AMD launches Nitro Diffusion for one-step image generation

AMD Nitro Diffusion is an open-source project that shrinks the traditional multi-step diffusion pipeline down to a single inference step. It lets you generate high-quality images from text prompts virtually instantly, cutting down the compute overhead whether you are running in a data center or locally on an AI PC.

How it works

Standard diffusion models typically need dozens of denoising steps to render a clean image. Nitro Diffusion delivers comparable visual quality in just one step by adapting two established open-source backbones.

The first variant is AMD Stable Diffusion 2.1 Nitro, which pairs a classic UNet architecture with a CLIP text encoder for lean and fast generation.

The second is AMD PixArt Sigma Nitro, built on a modern Diffusion Transformer (DiT) backbone combined with a larger T5 text encoder to handle more nuanced textual prompts and finer details.

Under the hood, the training and inference pipeline utilizes PyTorch alongside the Hugging Face Accelerate library, taking advantage of precomputed latent representations to maximize training throughput. Everything was trained and validated directly on AMD Instinct MI250 accelerators.

Where to get it

AMD made both the weights and code fully accessible to the community without waitlists.

You can pull the model cards directly from Hugging Face under AMD Stable Diffusion 2.1 Nitro and AMD PixArt Sigma Nitro, and inspect the codebase on GitHub. If you do not have compatible AMD silicon on hand, AMD also offers remote access to select GPUs via the AMD Developer Cloud for testing.

Why it matters

If you are building real-time generation features or trying to run image generation locally on edge laptops without draining the battery, one-step diffusion is the fastest path forward. Having ready-to-run checkpoints on two distinct architectures makes it easy to drop into existing pipelines.

Beyond raw inference speed, this is a clear signal from AMD that its Instinct MI250 hardware and ROCm software ecosystem are ready for serious generative AI workloads. Giving developers functional, open-weight models trained on their own silicon is exactly how AMD chips away at Nvidia's software moat.