Developing a single AI model that can both "see" and "create" has long been a challenge because understanding an image requires high-level meaning, while generating one requires fine-grained pixel details. These two goals often clash within the same neural network, making it difficult to optimize for both at once.

The newly introduced Cheers model solves this by separating patch-level details from semantic representations. By using a specialized vision tokenizer and a cascaded flow matching head, the system can process images more efficiently. This decoupling allows the model to stabilize its understanding of what an image represents (the semantics) while maintaining the ability to produce high-fidelity visuals through gated detail residuals that fill in the fine textures.

One of the most impressive aspects of the research is its efficiency. Cheers achieves a four-fold increase in token compression—the process of turning image data into smaller units for the AI to read—compared to existing methods. This means the model can handle high-resolution images while using significantly fewer computational resources. In testing, Cheers matched or exceeded the performance of the much larger Tar-1.5B model on key industry benchmarks.

Perhaps most notably, the researchers achieved these results using only 20% of the training cost typically required for such advanced multimodal systems. By unifying autoregressive decoding for text and diffusion decoding for images within a single Transformer, Cheers provides a scalable blueprint for the next generation of efficient, all-in-one AI assistants.

Developing a single AI model that can both "see" and "create" has long been a challenge because understanding an image requires high-level meaning, while generating one requires fine-grained pixel details. These two goals often clash within the same neural network, making it difficult to optimize for both at once.

The newly introduced Cheers model solves this by separating patch-level details from semantic representations. By using a specialized vision tokenizer and a cascaded flow matching head, the system can process images more efficiently. This decoupling allows the model to stabilize its understanding of what an image represents (the semantics) while maintaining the ability to produce high-fidelity visuals through gated detail residuals that fill in the fine textures.

One of the most impressive aspects of the research is its efficiency. Cheers achieves a four-fold increase in token compression—the process of turning image data into smaller units for the AI to read—compared to existing methods. This means the model can handle high-resolution images while using significantly fewer computational resources. In testing, Cheers matched or exceeded the performance of the much larger Tar-1.5B model on key industry benchmarks.

Perhaps most notably, the researchers achieved these results using only 20% of the training cost typically required for such advanced multimodal systems. By unifying autoregressive decoding for text and diffusion decoding for images within a single Transformer, Cheers provides a scalable blueprint for the next generation of efficient, all-in-one AI assistants.