Recent advancements in generative AI research have revolutionized the field of image generation and visual content creation, marked by significant breakthroughs in both quality and versatility. Various algorithms have been proposed to solve this problem, including Generative Adversarial Networks (GANs)1 and Variational Autoencoders (VAEs)2. Diffusion models have emerged as a leading technique in image generation, demonstrating impressive capabilities such as text-to-image synthesis, image-to-image transformation, and image inpainting3, 4, 5. Together, these advancements not only push the boundaries of artistic and practical applications but also pave the way for new possibilities in fields ranging from entertainment to scientific visualization.
AMD is excited to release one-step diffusion models that demonstrate the readiness of AMD Instinct™ MI250 accelerators for model training and further research. The models are designed to provide performance comparable to traditional full-step diffusion models while maintaining the efficiency needed for training on data center systems or deployment on edge devices, such as AI-enabled PCs and laptops.
AMD Nitro Diffusion Models:
The AMD Nitro Diffusio models are built from two popular open-source models, Stable Diffusion 2.1 and PixArt-Sigma. By utilizing a UNet architecture as the backbone, with a CLIP (Contrastive Language-Image Pre-Train) model as the text encoder for the former, and a Diffusion Transformer (DiT) with a larger T5 text encoder for the latter, AMD has created models combine efficiency with high image quality. The inference implementation leverages PyTorch, HuggingFace Accelerate library, and precomputed latent representations to enhance training throughput.
For more details on the generated images visual comparison and the inference performance measured on AMD Instinct MI250 accelerators, please refer to the full technical blog AMD Nitro Diffusion: One-Step Text-to-Image Generation Models. Appropriate prompts are also included.
To drive advancements in Generative AI, AMD has released the models and codes to the open-source community for more users to download and explore new possibilities in image generation and visual creation. Completed model files and code instructions are available via the AMD Hugging Face model cards AMD Stable Diffusion 2.1 Nitro, AMD PixArt Sigma Nitro and the GitHub repository. Additionally, we encourage developers to utilize the AMD Developer Cloud, which offers remote access to select AMD GPUs for testing and development.
For further inquiries, please do not hesitate to contact the AMD team at amd_ai_mkt@amd.com.
