ActAnywhere: Subject-Aware Video Background Generation

Generating video background that tailors to foreground subject motion is an important problem for the movie industry and visual effects community. This task involves synthesizing background that aligns with the motion and appearance of the foreground subject, while also complies with the artist's creative intention. We introduce ActAnywhere, a generative model that automates this process which traditionally requires tedious manual efforts. Our model leverages the power of large-scale video diffusion models, and is specifically tailored for this task. ActAnywhere takes a sequence of foreground subject segmentation as input and an image that describes the desired scene as condition, to produce a coherent video with realistic foreground-background interactions while adhering to the condition frame. We train our model on a large-scale dataset of human-scene interaction videos. Extensive evaluations demonstrate the superior performance of our model, significantly outperforming baselines. Moreover, we show that ActAnywhere generalizes to diverse out-of-distribution samples, including non-human subjects. Please visit our project webpage at https://actanywhere.github.io.

Paper

References (50)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ghM95/10 · confidence 5/52024-07-01

Summary

The task addressed by this paper is, given the the appearance and segmentation mask of a foreground subject in a video, for example a human running, to synthesize video backgrounds that are plausible and realistic, both in content and motion. For example, for a person running, if the ground is wet there should be splashes corresponding to footfalls, and the background scene should move along with the runner's movement. The authors address this problem as a generative video outpainting task. They model it using a latent video diffusion model. The foreground subject appearances (encoded) and masks are concatenated with the latent feature vector, while the conditioning image (what the background should look like) is encoded via CLIP and is provided via cross-attention to the U-net. A variety of qualitative results are included, demonstrating the ability to out-paint plausible scenes, including background motion, effects (splashing water) and even objects that the subject interacts with. Quantitative evaluations include ablations validating the main components of the model, as well as human rater studies comparing the quality of the output to competing methods.

Strengths

Originality & Significance. The paper addresses a minor new variant on the outpainting/inpainting problem, which introduces some new challenges such as picking up on foreground contextual cues and using that to inform the background. Quality & Clarity: The paper is clearly written, seems correct, and demonstrates visually appealing results with a variety of different foreground subjects (person, car, duck, etc.). The supplemental videos help illustrate the model's capabilities well. The empirical results also support the quality of the generative model.

Weaknesses

The main weakness of this paper is very limited novelty, which compromises the claimed contributions of the paper. Video inpainting/outpainting is a fairly well-known challenge. The main novelty here is that the masks are simply inverted: instead of deleting an unwanted person and generating the missing background pixels, here we retain only the desired foreground person and generate the missing pixels. I am not convinced that this difference is substantively novel. Similarly, the modeling approach bears considerable similarity to [13] "Structure and Content-Guided Video Synthesis with Diffusion Models". In particular, compare Fig.2 from the two papers. The primary difference between the two is that this work conditions on foreground appearances and masks, while [13] conditions on estimated depth images. Again, the novelty is minor and unsurprising.

Questions

Have you tried unconditional generation, i.e. not providing any signal as to what the desired background should be? This could be quite interesting, since it would challenge the model to understand exactly what the foreground subject is doing and to invent plausible backgrounds. For example, this child seems to be interacting with an object, what could that object be? This woman is riding something… what fits? Would you like to discuss the appearance of the Shutterstock watermarks in the generated backgrounds, particularly in figure 3? This could be an interesting example of the model using foreground context informing the background generation, as per my previous question.

Rating

5

Confidence

5

Soundness

3

Presentation

4

Contribution

2

Limitations

Yes.

Reviewer gAcX7/10 · confidence 4/52024-07-11

Summary

This paper introduces ActAnywhere, a video diffusion model designed to generate video backgrounds that adapt to the foreground subject's motion. By utilizing a sequence of foreground subject segmentation and a background image, the model produces realistic videos with coherent foreground-background interactions. Experiments on a large-scale dataset demonstrate the model's effectiveness, outperforming existing methods in generating realistic and dynamic backgrounds.

Strengths

S1: The paper introduces a novel problem of automated subject-aware video background generation. S2: The methodology shows improvements in generating coherent videos with realistic subject-background interactions. S4: The contributions are significant, particularly for applications in the movie industry and visual effects. S4: The paper is comprehensive and well-written.

Weaknesses

W1: The paper lacks sufficient comparison with a broader range of existing methods, particularly those leveraging recent advancements in video generation and editing (though they are not for background generation, they can do). W2: It seems that the model relies heavily on the quality of the foreground video segmentation masks.

Questions

Q1: I wonder about the impact of the quality of the foreground segmentation. It seems that the model relies heavily on the quality of the foreground segmentation. Q2: How scalable is your model for generating longer video sequences? Have you tested its performance in generating videos of varying lengths, and what are the results?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

I think there is no potential negative societal impact and the authors have addressed the limitations.

