AMD releases Nitro-P pixel-space text-to-image framework
AMD researchers present Nitro-P, a VAE-free pixel-space text-to-image diffusion framework that utilizes a hierarchical semantic-to-pixel design. The model combines patch-level semantic modeling with a lightweight U-Net decoder using modulation-guided semantic injection to suppress patch artifacts and achieve high inference throughput. Trained entirely on public data, Nitro-P reaches a throughput of 10.26 samples/s on a single AMD Instinct MI300X GPU with a 512px generation size, while its distilled one-step variant achieves 90.02 samples/s. Open article →