Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Latent Diffusion
Image generation process where noise is added and removed in a lower-dimensional latent space, making computation faster and more efficient than in pixel space.
Diffusion Process
Generation method that involves progressively destroying initial data by adding noise over multiple steps, then learning to reverse this process to reconstruct the data from noise.
Noise Schedule
Strategy defining the amount and nature of noise added at each step of the diffusion process, influencing the quality and speed of final generation.
U-Net Model
Convolutional neural network architecture with a 'U' shape, used in diffusion models to predict noise through its downsampling (analysis) and upsampling (synthesis) paths.
Latent Text-to-Image
Application of latent diffusion where a model generates images in a latent space from a text description, before decoding them into final pixel images.
Text Guidance (Classifier-Free Guidance)
Conditioning technique that uses a diffusion model trained simultaneously with and without text, and combines their predictions to strengthen adherence to the text description during generation.
Latent Decoder
Part of an autoencoder (like a VAE) that converts a compressed representation from the latent space into final data in the original space (for example, a pixel image).
Text Encoder
Model (often a Transformer like CLIP) that transforms a character string into a numerical vector in the latent space, providing conditioning for image generation.
Negative Embedding
Use of text embeddings trained to represent concepts to avoid, allowing to guide the generation of a diffusion model away from undesirable styles or objects.
Latent Inpainting
Process of modifying an existing image by masking an area and using a latent diffusion model to generate new coherent content in that area, while preserving the context.
Latent Outpainting
Extension of an existing image beyond its original boundaries using a latent diffusion model to generate plausible content that seamlessly blends with the base image.
Latent Image-to-Image
Technique where an input image is encoded into the latent space, then partially noised and denoised according to a text prompt, to transform the style or content of the original image.