Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Volumetric Diffusion
Iterative generation process that applies progressive noise to a 3D volumetric representation, then learns to reverse this process to reconstruct coherent 3D shapes from random noise.
Score Matching for 3D
Training method for 3D diffusion models that involves predicting the gradient of the log-probability (the score) of the data distribution, allowing to guide the denoising process towards realistic samples.
EDM (Elastic Diffusion Model)
Variant of diffusion model adapted for generating deformable 3D meshes, where the diffusion and denoising process operates on vertex positions while preserving mesh topology.
SDF (Signed Distance Function)
3D surface representation where each point in space is associated with the signed distance to the nearest surface, used in diffusion models to ensure closed and artifact-free 3D geometries.
Text-to-3D Guidance
Technique that conditions the 3D diffusion process on text embeddings from CLIP-like models, in order to generate 3D objects that conform to a textual description.
DDIM 3D (Denoising Diffusion Implicit Models 3D)
Deterministic variant of diffusion models for 3D that allows for faster sampling with fewer denoising steps, using a non-Markovian process to reconstruct 3D data.
Gaussian Splatting Diffusion
3D generation method that applies diffusion to a collection of 3D Gaussians positioned in space, enabling the creation of photorealistic scenes with efficient rendering.
Multi-view Conditioning
Strategy for 3D diffusion models that uses 2D images of the object from multiple angles as input to guide the generation of a coherent and detailed 3D model.
Topology Regularization
Constraint applied during training or denoising of 3D diffusion models to preserve desirable topological properties, such as manifold-ness or connectivity of generated surfaces.
Shap-E Diffusion
Specific diffusion model that generates 3D implicit functions based on a latent space trained on meshes and point clouds, optimized for shape diversity and fidelity.
DreamFusion
Method that uses a pre-trained 2D diffusion model (such as Imagen) as a differentiable supervisor to optimize a 3D NeRF from text, without requiring supervised 3D data.
Magic3D
Two-stage text-to-3D generation system that first uses a diffusion model to create low-resolution geometry, then uses a 2D diffusion model to enhance it with high-resolution details.
ProlificDreamer
Variational diffusion-based 3D generation framework that optimizes a NeRF representation by maximizing the log-likelihood under a 2D diffusion model, improving the quality and diversity of results.