Reviewer Pq5G5/10 · confidence 4/52024-07-13

Summary

This paper studies a new topic: automatic background generation of moving foreground subject. Different from video inpainting/outpainting and other video editing methods, the method in this paper can maintain the consistency of foreground moving subject, and maintain reasonable and realistic interactions, camera motion, lighting and shadows. It can also generalizes to diverse scenarios including non-human objects, gaming and animations, as well as videos with multiple moving subjects. Input a foreground segmentation image sequence and a foreground mask sequence, and under the guidance of a background reference image, generate a motion video of the interaction between the foreground and the background. Specifically, the spatial layer parameters of Denoising 3D Unet are loaded from a pre-trained SD-inpainting model, and the temporal layer(motion module) is loaded from the pre trained model of AnimateDiff V2. The input is a feature map of nine channels, which concatenates the nosie latent, the vae encoder feature of the foreground segmentation map, and the downsampled foreground mask in the channel dimension. The CLIP feature of the background reference map is injected into the Unet through cross attention.Trained on the HiC+ dataset (2.4M videos of human-scene interactions), Given the error of foreground segmentation, random rectangular cut-outs augmentation is applied to the foreground segmentation sequence and foreground mask sequence

Strengths

A new task was studied: automatic background generation of moving foreground, and reasonable and realistic foreground/background interaction, camera motion, lighting and shadows were required. Although it is a new task, the results of similar methods were compared under fair conditions as much as possible, which indeed proved the superiority of the method in this paper.

Weaknesses

The training resolution is low (256x256), and it is unclear whether the method in this paper still has advantages and reliability under high-score conditions? Ablation experiment: The method of using background reference image for guidance does not seem to be fully considered, such as the method similar to IPAdapter? The training details are not clear: is it fine-tuning the spatial layer and the motion module at the same time? Or only fine-tuning the motion module? Or is cotraining still required? This article spends a certain amount of space to explain the difference in results between the method proposed in this paper and the SD-inpainting method, but in terms of the algorithm pipeline, except for using the CLIP features of the background reference image for guidance, the rest is almost exactly the same as the SD-inpainting model, and even the spatial layer of Unet directly loads the pre-trained model of SD-inpainting. The article does not seem to explain the difference in principle between the method proposed in this paper and the SD-inpainting method?

Questions

Please refere weakness

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Yes.

Reviewer Aquh5/10 · confidence 5/52024-07-15

Summary

This paper study to automatically generate video background that tailors to foreground subject motion. It proposes ActAnywhere, a video diffusion model that takes as input a sequence of foreground subject segmentation and an image of a novel background and generates a video of the subject interacting in this background. Both quantitative and qualitative comparisons demonstrate that the model significantly outperforms existing methods, which fail to accomplish the studied task.

Strengths

1. The task is interesting and the result seems to be very competitive. 2. The paper is written clear.

Weaknesses

1. Only encoder condition frame by a clip encoder cannot contains details information, especially for high-resolution input images. Please refer AnimtaeAnyone for human image animation. And please compare such a referencenet attention way. 2. Can you provide some video demos. I suppose there may exists blur boundary results. 3. Since the task seems to be very related to human animation task, could you please show some results on TikToK dataset? Please see AnimateAnyone paper for comparision details.

Questions

See weakness

Rating

5

Confidence

5

Soundness

4

Presentation

3

Contribution

2

Limitations

See weakness

Area Chair Z54u2024-08-09

Dear Reviewers, Thank you very much again for your valuable service to the NeurIPS community. As the authors have provided detailed responses, it would be great if you could check them and see if your concerns have been addressed. Your prompt feedback would provide an opportunity for the authors to offer additional clarifications if needed. Best regards, AC

Reviewer ghM92024-08-09

Thank you for exploring unconditional background generation. I think this is an intriguing variant of the problem to showcase. I like how the model attempts to correct for the overly-tight cropping of the foreground person, by generating the missing shoe. Given the general agreement from other reviewers that this problem is novel and interesting, I'll retract that part of my assessment, and boost the score.

Authorsrebuttal2024-08-09

We are glad to hear that we addressed your concerns. We will include the unconditional generation results in the final version of the paper. Thank you for raising the score! And thanks again for your efforts spent reviewing our paper!

Area Chair Z54u2024-08-12

Has the rebuttal addressed your concerns?

Dear Reviewer Aquh, Thank you again for your time to review this paper. Could you please check if the authors' rebuttal has addressed your concerns at your earliest convenience? The deadline of the discussion period will end in about 24 hours. Thank you! Best regards, AC

Area Chair Z54u2024-08-12

Has the rebuttal addressed your concerns?

Dear Reviewer Pq5G, Thank you again for your time to review this paper. Could you please check if the authors' rebuttal has addressed your concerns at your earliest convenience? The deadline of the discussion period will end in about 24 hours. Thank you! Best regards, AC

Reviewer Pq5G2024-08-12

Thanks authors for providing answers. Most of my concerns was addressed. But I still think the technical contribution is not very significant. I kept my score as 5.

Area Chair Z54u2024-08-12

Has the rebuttal addressed your concerns?

Dear Reviewer gAcX, Thank you again for your time to review this paper. Could you please check if the authors' rebuttal has addressed your concerns at your earliest convenience? The deadline of the discussion period will end in about 24 hours. Thank you! Best regards, AC

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC