METHOD FOR ADAPTING TRAINED GENERATIVE MACHINE LEARNING MODELS

Smart overview of the Invention

The method described focuses on generating high-resolution images using a generative machine learning (ML) model on an electronic device. It involves obtaining a text prompt and a noisy image, which is then processed to create a high-resolution image through a series of denoising and upsampling steps. The approach allows for the adaptation of trained diffusion-based models to produce images with higher resolutions than those used during their training, without requiring additional training.

Technical Field

The application pertains to adapting trained generative ML models, specifically diffusion-based models, to generate high-resolution images. This method addresses the challenges of generating images at resolutions higher than the model's training resolution, which can lead to issues like object repetition and unrealistic patterns. The aim is to enhance the model's capabilities without the need for retraining, which is computationally intensive.

Background

Diffusion models are powerful but struggle to scale to higher resolutions seamlessly. Directly generating images at higher resolutions can lead to repetitive patterns and unrealistic textures. Existing solutions either improve global consistency with mixed results or alter the model architecture, compromising image quality. The disclosed method seeks to adapt diffusion models to generate high-resolution images efficiently, addressing these limitations.

Method Summary

The method involves generating an image from a text prompt at an initial resolution, then upsampling it to a higher target resolution using an upsampling module. Noise is added to the upsampled image, creating a noisy version at the target resolution. This is followed by a denoising process that maintains global structural features, ensuring the final image is coherent and free from artifacts. The process is executed on an electronic device with a processor and memory storing necessary instructions.

Advantages

The technique offers a straightforward, training-free way to enable pre-trained diffusion models to produce high-resolution images. It overcomes issues with repetitive patterns and structural distortions seen in other methods. By maintaining global and local structural features, the approach ensures high-quality images without altering the trained model itself. This is achieved through modules controlling inputs during image generation, making it efficient and practical for deployment on devices like smartphones and laptops.