follow, and the human preference results consistently favor our system. [Vdqa, qUT7]

Following your suggestions, we have made the following revisions in the updated manuscript (highlighted in red):

  • Revised Related Work to clarify the differences from prior systems. Section 2. [Vdqa, qUT7, WbW6]
  • Clarified the agentic workflow and method descriptions. Section 3 and Figure 2. [Vdqa, qUT7, WbW6]
  • Added WorkSwarm as an additional baseline, evaluated across all nine automatic metrics. Sections 4.1–4.2, Tables 1–2. [Vdqa]
  • Expanded the evaluation of generation stability and scalability, reporting three-run generation results and evaluation on all 75 DirectorBench instructions. Appendix D, Tables 6–7. [qUT7, WbW6]
  • Added component-level ablations and clarified their settings. Appendix F, Table 9. [WbW6]
  • Added computational cost measurements, reporting API costs and model calls. Appendix E, Table 8. [qUT7, WbW6]
  • Expanded the human evaluation details. Appendix C. [qUT7]

We sincerely appreciate your constructive suggestions, which help strengthen our paper a lot. We are happy to provide additional clarifications. Thank you for your time!

Sincerely,

The Authors

AuthorsOpenReview ↗

Summary of Main Contributions and Rebuttal Outcomes

Dear AC, SAC, and PC,

We sincerely thank you and the reviewers for your time and constructive feedback. To facilitate your final assessment, we summarize the main contributions of our work and the key outcomes of rebuttal as follows:


Part I: Main Contributions of Our Work

  1. An automated framework for end-to-end video production. We introduce Pixelle-Video, which transforms user requests into complete multi-shot videos through two coordination mechanisms operating on a shared Stateful Storyboard of multimodal artifacts and execution metadata.
  2. Two complementary coordination mechanisms. Persistent Visual Memory preserves cross-shot consistency through reusable visual references and VLM-based inspection with targeted revision and regeneration. Cross-Modal Temporal Coordination aligns text, audio, and video within each shot, using measured narration duration as a shared temporal clock to synchronize video generation and composition.
  3. Strong performance over recent baselines. With all compared systems using a shared backbone stack, Pixelle-Video achieves the best performance across all nine automatic metrics and human preference over all baselines.

Part II: Key Outcomes of Rebuttal

  1. Reviewer qUT7 (Rating: 1 Rejection → 2 Weak Rejection)

    • [Review] Recognized that the shared backbone stack provides “a good control” and that the system is “practical and complete.” Raised concerns about technical novelty, evaluation scale and variance, computational cost, and human evaluation details.
    • [Response] We clarified the differences from composition-time duration matching and prior reference-based pipelines. We also added three-run evaluation and results on all 75 DirectorBench instructions, and reported computational costs and human evaluation details.
    • [Feedback] The reviewer raised the review rating by one level.
  2. Reviewer WbW6 (Rating: 2; confirmed no further questions)

    • [Review] Recognized the framework's “practical value” and “strong performance.” Raised concerns about novelty, implementation details, experimental coverage, longer-video generation, evaluation fairness, ablation settings, and computational efficiency.

    • [Response] We revised Related Work and the method description to clarify the differences from prior pipelines in cross-shot visual consistency and narration-driven temporal coordination. Beyond the expanded robustness and cost evaluations, we provided:

      • Longer-video experiments: Extending the target duration from one to three minutes preserves strong Background Consistency and Subject Consistency.
      • Component-level ablations: Separate variants isolate VLM inspection and regeneration from cross-shot references, supporting the complementary benefits. We also clarified the replacement timing strategy used in the temporal coordination ablation.
      • Evaluation fairness clarification: Automatic evaluation uses only generated visual content, excluding presentation backgrounds, titles, and subtitles.
    • [Feedback] The reviewer stated "I have no questions."

  3. Reviewer Vdqa (Rating: 3, maintained Weak Acceptance)

    • [Review] Described the paper as "clear and easy to follow." Raised questions about the disappearing bowl in Figure 1, the agentic workflow, and comparisons with recent systems, including CANVAS and WorkSwarm.
    • [Response] We clarified the shot transition in Figure 1 and the workflow and agent responsibilities in Section 3 and Figure 2. We added CANVAS to Related Work and WorkSwarm as a baseline under the same settings, with Pixelle-Video outperforming it across all nine automatic metrics.
    • [Discussion status] No follow-up feedback has been posted.

Best,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.