GLOBER: Coherent Non-autoregressive Video Generation via GLOBal Guided Video DecodER

Video generation necessitates both global coherence and local realism. This work presents a novel non-autoregressive method GLOBER, which first generates global features to obtain comprehensive global guidance and then synthesizes video frames based on the global features to generate coherent videos. Specifically, we propose a video auto-encoder, where a video encoder encodes videos into global features, and a video decoder, built on a diffusion model, decodes the global features and synthesizes video frames in a non-autoregressive manner. To achieve maximum flexibility, our video decoder perceives temporal information through normalized frame indexes, which enables it to synthesize arbitrary sub video clips with predetermined starting and ending frame indexes. Moreover, a novel adversarial loss is introduced to improve the global coherence and local realism between the synthesized video frames. Finally, we employ a diffusion-based video generator to fit the global features outputted by the video encoder for video generation. Extensive experimental results demonstrate the effectiveness and efficiency of our proposed method, and new state-of-the-art results have been achieved on multiple benchmarks.

Paper

Similar papers

Peer review

Reviewer P49b4/10 · confidence 5/52023-07-07

Summary

This paper proposes a new video generation framework based on extracting the global features of the video and conditional diffusion model to predict frame features, leading to the frame. The paper argues the proposed method outperforms prior video generation methods on various benchmarks, including UCF-101, Taichi-HD, and SkyTimelapse.

Strengths

- Compared with the prior video generation methods, the proposed method considers non-autoregressive approach for generating the video, which can improve the efficiency in inference time. - The proposed method shows better performance compared with prior works.

Weaknesses

- The overall framework is quite complex, including so many notations, and a bit difficult to follow. For instance, Why $I_j$ is put into the video decoder model as well as $I_j$ in Figure 2? Moreover, are KL-VAE and video encode/decoder, discriminator jointly trained or not? What is the intuition of letting the video decoder network as a conditional diffusion model instead of letting simple 2D CNNs? Why do we need to consider "keyframes" for extracting global features from a given video? Does DiT for modeling global features is trained in post-hoc manner after the training of the entire framework? - The paper misses an efficiency comparison with recent latent video diffusion models to improve the efficiency in training and efficiency: e.g., LVDM [He et al., 2023] and PVDM. Compared with these frameworks, what is the advantage and disadvantages of the method? - Typo: Specificcally -> Specifically in L197. --- [He et al., 2023] Latent Video Diffusion Models for High-Fidelity Long Video Generation [Yu et al., 2023] Video Probabilistic Diffusion Models in Projected Latent Space, CVPR 2023

Questions

- I guess the proposed method may show a worse performance if the targeting video length becomes large because the quality of global features might have a limitation and the decoder that synthesizes a frame in frame-index conditioned manner has a limited capacity. What is the (empirical) maximum length for high-quality modeling with this framework?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

The paper adequately addresses the limitations in Conclusion section.

Reviewer JPZt6/10 · confidence 4/52023-07-07

Summary

This work introduces a novel non-autoregressive method, GLOBER, that first generates global features for comprehensive global guidance and then synthesizes video frames based on these global features to produce coherent videos. The authors propose a video auto-encoder to encode videos into global features and a video decoder to decode the global features and synthesize video frames in a non-autoregressive manner. Notably, the video decoder uses normalized frame indexes to perceive temporal information, allowing it to synthesize any video clips with predetermined frame indexes. The authors also introduce a unique adversarial loss to enhance global coherence and local realism of the synthesized video frames. Finally, a diffusion-based video generator is employed to fit the global features produced by the video encoder for video generation. The effectiveness and efficiency of the proposed method are demonstrated through extensive experiments, and it sets new state-of-the-art results on multiple benchmarks.

Strengths

1. The inclusion of Coherence and Realism Adversarial Loss is a novel approach compared to previous diffusion-based architectures. 2. Extensive experiments have been performed on various benchmarks, all demonstrating the significance of GLOBER.

