Core Concepts
DreamScape leverages the strengths of Gaussian Splatting and Large Language Models to generate high-fidelity 3D scenes from textual descriptions, achieving both instance-level realism and global consistency.
Abstract
DreamScape presents a novel approach for creating 3D scenes from textual prompts. The key component is the 3D Gaussian Guide (3DG2), which is derived from text prompts using Large Language Models (LLMs) and serves as a comprehensive representation of the scene. 3DG2 includes semantic primitives (objects), their spatial transformations, and scene correlations, enabling DreamScape to employ a local-global generation strategy.
During local optimization, DreamScape uses a progressive scale control technique to ensure the scale of each object aligns with the overall scene. At the global level, a collision loss between objects is used to prevent intersection and misalignment, addressing potential spatial biases of 3DG2 and ensuring physical correctness.
To model pervasive objects like rain and snow, DreamScape introduces sparse initialization and corresponding densification and pruning strategies. This allows for more realistic representation of such objects distributed extensively across the scene.
Experiments demonstrate that DreamScape can generate high-fidelity 3D scenes from text prompts, outperforming existing state-of-the-art methods in terms of semantic accuracy, visual quality, and multi-view consistency. The approach also supports flexible editing capabilities, enabling users to modify object positions, scales, and rotations.
Stats
An astronaut stood under a big tree.
Some pink petals were flying in the air, and some petals were piled up on the grass.
A modern style bedroom with a nightstand in dark, a lamp illuminated the surroundings.
An amusement park swing, a seesaw, a glowing street lamp on the grass.
A snowman wearing a Santa hat is standing on a cobblestone road. There is a lot of snow on the road. Snow is flying in the air.
Quotes
"DreamScape leverages the strengths of Gaussian Splatting and Large Language Models to enhance 3D fidelity and reduce discrepancies with textual descriptions."
"DreamScape employs a progressive scale control technique during local object generation, considering the scale of each object in relation to the overall scene."
"DreamScape introduces the concept of pervasive objects, proposing sparse initialization and developing corresponding densification and pruning strategies for such objects."