Chameleon: Meta's New Multi-Modal LLM

We present Chameleon, a family of early-fusion token-based mixed-modal models capable of understanding and generating images and text in any arbitrary sequence. The models demonstrate state-of-the-art performance in image captioning and competitive results in other tasks, outperforming several existing models in diverse evaluations.

• Chameleon models can process images and text.
• Training approach and alignment recipe provided.
• Outperforms Llama-2 in text tasks.
• Competitive with Mixtral 8x7B and Gemini-Pro.
• Exceeds GPT-4V in mixed-modal evaluation.