Any-to-Any Generation via Composable Diffusion

We present Composable Diffusion (CoDi), a novel generative model capable of generating any combination of output modalities, such as language, image, video, or audio, from any combination of input modalities. Unlike existing generative AI systems, CoDi can generate multiple modalities in parallel and its input is not limited to a subset of modalities like text or image. Despite the absence of training datasets for many combinations of modalities, we propose to align modalities in both the input and output space. This allows CoDi to freely condition on any input combination and generate any group of modalities, even if they are not present in the training data. CoDi employs a novel composable generation strategy which involves building a shared multimodal space by bridging alignment in the diffusion process, enabling the synchronized generation of intertwined modalities, such as temporally aligned video and audio. Highly customizable and flexible, CoDi achieves strong joint-modality generation quality, and outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis. The project page with demonstrations and code is at https://codi-gen.github.io

Paper

References (59)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer hGLR7/10 · confidence 3/52023-06-27

Summary

They present Composable Diffusion (CoDi), a novel generative model capable of generating any combination of output modalities, such as language, image, video, or audio, from any combination of input modalities. Unlike existing generative AI systems, CoDi can generate multiple modalities in parallel and its input is not limited to a subset of modalities like text or image. Despite the absence of training datasets for many combinations of modalities, they also propose to align modalities in both the input and output space.

Strengths

1 One model that takes any combination of modalities as input or output is novel and promising. 2 As the lack of training data, the alignment of different modalities is very difficult. The proposed method for the alignment is very interesting.

Weaknesses

1 The simple weighted interpolation of different representations is not so convincing. Why does this method work?

Questions

see above

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

not addressed

Reviewer mZPw5/10 · confidence 5/52023-07-04

Summary

This paper presents a method that can generate any combination of output modalities, including language, audio, image, or video, from any combination of input modalities. The idea here is to align four modalities in a shared feature space first, and then learn to generate one or more modalities based on the shared feature space. This design enables many combinations of modalities despite the lack of training datasets. Since the feature space is shared, it also flexible to extend to other modalities.

Strengths

* The idea of any-to-any generation is interesting, and it enables many different tasks in one model. * The framework is flexible and customizable to many other potential modalities, such as semantic maps, heat map, depth map and so on. * The performance of the proposed method achieves comparable or better results than previous SOTA methods.

Weaknesses

* The method part is not clear. The relation among image diffusion model, video diffusion model, vision encoder and vision unet is confusing. Since 4 diffusion models are introduced and only 3 types of encoders and unet are shown in Figure 2, It s not clear whether image and video models share the parameters or not. * The evaluation of Table 3 is not sufficient. Only the text-video faithfulness (CLIPSIM) is evaluated, while the video quality (FVD) is not evaluated. * The proposed framework enables many different tasks. However, it does not outperform previous SOTA methods in many tasks, such as text-to-video generation, text-to-image generation, image captioning and video captioning.

Questions

* From Table 8, using both text and audio as input achieves higher FID compared to using each single modality as input. Could you explain why model achieves worse performance with more information as input? * From table 2 and table 3, CoDi does not outperform previous SOTA results. Do you think a model that can do all tasks need to sacrifice its performance on each specific task? * During training, the text encoder weights are frozen after training with images, would it result to a suboptimal problem when training with other modalities? * In Sec 3.3, image diffusion model and video diffusion model are introduced separately. However, in Figure 2, only vision UNet and Vision Encoder are shown. Does it mean image diffusion model share parameters with video diffusion model during training? * In table 4, why CoDi can outperform other diffusion-based method in image captioning?

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

3 good

Limitations

The authors adequately address the limitations and potential negative sosietal impact.

Reviewer dG2H6/10 · confidence 4/52023-07-06

Summary

The paper introduces Composable Diffusion (CoDi), an innovative generative model capable of producing any combination of output modalities, such as language, image, video, or audio, from any combination of input modalities. Unlike existing generative AI systems, CoDi can generate multiple modalities simultaneously and is not limited to a subset of modalities like text or images. To address the challenge of lacking training datasets for many modalities combinations, the authors propose a modality alignment approach in both the input and output space. This enables CoDi to condition freely on any input combination and generate any group of modalities, even if they are not present in the training data. CoDi employs a unique composable generation strategy that establishes a shared multimodal space through alignment in the diffusion process. This allows for the synchronized generation of intertwined modalities, such as temporally aligned video and audio. Offering high customization and flexibility, CoDi achieves impressive quality in joint-modality generation and either outperforms or matches the state-of-the-art unimodal models for single-modality synthesis.

Strengths

1. The paper is addressing an important problem of mapping modalities from any domain to any domain without fully paired data. 2. The proposed method is novel and reasonable. It is good to see that each different component can be trained separately. 3. The proposed bridging alignment is interesting.

Weaknesses

The proposed method shares some similarities with previous works. Nevertheless, this paper still contributes to the community in my opinion. It could be better to have a more specific discussions on the difference with the related work.

Questions

Please refer to 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

Yes.