Weaknesses

1. The authors identify VideoFusion as the most closely related work due to its use of non-autoregressive generation. However, there are other public models, such as ModelScope Text-to-Video, that use non-autoregressive generation in the latent space similarly to the authors' work. I suggest that the authors compare their work to these models as well. 2. It appears that GLOBER outperforms VideoFusion in all tasks, which generates videos in the pixel space. This superiority seems to result from the CRA loss proposed by the authors. Therefore, a direct comparison between GLOBER (without CRA loss) and VideoFusion would be intuitive. However, inconsistent results in Table 1 and 3 make this comparison unfeasible. Could the authors explain this inconsistency and provide justifications for GLOBER's superiority over VideoFusion, aside from the CRA loss? 3. My interpretation of Equation 9 suggests it's an estimation of the frame feature. However, this estimation might not be accurate because, like in DDPM (or DDIM), one could perform T steps of reverse denoising to generate images. What is the quality difference between these two types of images? I assume that the frame feature generated by Equation 9 will be of lower quality. 4. In Table 2, why does GLOBER use 50+50 diffusion steps? 5. The optimization objective of Equation 2 in the Video Encoding section appears to be derived from the Variational Autoencoder. Could the authors provide a justification for this design? My understanding is that video encoding trains a dataset-dependent distribution to be sampled as z_t. 6. In Line 127, "Gauss distribution" appears to be misspelled. 7. The authors seem to have overlooked specifying the dimension of C'. My score could be revised upward if my concerns are adequately addressed. **Reference:** [1] ModelScope Text-to-Video Technical Report, arXiv. (Model: **https://modelscope.cn/models/damo/text-to-video-synthesis/summary**).

Questions

Please refer to the weaknesses.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

4 excellent

Limitations

In this work, the authors mentioned limitations and broader impact.

Reviewer xuBL5/10 · confidence 5/52023-07-07

Summary

In this paper, the author studies the text-to-video task and proposes a method called GLOBER. The proposed method first generates a global guidance feature, then the video frames are generated through a diffusion model that takes the frame index as a condition. An adversarial loss is also proposed to improve global coherence and local realism.

Strengths

1. The overall presentation of the proposed method is clear and easy to follow. 2. The author conducts experiments on three widely used datasets (e.g., Sky Time-lapse, TaiChi-HD and UCF-101).

Weaknesses

1. As the author claims their method is capable of generating videos from text. It would be great if the author could compare their methods with SOTA open-sourced methods (DAMO-text2video, VideoCrafters, CogVideo and VideoFactory) on WebVid-10M. 2. The author should also consider comparing their methods with PVDM.

Questions

Please refer to the weakness part.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Please refer to the weakness part.

Reviewer 4poX6/10 · confidence 4/52023-07-08

Summary

The study introduces a unique non-autoregressive approach called GLOBER. This method initially generates global features, offering a thorough global guidance, which then synthesizes video frames using these global features to produce cohesive videos. Furthermore, the study suggests a coherence and realism adversarial loss to improve the quality of the videos.

Strengths

The suggested non-autoregressive technique is simple and effective. The empirical tests and ablation studies conducted are adequate

Weaknesses

For Table 2, the absence of some contemporary methods such as modelscope implies that the assertion regarding inference time and GPU memory may not be as robust as claimed. What strategies are in place to ensure that the distribution of global features produced by DiT during inference aligns with the features acquired by the video encoder during the training phase? I'm intrigued to find out whether this non-autoregressive approach is effective with lengthy videos, for example, those consisting of 128 or 256 frames. If it's not, the benefits of this method could be significantly reduced.

Questions

Please refer weakness

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have discussed limitations.

Reviewer P49b2023-08-17

Response

Thanks for the response. It helps me a lot to understand the details of the method. However, at the current status, it is difficult for me to recommend acceptance. Specifically, I still have a doubt about the capability of your video autoencoder to encode possibly long videos through relatively low-dimensional latents (4,096). The response states that the proposed method can compress video (128, 256, 256, 3) to 4,096, but I really don't think this can achieve high-quality reconstructions, especially on complex datasets (e.g., UCF-101, Kinetics, and so on). In the authors' response to my review, the authors do not provide any reconstruction/generation results on such a long video with complex datasets. Without this result, it is hard for me to believe whether this method indeed scales up well to large-scale and complex datasets. In addition, the authors state "PVDM and LVDM obtain better performance than GLOBER on short-video generation on simple domain datasets like SkyTimelapse and long-video generation on multi-motion datasets like UCF-101.", I think short-video generation on simple domain dataset and long video generation on multi-motion datasets has no similarity and thus think the analysis provided in the response is not that insightful. Considering all of these aspects, I will retain my score.

