DistScene
An object-to-scene distillation framework for generating 3D scenes from a single image. It explicitly models environmental components, jointly generates objects and surroundings, and uses knowledge distillation to transfer pretrained priors, achieving compositional 3D scene generation with coherent spatial relationships.
Timeline (last 90 days)
DistScene is an object-to-scene distillation framework for generating compositional 3D scenes from a single image.
DistScene's generation pipeline consists of three modules: Scene-Frame Generation, Object-Centric Refinement, and Object-to-Scene Distillation.
Scene-Frame Generation separately generates environment components and object components within a shared coordinate system, enabling their geometric shapes and relative positional relationships to be learned simultaneously.
Object-Centric Refinement performs refinement in each object's local coordinate system while preserving scene context information.
Object-to-Scene Distillation transfers pretrained object generation prior knowledge to the scene generation task through automatically synthesized and rendered composite scenes.
DistScene's distillation approach alleviates the problem of scarce scene-level training data.
DistScene was evaluated on both indoor and outdoor benchmarks, achieving improvements in scene-level spatial coherence compared to the evaluated baseline methods.
The claim of 'outperforming evaluated baselines' in the DistScene paper is limited in scope to a specific set of baselines, and generalization in broader scenarios remains to be further verified.
The DistScene project page is located at https://coolbeam.github.io/DistScene/.
All Facts (9)
DistScene is an object-to-scene distillation framework for generating compositional 3D scenes from a single image.
50%UnverifiedDistScene's generation pipeline consists of three modules: Scene-Frame Generation, Object-Centric Refinement, and Object-to-Scene Distillation.
50%UnverifiedScene-Frame Generation separately generates environment components and object components within a shared coordinate system, enabling their geometric shapes and relative positional relationships to be learned simultaneously.
50%UnverifiedObject-Centric Refinement performs refinement in each object's local coordinate system while preserving scene context information.
50%UnverifiedObject-to-Scene Distillation transfers pretrained object generation prior knowledge to the scene generation task through automatically synthesized and rendered composite scenes.
50%UnverifiedDistScene's distillation approach alleviates the problem of scarce scene-level training data.
50%UnverifiedDistScene was evaluated on both indoor and outdoor benchmarks, achieving improvements in scene-level spatial coherence compared to the evaluated baseline methods.
50%UnverifiedThe claim of 'outperforming evaluated baselines' in the DistScene paper is limited in scope to a specific set of baselines, and generalization in broader scenarios remains to be further verified.
50%UnverifiedThe DistScene project page is located at https://coolbeam.github.io/DistScene/.
50%