Iris-3B does not use a VAE (non_vae = true).
I announced the release of Iris‑3B, a 3 billion‑parameter pixel‑space text‑to‑image (T2I) model that generates RGB pixels directly without a variational auto‑encoder. The model was trained from scratch on a mixed image‑text and unlabeled image dataset to serve both generative (T2I) and discriminative (vision) tasks. All weights, code, and a demo are released under the Apache 2.0 license, enabling the community to explore detail‑critical applications such as depth estimation and image restoration.