HIGH-RESOLUTION IMAGE GENERATION USING DIFFUSION MODELS

Smart overview of the Invention

The patent application outlines methods, systems, and apparatuses for generating high-resolution images using diffusion models. It involves obtaining a prompt and using a first diffusion model to generate a predicted denoised image at a low resolution. This image is then upsampled to a higher resolution, and a second diffusion model generates a final output image based on this upsampled version. This process aims to improve image quality and reduce processing time compared to conventional methods.

Background

Diffusion models are machine learning models particularly suited for image generation, where they progressively remove noise from an image through a series of steps. These models typically face challenges in generating high-resolution images due to increased processing time and memory demands. Traditional approaches involve generating a low-resolution image and then upsampling it, which often results in less accurate outputs.

Innovation

The proposed method first creates a noiseless prediction at a low resolution, which is then upsampled before being processed by a second diffusion model to produce the high-resolution output. This approach contrasts with conventional techniques that might add noise correction steps, potentially introducing errors. By avoiding these corrections, the new method achieves higher quality images more efficiently.

Technical Details

• A first diffusion model generates a low-resolution denoised prediction during an intermediate reverse diffusion step.
• This prediction is upsampled to a higher resolution without adding noise.
• A second diffusion model uses the upsampled prediction to generate the final high-resolution output image.
• The process reduces processing time and memory costs by leveraging larger models than conventional methods allow.

Advantages

By eliminating the need for noise correction in the upsampling process, the method enhances image quality and accuracy. It also optimizes processing efficiency, enabling the use of larger diffusion models within existing hardware constraints. This leads to better high-resolution images without the drawbacks of traditional approaches that rely on noisy intermediate images.