Authorsrebuttal2023-08-17

Response to Reviewer P49b

Thank you for your response! In response to your two questions, we have the following explanations: ### R4.5: Generation results on long videos with complex datasets **We have provided the qualitative and quantitative results for 128-frame video genration on the UCF-101 dataset in the A.2 section (L9-18) in the appendix.** As reported in the Table 1 of the appendix (which is copied in the following), our GLOBER outperforms previous methods StyleGAN-V (CVPR2022) and VIDM (AAAI2023) by a large margin. We also visualize the generated long videos on the UCF-101 and SkyTimelapse datasts in the link provided in the L21 of the appendix. Moreover, for short video generation, our method can obtain comparable performance with current SOTA models on the much more complex dataset WebVid-10M as reported in the Table Q1.4 of the common question 1. Table 1: Quantitative Results of FVD comparison on the SkyTimelapse and UCF-101 datasets for 128-frame long video generation. | Method | UCF-101 | Sky Timelapse | |:-|:-:|:-:| | MoCoGAN [CVPR18] | 3679.0 | 575.9 | | +StyleGAN2 backbone | 2311.3 | 272.8 | | MoCoGAN-HD [ICLR21] | 2606.5 | 878.1 | | DIGAN [ICLR22] | 2293.7 | 196.7 | | StyleGAN-V [CVPR22] | 1773.4 | 197.0 | | VIDM [AAAI23] | 1531.9 | 140.9 | | GLOBER (ours) | **1177.4** | **125.5**| ### R4.6: Explanation of model performance compared to PVDM and LVDM **Performance on the multi-motion dataset** Since PVDM and LVDM auto-encode a fixed number of video frames while our GLOBER pursue flexible decoding and use much less number of latent elements, PVDM and LVDM can put stronger constraint on the consistency of decoded video frames than our GLOBER when video motion is dramatic, e.g. long videos in a multi-motion dataset. Thus they obtain better performance on 128-frame video generation on UCF-101. However, for short video generation on the multi-motion dataset, our GLOBER can obtain comparable video consistency and much better video realism since PVDM and LVDM requires much more number of elements to represent a video clip than GLOBER (Table. Q1.3 in the common question 1), making their video generators difficult to fit the distribution of video latents, and our GLOBER employs the powerful diffusion model as the video decoder. **Performance on the simple domain dataset** The key reason of PVDM outperforming our GLOBER for short-video generation on simple domain dataset like SkyTimelapse lies in that since videos in such dataset contain mostly simple and statistic scenes (city or nature scenes), the determinstic video decoder in PVDM may obtain better video reconstruction than our diffusion video decoder. Notably, despite that LVDM also employs a determinstic video decoder, it performs inferior to our GLOBER (95.2 vs 78.1 FVD) since it requires three times the number of latent features of this method, thus being too difficult to fit the latent distribution well. When the length of video increases, the video consistency of our GLOBER is still comparable with them due to videos containing small motions (clouds floating and other variation of sky), while the video realism of PVDM and LVDM drops due to error accumulation (both of them employ the auto-regressive generation strategy). Thus our GLOBER can obtain a better score against PVDM and LVDM in such case.

Authorsrebuttal2023-08-21

For Reviewer P49b

Dear Reviewer P49b, There is not so much time left for the discussion stage, if you still have questions about our work, please let us know and we will reply as soon as possible, thanks for your effort and time! Best wishes, Author

Reviewer P49b2023-08-21

Response

Thanks for the detailed follow-up. For me, the limitation of GLOBER on long-video, multi-motion setup is quite critical issue. Specifically, it seems the scalability of the method toward real-world complex and long videos is limited, in contrast to other methods; so forcing the latent dimension to be small (4,096) is not persuable. If scalability can be ensured with low-dimensional latents for video encoding, it is definitely a good contribution and a good method, but if not, I think the strength of the method is limited. Therefore it is a bit difficult to recommend acceptance. I think in the future, the authors can provide long vid. experiments using multi-motion datasets with higher-dimensional latents (like 16,384) and check whether such change can outperform previous arts such as PVDM to improve the soundness of the method, not just stating the method shows limited performance due to the usage of low-dimensional latents.

