US20260320867
2026-10-15
Physics
G06T11/60
The patent application outlines a method for generating images from text prompts using a text-to-image generative model. This model incorporates two sets of parameters: one shared across various styles and another specific to the style specified in the text prompt. The system is designed to produce images that align with the user's specified style and object, offering flexibility in image generation.
The technology leverages neural networks to process images, utilizing machine learning models that predict outputs based on input data and model parameters. Deep models, such as deep neural networks, use multiple layers to transform inputs into outputs. These layers apply non-linear transformations, allowing the model to generate complex outputs from simple inputs.
One significant advantage of the described system is its adaptability across diverse image styles. Users can adjust the model to a specific style with minimal reference images, sometimes just one. This is achieved through style-specific adapters within the model's layers, which allows efficient fine-tuning. The process is less computationally demanding than traditional pre-training, enabling it to run on less powerful hardware like mobile devices or home computers.
The image generation system operates by receiving a text prompt from a user, which specifies the desired image style and, optionally, specific objects to include or exclude. The system processes the text prompt to produce an image that matches the described style and content. This flexibility enables the creation of images ranging from cartoons to photorealistic renderings, based on user instructions.
The system employs a reverse diffusion process to generate images, either in pixel space or a lower-dimensional latent space. This process involves progressively removing noise from a noisy version of the image to achieve the final output. The model can map text prompts to images using intermediate representations, updating these representations at each diffusion step to refine the image according to the specified style and content.