EucliDreamer is a method for automatically generating textures for 3D meshes using a high-capacity text-to-image diffusion model, such as Stable Diffusion. The texture is randomly initialized and iteratively refined through a process called score distillation sampling (SDS) by the diffusion model. Depth conditioning is introduced to enhance texture quality and speed up convergence by leveraging a version of Stable Diffusion that conditions on a depth image in addition to text prompts. The depth image is acquired by rendering from the 3D mesh. The optimization process involves updating a hash grid conditioned on text prompts and depth images.

In the experiments, 3D models from Objaverse are used to test EucliDreamer. The method is visually compared with other 3D texturing methods, showing superior quality in terms of realism, detail, cross-view consistency, and aesthetics. A user study involving professionals in 3D content creation ranks the textures generated by different methods, with EucliDreamer receiving high acceptance. The convergence speed of EucliDreamer is faster compared to a similar method, DreamFusion, highlighting its computational efficiency. Depth conditioning is shown to be important for ensuring correctness, improving detail quality, and texture sharpness in the generated textures.

Implementation details include parameter selection, infrastructure using the ThreeStudio framework, and hardware specifications. A series of experiments are conducted to understand the model and select parameters, such as elevation range, batch size, guidance scale, negative prompts, and data augmentation. The results support the effectiveness of EucliDreamer in generating high-quality textures for 3D models using depth-conditioned Stable Diffusion.