Authorsrebuttal2023-08-21

Response to Reviewer P49b

Thanks for your response! ### R4.7: Firstly, we did not force the number of latent elements to be small (4096). The focus is that dispite of small number of latent elements, our proposed method can still obtain SOTA results on multiple multi-motion benchmarks for short video generation, i.e. UCF-101 and Webvid-10M, and outperform most contemporary methods like VIDM for long video generation. Secondly, since we employ much less number of latent elements than PVDM (4096 vs 65536) to represent a long video, we are afraid that harse comparison with PVDM is unfair for our proposed method, especially when we have outperform PVDM for 16-frame video genration on the multi-motion dataset UCF-101 by ~100 FVD and obtain other advantages like flexible decoding and efficiency. We will add the experiment you mentioned on multi-motion datasets with higher-dimensional latents in our paper later. Thanks again for your effort and time to serve as a reviewer! Best regrads, Author

Reviewer xuBL2023-08-18

Thanks for the rebuttal. I think most of my concerns have been well addressed. I will raise my score.

Authorsrebuttal2023-08-18

Response to Reviewer xuBL

Thank you for your recognition of our job! We will incorporate the above discussions when we revise our paper.

Area Chair 316v2023-08-18

Discussion period

Dear reviewers 4poX and JPZt You haven't provided your comments on the authors' response. Please do so before the deadline. Does the response address any of your concerns? The AC thanks to reviewers xuBL and P49b for doing so! AC

Reviewer JPZt2023-08-19

Post-rebuttal discussions

Dear the authors of Paper 6343, Many thanks for your detailed reply. It solves most of my concern. I have one follow-up question about Question3.5. Is the variance prediction in Table R3.5.1 indicates the sampling process in Eq.1? Btw, is the auto-encoding stage in [1] trained in an end-to-end manner with the denoising UNet using KL loss, like what GLOBER does? I look forward to your reply. Kind regards, Reviewer JPZt

Authorsrebuttal2023-08-19

Response to Reviewer JPZt

Thank you for your response! For the first question, **yes**, the variance prediction in Table R3.5.1 is the sampling process in Eq. 1. For the second question, **no**, the auto-encoding stage in [1] utilized simple CNNs as its encoder and decoder. The overall structure of the auto-encoder in [1] (i.e. KL-VAE) is similar to the traditional VQ-VAE [2] or VQ-GAN [3] except for that KL-VAE represents images with continuous latent features while VQ-VAE and VQ-GAN represents images with discrete tokens through vector-quantization, which are difinitely different from our model. [2] Zero-Shot Text-to-Image Generation. [3] Taming Transformers for High-Resolution Image Synthesize.

Reviewer JPZt2023-08-19

Post-rebuttal discussions (2)

Dear the authors of Paper 6343, Thank you for your thorough response. Your rebuttal has addressed my concerns, and as promised, I have decided to raise the score. Best regards, Reviewer JPZt

Reviewer 4poX2023-08-19

Thanks for your responses. My concerns have been well addressed. But as a video generation paper, it is necessary to provide the video results, in addition to quantitative results. Could you please provide an anonymous links containing video results? As indicated in PC emails, such a link is allowed and encouraged for video generation papers.

Authorsrebuttal2023-08-19

Response to Reviewer 4poX

We have provide an anonymous link in Section A.3 (L20-21) of the appendix, and that is https://anonymouss765.github.io/GLOBER. Please refer to this link for both short and long video samples! Thank you for your response!

Authorsrebuttal2023-08-19

Response to Reviewer 4poX

Thank you for recognization! To facilitate following researches, we will open source our codes and models once the paper is accepted.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC