AMD Publishes Nitro-P, a VAE-Free 679M Pixel-Space T2I Model
TL;DR
- Nitro-P is a 679M-parameter pixel-space text-to-image diffusion framework that drops the VAE decoder and uses a lightweight U-Net instead.
- On a single AMD Instinct MI300X, the full model hits 0.88 GenEval at 10.26 samples/s at 512×512, and 0.87 at 1024×1024.
- A one-step distilled variant pushes 512×512 throughput to 90.02 samples/s with GenEval slipping to 0.85.
At 1024×1024, AMD's new Nitro-P scores 0.87 on GenEval against PixelDiT's 0.74. The model has 679M parameters and skips the VAE decoder entirely.
Published November 20 on the ROCm Blogs by Jingai Yu, Tong Shen, Dong Zhou, Dong Li and Emad Barsoum, Nitro-P pairs a patch-level E-MMDiT backbone with a lightweight U-Net for pixel reconstruction. "We present Nitro-P, an end-to-end pixel-space text-to-image (T2I) diffusion framework," the authors write. The post describes it as a model that "extends the Nitro family from latent-space T2I generation to a VAE-free pixel-space design."
The throughput numbers are the pitch. On a single AMD Instinct MI300X with batch size 32, the full model runs 10.26 samples/s at 512×512 while scoring 0.88 GenEval; PixelDiT on the same resolution comes in at 0.78 and 5.54 samples/s. "The one-step distilled model further increases throughput to 90.02 samples/s with moderate GenEval drop (0.85)," the post reports. At 1024×1024 the full model drops to 2.44 samples/s and the one-step variant to 26.00.
Text conditioning uses Qwen3-1.7B. "Compared with Llama 3.2 1B, Qwen3-1.7B demonstrates a better understanding of color and spatial relationships," the authors note. Training ran on 25M text-image pairs under a resolution curriculum from 256×256 to 512×512 and finally to 1024×1024.
Nitro-P lands inside our wider thread of AMD coverage: 23 tracked stories in the last 90 days, most of them about export policy rather than research. A pixel-space T2I paper shipping on AMD's own silicon is a different beat than the Commerce-approval headlines that have dominated the tracker lately.
Originally reported by rocm.blogs.amd.com
Read the original article →Original headline: AMD Ships Nitro-P, a VAE-Free 679M Pixel-Space Text-to-Image Diffusion Model That Beats PixelDiT