AI 용어집
인공지능 완전 사전
Diffusion Models
Generative architecture that uses a progressive process of adding and removing noise to transform data into pure noise and then reconstruct realistic samples. These models excel in generating high-quality images with superior training stability compared to GANs.
Forward Diffusion Process
Training phase where Gaussian noise is progressively added to the original data over several time steps to reach a pure noise distribution. This Markov process is designed to be mathematically reversible and predictable.
Reverse Diffusion Process
Generation phase where a neural network learns to progressively denoise a pure noise distribution to reconstruct coherent data. This process is guided by estimating the score gradient of the data distribution.
DDPM (Denoising Diffusion Probabilistic Models)
Fundamental class of diffusion models introduced by Ho et al. in 2020, using a linear variance schedule and a noise prediction objective. DDPM establishes the mathematical foundations for modern diffusion architectures.
Noise Schedule
Temporal parameter controlling the amount of noise added at each step of the forward diffusion process. A well-designed schedule optimizes the balance between information preservation and denoising efficiency.
Classifier-Free Guidance
Generation control technique that combines conditional and unconditional predictions to improve text fidelity without requiring an external classifier. This method allows fine control of generation while preserving diversity.
Latent Diffusion Models
Architecture optimizing diffusion models by working in a compressed latent space rather than directly in pixel space. This approach drastically reduces computational costs while maintaining high generation quality.
Ancestral Sampling
Stochastic sampling method that combines deterministic denoising with controlled noise addition to improve the diversity of generated samples. This technique balances quality and creativity in generation.
Diffusion Timesteps
Discrete number of steps used in the diffusion process, typically between 100 and 1000 steps for an optimal quality-performance tradeoff. The selection of timesteps directly influences the fine detail of generated samples.
Denoising U-Net
Neural architecture with residual connections and attention, specifically designed to predict and remove noise at each diffusion step. The U-Net structure effectively preserves spatial information while capturing global dependencies.
Stochastic Differential Equations
Continuous mathematical formulation unifying the forward and reverse diffusion processes within a rigorous theoretical framework. This approach enables theoretical analysis and the development of new sampling algorithms.
Conditional Diffusion
Extension of diffusion models incorporating external conditions like text, images, or classes to guide the generation. This approach allows precise control over the characteristics of generated samples.
Guidance Scale
Parameter controlling the influence of conditions on the generation process, allowing adjustment of the tradeoff between fidelity and creativity. A high scale strengthens adherence to instructions while a low scale promotes diversity.
Progressive Resampling
Iterative improvement technique applying multiple denoising cycles to refine details and correct artifacts. This method optimizes the final quality of generations at the cost of increased computation time.
Cross-Attention
Attention mechanism allowing diffusion models to effectively merge textual and visual information during denoising. This architecture is crucial for semantic coherence in text-to-image generation.
DDIM (Denoising Diffusion Implicit Models)
Deterministic variant of diffusion models enabling accelerated sampling with fewer steps while preserving quality. DDIM transforms the stochastic process into a non-Markovian deterministic mapping.