Reviewer 7M7p6/10 · confidence 4/52023-07-07

Summary

The paper presents a new generative model called Composable Diffusion (CoDi). This model is capable of generating any combination of output modalities from any combination of input modalities, including language, image, video, or audio. Unlike other models that are limited to a subset of modalities like text or image, CoDi can generate multiple modalities in parallel. The authors have designed CoDi to align modalities in both the input and output space. This allows the model to condition on any input combination and generate any group of modalities, even if they are not present in the training data. A key feature of CoDi is its novel composable generation strategy. This involves building a shared multimodal space by bridging alignment in the diffusion process. This feature enables the synchronized generation of intertwined modalities, such as temporally aligned video and audio. The paper reports that CoDi achieves strong joint-modality generation quality. It either outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis.

Strengths

1. Originality: The paper introduces Composable Diffusion (CoDi), a new model in multimodal generation. This model is designed to process and generate modalities across text, image, video, and audio simultaneously. This is a novel contribution as it enables the generation of various output modalities from different combinations of input modalities. 2. Quality: The authors have conducted extensive experiments to demonstrate the capabilities of CoDi. The results show that CoDi can generate single or multiple modalities from a wide range of inputs. The model's performance is competitive with state-of-the-art models in tasks such as image and video generation, video captioning, and image synthesis from multiple input modalities. 3. Clarity: The paper is well-structured and provides clear explanations of the model's architecture and its generation strategy. The use of figures and tables helps to understand the model's capabilities and performance. 4. Significance: This work represents a step towards more comprehensive human-computer interactions by enabling the generation of multiple modalities in parallel. CoDi has potential applications in various areas, from content creation to human-computer interaction. The authors also provide a basis for future research in generative artificial intelligence. In summary, the paper presents a significant and original contribution to the field of multimodal generation, demonstrating high-quality research and clear presentation.

Weaknesses

The paper presents a novel approach to multimodal generation, but there are several areas where it could be improved: 1. Evaluation Metrics: The evaluation of the model's performance is primarily based on quantitative metrics such as Frechet Inception Distance (FID) and CLIPSIM. These metrics, while useful, may not fully capture the perceptual quality or coherence of the generated outputs. Incorporating user studies or other qualitative evaluations could provide a more comprehensive understanding of the model's performance. 2. Quality of Generated Results: The quality of the generated results could be improved. The generated videos are relatively short, the quality of the images is perceptually low, and the generated text is often short and discontinuous. These factors could limit the practical utility of the generated outputs. 3. Preservation of Input Modality: The model primarily focuses on understanding between modalities, but it does not always preserve the faithfulness of the input modality. For instance, the output video and images do not consistently preserve the identity of the input image. This could limit the model's ability to generate accurate and coherent outputs across different modalities. 4. Cross-Modality Benefits: The paper does not convincingly demonstrate that the generation results benefit from cross-modality conditions. For example, Table 8 shows that the quality of image generation can even degrade when using conditions from two modalities. Similarly, Table 9 shows only marginal improvements in video quality when using multiple modalities. The authors should establish a benchmark that clearly demonstrates the benefits of using multiple modalities for generation. Without such evidence, the necessity of the proposed architecture could be questioned. 5. Omission of Baselines: In Table 2, the authors omit the StableDiffusion v1.5 baseline, which is the image Latent Diffusion Model (LDM) they used. Including this baseline could provide a more comprehensive comparison of the model's performance.

Questions

1. Evaluation Metrics: Could you provide more details on why you chose FID and CLIPSIM as the primary evaluation metrics? Have you considered incorporating user studies or other qualitative evaluations to assess the perceptual quality and coherence of the generated outputs? 2. Quality of Generated Results: Could you elaborate on the factors that might be contributing to the short and discontinuous text, short video length, and perceptually low-quality images? Are there potential improvements or modifications to the model that could address these issues? 3. Preservation of Input Modality: How does the model ensure the preservation of the identity or characteristics of the input modality in the generated outputs? Are there specific mechanisms in place to ensure this, or is it an area for future work? 4. Cross-Modality Benefits: Could you provide more evidence or a clearer explanation of how the generation results benefit from cross-modality conditions? The results in Tables 8 and 9 suggest that the benefits might be marginal or even negative in some cases. Could you clarify this? 5. Omission of Baselines: Why was the StableDiffusion v1.5 baseline omitted from the comparisons in Table 2? Including this baseline could provide a more comprehensive view of the model's performance relative to existing methods.

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

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

The authors have adequately addressed the limitations and potential negative societal impact of their work.

Authorsrebuttal2023-08-11

Thank you for your review. We acknowledge the importance of comparing our method with previous work using the FVD metric on MSR-VTT. However, almost no previous work tested FVD on MSRVTT. Plus most of these works are not open-source and unable to test. CogVideo stands out as an exception, being the only open-source video generation model (that is already in Table 3) available, and as such, we have included its FVD score in our table. Below, please find the extended Table 3 showcasing MSR-VTT text-to-video generation performance, now including FVD scores. ### Table 3: MSR-VTT text-to-video generation performance extended to FVD. | Method | Zero-Shot | CLIPSIM $\uparrow$ | FVD $\downarrow$ | |-----------------|:---------:|:---------:|:---------:| | GODIVA | No | 0.2402 | _ | | NÜWA | No | 0.2439 | - | | CogVideo | Yes | 0.2631 | 801.98 | | Make-A-Video | Yes | 0.3049 | - | | Video LDM | Yes | 0.2929 | - | | CoDi (Ours) | Yes | 0.2890 | 612.02 |

Reviewer dG2H2023-08-19

Thanks for the response. I have no other questions.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC