AMD launches Nitro-P: image generator works directly in pixels, up to 90 images per second

AMD Nitro-P is a text-to-image model that generates images directly in pixel space, without the VAE (variational autoencoder) decoder that latent models use. The announcement was published on Friday (20/11) on AMD's ROCm blog, together with the model weights and the complete training code — all open for the community to reproduce and use.

How Nitro-P works

Most image generators today work in latent space: a VAE first compresses the image into a smaller representation, diffusion happens in that compressed version, and in the end the VAE decoder converts everything back to RGB pixels. The process is efficient, but it depends on the quality of that compression — and the final decoder costs time, especially in few-step generation.

Nitro-P extends Nitro-E, AMD's earlier latent-space model, to direct pixel generation. It uses a hierarchical two-stage design: a semantic component based on the E-MMDiT transformer models the image structure in 16x16 pixel patches, using the Qwen3-1.7B text encoder; then a lightweight U-Net decoder reconstructs the pixels at full resolution, guided by modulation-guided semantic injection. In practice, the model avoids the VAE bottleneck and reduces the patch-boundary artifacts that appear when each piece of the image is decoded separately.

Results: high quality and 90 images per second in one step

In AMD's tests, the full Nitro-P reached 0.88 on GenEval at 512x512 images and 0.87 at 1024x1024, with a throughput of 10.26 samples per second on a single AMD Instinct MI300X GPU (batch size 32). The distilled one-step variant reaches 90.02 images per second at 512x512 and 26 per second at 1024x1024, with a GenEval of 0.85 — a small quality drop for a large speed gain.

The numbers need to be read carefully: at 20 steps, Nitro-P is slower than Nitro-E (10.26 versus 18.83 samples per second). The pixel-space model's advantage shows up in one-step generation, where it is about twice as fast as one-step Nitro-E (90.02 versus 44.13 samples per second) — because the VAE decoder, which weighed on the latent model, simply no longer exists in the pipeline.

Open training, public data

Nitro-P was trained from scratch on 25 million images from public sources: 11.1 million real images from Segment Anything 1B, 4.4 million synthetic images from JourneyDB, 9.5 million from FLUXDB and 120,000 high-quality images from MiniT2I. According to AMD, the model is fully supported by ROCm software, and the weights and training code were released as part of the company's commitment to open-source research.

Image: artwork from AMD's ROCm blog for Nitro-P (AMD / ROCm Blog).

Sources: AMD ROCm Blog — Nitro-P: Pixel-Space Text-to-Image; AI.info — AMD publishes Nitro-P for pixel-space image generation.

Transparency note: this article was produced with the assistance of artificial intelligence, cross-checking public posts on X and sources on the internet, and reviewed before publication. Check the original sources to confirm the information.

Por Muse