Magic3DModel<T>
Models & Types · AiDotNet.Diffusion.ThreeD
Magic3D model for high-quality text-to-3D generation using score distillation.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → ThreeDDiffusionModelBase<T> → Magic3DModel<T>
For Beginners
Magic3D creates 3D models from text descriptions:
How Magic3D works:
- Coarse stage: Optimize a neural radiance field (NeRF) using SDS from a low-res diffusion model
- Fine stage: Convert NeRF to a mesh and optimize with SDS from a high-res latent diffusion model
- Result: High-quality textured 3D mesh
Key characteristics:
- Two-stage coarse-to-fine optimization
- Coarse stage: NeRF + low-res SDS (64x64 base model)
- Fine stage: DMTet mesh + high-res SDS (latent diffusion)
- 2x faster than DreamFusion
- 8x higher resolution meshes than DreamFusion
- Uses both pixel-space and latent-space diffusion guidance
Advantages:
- High-quality textured 3D meshes
- Much faster than DreamFusion
- Better geometry through mesh refinement
- Supports both NeRF and mesh representations
Limitations:
- Multi-view consistency (Janus problem)
- Optimization takes minutes per object
- Quality depends on 2D diffusion prior
How It Works
Magic3D is a two-stage coarse-to-fine text-to-3D generation framework by NVIDIA. It uses Score Distillation Sampling (SDS) from a 2D diffusion model to optimize a 3D representation, first as a coarse NeRF then refined as a textured mesh.
Technical specifications:
- Coarse stage: Instant-NGP NeRF, 64x64 diffusion guidance, ~40 min optimization
- Fine stage: DMTet mesh, latent diffusion guidance, ~20 min optimization
- Diffusion prior: eDiff-I (coarse) + Stable Diffusion (fine)
- SDS guidance scale: 100 (coarse) → 7.5 (fine)
- NeRF resolution: 128^3 hash grid
- Mesh resolution: 512^3 DMTet grid
Reference: Lin et al., “Magic3D: High-Resolution Text-to-3D Content Creation”, CVPR 2023
Example
var magic3d = new Magic3DModel<float>();
var mesh = magic3d.GenerateMesh(
prompt: "A blue poison dart frog sitting on a lily pad",
resolution: 256,
numInferenceSteps: 64,
guidanceScale: 100.0,
seed: 42);
Constructors
| Constructor | Summary |
|---|---|
Magic3DModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, UNetNoisePredictor<T> coarseUnet, UNetNoisePredictor<T> fineUnet, StandardVAE<T> vae, IConditioningModule<T> conditioner, int defaultPointCount, int? seed) |
Initializes a new instance of Magic3DModel with full customization support. |
Properties
| Property | Summary |
|---|---|
IConditioningModule<T> Conditioner { get; } |
|
INoisePredictor<T> CoarseModel { get; } |
Gets the coarse-stage noise predictor (pixel-space SDS). |
INoisePredictor<T> NoisePredictor { get; } |
|
IVAEModel<T> VAE { get; } |
|
bool SupportsMesh { get; } |
|
bool SupportsNovelView { get; } |
|
bool SupportsPointCloud { get; } |
|
bool SupportsScoreDistillation { get; } |
|
bool SupportsTexture { get; } |
|
int LatentChannels { get; } |
Methods
| Method | Summary |
|---|---|
Mesh3D<T> ImageTo3D(Tensor<T> inputImage, int numViews, int numInferenceSteps, double guidanceScale, int? seed) |
Not supported: image-to-3D is not part of this model’s published pipeline. |
ModelMetadata<T> GetModelMetadata() |