Meta is about to release a pixel space model (Tuna-2)

Meta has introduced Tuna-2, a cutting-edge pixel space model designed for image generation that operates directly in pixel space rather than latent space, offering potential advantages in fidelity and control. However, the release comes with a significant catch: due to organizational policy constraints, Meta cannot release the fully trained production model weights. Instead, they provide a foundation checkpoint with a small number of layers removed from both the large language model (LLM) backbone and the diffusion head (flow head). The remaining components—including the vision encoder, projections, and embeddings—are fully preserved, allowing the community to fine-tune the model and re-learn the removed layers to restore full quality.

This deliberate breaking of the model is a strategic move to balance Meta's internal policies with its commitment to open research. By releasing a partially disabled version, Meta invites the research community to engage in a hands-on restoration process, effectively turning the release into a collaborative challenge. The Tuna-2 project, hosted on GitHub, emphasizes that with a short fine-tuning pass on custom data, the model can be quickly restored to its intended performance. This approach not only circumvents policy hurdles but also fosters innovation, as developers gain insights into the model's architecture and training dynamics while contributing to its improvement. The move is likely to spark interest among AI researchers and hobbyists eager to work with state-of-the-art pixel space generation techniques.