Texte vers 3D
Transformer une phrase en modèle 3D
Le texte intégral est présenté en anglais ; le titre et le résumé sont localisés.
CE QUE DÉSIGNE CETTE CAPACITÉ
Takes a text description and outputs a 3D asset: a mesh, a textured model or a renderable 3D representation. The output is neither an image nor a video but geometry that can be viewed from any angle and placed in a scene. Unlike image-to-3D it has no reference image at all, and unlike text-to-image its output carries a genuine third dimension.
Comment c'est fait
One route uses 2D generative models as supervision: a 2D diffusion model scores renderings from many viewpoints and that score optimises a 3D representation such as a neural radiance field or a Gaussian splat — score distillation. Another route trains a generator directly on large 3D datasets and emits mesh and texture in one or several steps. Output meshes often need retopology and decimation before entering standard rendering or game pipelines.
Produits représentatifs
2Organisations concernées
Usages typiques
- Draft props and sets for games and film
- Quick 3D showcases for products
- Turning industrial concepts into physical form
- 3D illustration for education
Comment on l'évalue
- CLIP similarity
- Semantic agreement between rendered views and the prompt
- Chamfer distance
- Mean surface-point distance between generated and reference meshes
- Human rating
- Ratings for shape completeness and texture quality
Limites et points difficiles
- Generated meshes often come out as fragmented triangles and need heavy retopology before use
- Backfaces and interiors are guessed, producing hollows and intersections when you orbit the model
- Material, lighting and geometry are not properly separated, so the look breaks when exported to another engine
Concepts sous-jacents
Modèles de diffusion
Apprenez mille petits pas de débruitage et vous créerez une image à partir de bruit pur
Génération multimodale
Un seul modèle qui apprend à parler, dessiner, bouger — et même à modéliser le monde 3D
Diffusion latente et contrôle conditionnel
Diffuser non pas sur les pixels, mais dans un espace sémantique comprimé