Summary CityDreamer Compositional Generative Model of Unbounded 3D Cities arxiv.org
8,763 words - PDF document - View PDF document
One Line
The CityDreamer model creates limitless 3D cities by separating building creation from background objects and utilizing a bird's eye view scene representation along with MaskGIT and VQVAE for layout generation.
Slides
Slide Presentation (9 slides)
Key Points
- The CityDreamer model is a generative model designed for creating unbounded 3D cities.
- The model separates the generation of building instances from other background objects and uses a bird's eye view scene representation.
- The CityDreamer model utilizes MaskGIT and VQVAE for layout generation and tokenizing the maps and fields.
- The model employs volumetric rendering for generating accurate 3D geometry and photorealistic images of unbounded 3D cities.
- The evaluation of the CityDreamer model involves generating distinct city layouts and sampling different styles for each scene.
Summaries
30 word summary
The CityDreamer model generates unbounded 3D cities by separating building generation from background objects. It uses a bird's eye view scene representation and incorporates MaskGIT and VQVAE for layout generation.
37 word summary
The CityDreamer model is a generative model for creating unbounded 3D cities. It separates building generation from other background objects and uses a bird's eye view scene representation. The model utilizes MaskGIT and VQVAE for layout generation
397 word summary
The CityDreamer model is a generative model designed for creating unbounded 3D cities. It separates the generation of building instances from other background objects and uses a bird's eye view scene representation. The model employs a volumetric renderer to generate
Generative models for scene-level content generation face challenges due to the high diversity of scenes. Some approaches have achieved 3D-aware scene synthesis but lack full 3D consistency or support for feed-forward generation of novel worlds. Other works focus on indoor
The CityDreamer model focuses on generating unbounded 3D cities by creating extendable semantic maps and height fields. It utilizes MaskGIT and VQVAE for layout generation and tokenizing the maps and fields. The bird's-eye-view (
The document describes a compositional generative model for creating unbounded 3D cities. The city background generator is trained using a combination of reconstruction loss and adversarial learning loss. Volumetric rendering is used for the building instance generator, which incorporates
The article discusses the evaluation protocols and results of the CityDreamer compositional generative model of unbounded 3D cities. The evaluation involves generating 1024 distinct city layouts and sampling 20 different styles for each scene. Evaluation metrics include F
The CityDreamer model is capable of generating accurate 3D geometry and photorealistic images of unbounded 3D cities. A user study was conducted to assess the quality and consistency of the generated cities, showing that the proposed method outper
The entertainment industry has a high demand for generating content for computer games and movies. However, there are limitations to the current city layout generation process, as it cannot model and generate concave geometries like caves and tunnels. Additionally, the individual generation of
The document is a list of references to various papers and studies related to generative models and 3D city modeling. These references include papers on generative adversarial nets, 3D image synthesis, neural rendering of Minecraft worlds, human sensing in
This text excerpt provides references to various papers and studies related to 3D avatar generation, image synthesis, and city layout generation. The papers mentioned cover a range of topics including data platforms for learning 3D structures, image synthesis techniques, 3
The excerpt discusses the qualitative comparison of different building instance generator variants and the effectiveness of different generative scene parameterization. The results show that using the Global Encoder and Hash Grid as scene parameterization produces more natural city backgrounds but decreases the quality of generated buildings