Table of Links

Abstract and 1 Introduction

2 Pre-Training

2.1 Tokenization

2.2 Pre-Training Data

2.3 Stability

2.4 Inference

3 Alignment and 3.1 Data

3.2 Fine-Tuning Strategy

4 Human Evaluations and Safety Testing, and 4.1 Prompts for Evaluation

4.2 Baselines and Evaluations

4.3 Inter-annotator Agreement

4.4 Safety Testing

4.5 Discussion

5 Benchmark Evaluations and 5.1 Text

5.2 Image-To-Text

6 Related Work

7 Conclusion, Acknowledgements, Contributors, and References

Appendix

A. Samples

B. Additional Information of Human Evaluations

Chameleon builds upon the lineage of works exploring token-based approaches for multimodal learning. The idea of using discrete tokens to represent continuous modalities like images was first explored in works like BEiT (Bao et al., 2021), which proposed a self-supervised vision representation learning method based on tokenized image patches. Aghajanyan et al. (2022) extended this idea to learning from mixed-modal documents through interleaved image and text tokens, allowing for joint reasoning over both modalities within a unified architecture. CM3Leon (Yu et al., 2023) further scaled up this approach to autoregressive text-to-image generation, building on the initial proposal of token-based image generation in DALL-E (Ramesh et al., 2021).

As a fully token-based early-fusion model, Chameleon differs from late-fusion approaches like Flamingo (Alayrac et al., 2022) which encode images and text separately before combining them at a later stage. Other models like LLaVA (Liu et al., 2023a), IDEFICS (Laurençon et al., 2023), and VisualGPT (Chen et al., 2022) also maintain separate image and text encoders. In contrast, Chameleon’s unified token space allows it to seamlessly reason over and generate interleaved image and text sequences, without the need for modality-specific components. This early-fusion approach, however, comes with significant challenges in terms of representation learning and alignment, as discussed in Baltrušaitis et al. (2018).

The most similar model to Chameleon is Gemini (Gemini et al., 2023), which also uses an early-fusion token-based approach. However, a key difference is that Gemini uses separate image decoders, whereas Chameleon is an end-to-end dense model without any routing components. This makes Chameleon a more general-purpose model for both multimodal understanding and generation tasks, similar in spirit to the Perceiver (Jaegle et al., 2021) architecture which also aims for a unified model across modalities and tasks.

In summary, Chameleon builds on a rich history of work in multimodal learning and token-based architectures, while pushing the boundaries in terms of model scale and architecture design. By demonstrating strong performance across a wide range of vision-language tasks and enabling new capabilities in mixed-modal reasoning and generation, Chameleon represents a significant step towards realizing the vision of general-purpose multimodal foundation models.