What changed
AMD has announced the release of Nitro-P, an end-to-end pixel-space text-to-image (T2I) diffusion framework. Built upon their previously released Nitro-E, Nitro-P extends the Nitro family from latent-space T2I generation to a VAE-free pixel-space design. The model uses a hierarchical semantic-to-pixel approach, combining a patch-level semantic modeling component based on E-MMDiT with a lightweight U-Net decoder for pixel-level reconstruction.
What this means in practice
Nitro-P operates directly in pixel space, removing the VAE decoder from the inference pipeline. This design allows for faster sampling, particularly in few-step scenarios where latent-space models still require a final decoding pass. On a single AMD Instinct™ MI300X GPU, Nitro-P achieves a throughput of 10.26 samples/s for 512px image generation with a batch size of 32. Its distilled one-step variant further boosts throughput to 90.02 samples/s. The model also achieves competitive GenEval scores of 0.88 at 512×512 and 0.87 at 1024×1024 resolutions.
Limitations and availability
The model is fully supported by AMD ROCm™ Software and is trained from scratch on public data. Both the model weights and complete training code are released to support reproducibility and broader adoption. The model is available for developers and researchers interested in pixel-space generation, with no specific hardware requirements beyond the supported ROCm environment.
