AI 용어집
인공지능 완전 사전
3D Diffusion
Generative process applying diffusion principles to three-dimensional data, such as point clouds or meshes, to create new geometric shapes from initial noise.
Score Distillation Sampling (SDS)
Optimization technique using score gradients from a pre-trained 2D diffusion model to guide the synthesis of 3D assets, without requiring supervised 3D data.
3D Latent Representation
Encoding of 3D geometry into a lower-dimensional space where the diffusion process is more efficient and stable, often via neural networks such as autoencoders.
Signed Distance Field (SDF)
Geometric representation where each point in space is assigned a value equal to its distance to the nearest surface, signed to indicate whether the point is inside or outside the object.
Mesh
Set of vertices, edges, and faces defining the shape of a polyhedral object in 3D, often used as structured output for geometric diffusion models.
Geometric Constraints
Conditions imposed during the diffusion process to ensure that generated shapes respect specific properties such as manifoldness, watertightness, or symmetry.
Neural Implicit Modeling
Approach where 3D geometry is defined by a neural network that maps spatial coordinates to a field value (e.g., SDF), enabling a continuous and differentiable representation.
Geometric Denoising
Step in the 3D diffusion process where a neural network learns to reverse the progressive addition of noise to a geometric representation to reconstruct a coherent shape.
Conditional 3D Generation
Generation of 3D shapes guided by additional inputs such as text, a 2D image, a sketch, or a partial shape to control the final result.
Radiance Field
Volumetric representation that encodes the color and density of light for each point in space, often used in conjunction with diffusion to generate photorealistic 3D scenes.
Topology Preservation
Objective aimed at maintaining stable topological characteristics (such as the number of holes or connected components) throughout the diffusion and generation process.
Diffusion on Voxel Grid
Application of diffusion models to discretized 3D representations on a voxel grid, where each voxel encodes a property like occupancy or density.
Projection Guidance
Method where 2D views generated by a diffusion model are projected into 3D space to constrain and refine the consistency of a 3D model under construction.
Diffusive Autoregression
Generation strategy where diffusion is applied sequentially to parts of a 3D object (e.g., mesh patches), with each part conditioned on the previous ones.
3D Consistency Loss
Loss function used during training to penalize inconsistencies between different representations or views of the same 3D object generated by the model.
Random Walk Sampling
Sampling technique for point clouds or meshes in the context of diffusion, where noise is added or removed by following a random walk on the geometry structure.
Geometric Component Decomposition
Process of separating a 3D shape into semantic substructures (e.g., legs, torso) before diffusion, allowing finer control and more coherent generation.