# AMD releases Nitro-P pixel-space text-to-image framework with ROCm support

> **Key takeaway:** Nitro-P is a VAE-free, pixel-space diffusion model that improves generation efficiency by removing the latent-space decoding step.

**Published:** 2026-11-20T18:03:12.042415+00:00 
**Source:** AMD ROCm Official Blog 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/amd-releases-nitro-p-pixel-space-text-to-image-framework-wit.html 

## Summary
AMD has announced Nitro-P, a new end-to-end text-to-image (T2I) diffusion framework that operates directly in pixel space. Unlike traditional latent-space models that rely on a Variational Autoencoder (VAE) to compress and reconstruct images, Nitro-P utilizes a hierarchical semantic-to-pixel design to eliminate the VAE decoder from the inference pipeline.

To support open-source research and adoption, AMD has released both the model weights and the complete training code. The framework is fully supported by the AMD ROCm software stack and has demonstrated high throughput on Instinct MI300X GPUs, particularly in its distilled one-step variant.

## Technical architecture of Nitro-P
- **Hierarchical Design:** The framework combines a patch-level semantic modeling component based on E-MMDiT with a pixel-level modulation-guided reconstruction component using a lightweight U-Net decoder.
- **VAE-free Pipeline:** By operating in pixel space rather than latent space, the model avoids the computational overhead and potential visual fidelity loss associated with the final VAE decoding pass during few-step sampling.
- **Performance Benchmarks:** On a single AMD Instinct MI300X GPU with a batch size of 32, the model achieves 10.26 samples/s for 512px images, while the distilled one-step version reaches 90.02 samples/s.
- **Generation Quality:** The full-step model achieved GenEval scores of 0.88 at 512x512 resolution and 0.87 at 1024x1024 resolution.

## Why it matters
The release of Nitro-P provides a high-performance, ROCm-native alternative to CUDA-centric latent-space diffusion models. By providing open weights and training code, AMD enables developers to build efficient, VAE-free generative AI applications that run natively on AMD hardware without vendor lock-in.