ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial2 min readDisconfirmAI Editorial

AMD Releases Nitro-P: Pixel-Space Text-to-Image Model Achieves 90 Samples/s in One-Step Generation

AMD releases Nitro-P, a VAE-free pixel-space text-to-image model. By removing the final decoding step, it achieves 90 samples/s throughput in one-step generation on MI300X GPUs.

2026-11-20 08:00 GMT+8

On November 20, 2026, AMD released Nitro-P, an end-to-end pixel-space text-to-image (T2I) diffusion framework. Built upon its previously released latent-space model Nitro-E, Nitro-P extends the family to a VAE-free design by adopting a hierarchical semantic-to-pixel architecture. This structure consists of a patch-level semantic modeling component based on E-MMDiT and a pixel-level modulation-guided reconstruction component implemented as a lightweight U-Net decoder.

The core advantage of operating directly in pixel space is the removal of the Variational Autoencoder (VAE) decoder from the inference pipeline. In traditional latent-space models, generated latent representations must be converted back to RGB images by a VAE decoder, which incurs non-negligible computational overhead. Nitro-P bypasses this final decoding pass entirely. According to AMD's self-reported benchmarks on a single AMD Instinct™ MI300X GPU with a batch size of 32, the full-step Nitro-P model achieves a throughput of 10.26 samples/s for 512px image generation. Its distilled one-step variant further boosts throughput to 90.02 samples/s.

To isolate the effect of removing the VAE decoder, AMD compared the end-to-end throughput of Nitro-P against Nitro-E. The results show that the performance gap is largest in the few-step regime. At 512×512 resolution, the one-step distilled Nitro-P reaches 90.02 samples/s, versus 44.13 samples/s for Nitro-E. However, as the number of sampling steps increases, the gap narrows. At the full 20-step setting, Nitro-E is actually faster (18.83 samples/s vs. 10.26 samples/s), because the denoising backbone dominates runtime and the latent-space path remains cheaper per step. These findings indicate that Nitro-P is particularly well suited to low-NFE (Number of Function Evaluations), high-throughput serving scenarios.

Sign in / Sign up
ReadingResearchRadarInvestment framework

In terms of generation quality, Nitro-P achieves competitive GenEval scores of 0.88 at 512×512 and 0.87 at 1024×1024. The one-step distilled variant sees a moderate drop in quality, scoring 0.85 at 512×512. All models were trained from scratch on public data, including a curated dataset of 25 million real and synthetic images. AMD has released both the model weights and the complete training code to support reproducibility.

Sources

  • rocm.blogs.amd.comROCm Blogs↗
Continue researching →