Poor Results with Building Obj Model

[bdcms1]

Thank you for your excellent work.
We applied your TEXGen to an OBJ format of a building, but the results were not up to our expectations. Specifically, the front doors and windows were not generated correctly, despite using brick walls for the structure.
Due to some limitations, we are unable to share the results directly, but we can inform you that we used data sourced from the website listed below.
https://www.cgtrader.com/free-3d-models/architectural/architectural-street/townhouse-7ea1baf2-c095-4959-a878-7f2ffcab1bbc
We appreciate your understanding and would welcome any suggestions you may have to improve the outcome.

[XinYu-Andy]

Thank you for testing our work and sharing your feedback.

Our method does have limitations at this stage. Since we perform diffusion directly in UV space and train from scratch with limited open-source data, textures may not always align perfectly with semantic structures. Unlike methods using image-pretraining, our approach prioritizes speed and end-to-end processing, which can sometimes sacrifice high-level alignment. We suggest using the method mainly for inpainting low-level details rather than generating structure-aligned textures. Feed-forward diffusion models are still in an early phase and less stable than test-time optimization methods, but we are committed to improving this approach in the long term by scaling data and with better algorithm (we are working on this).

For reference, if you are not constrained by generation time, you might want to look into our previous work, CSD, which produces highly harmonious and detailed textures, albeit with a longer processing time.

Thank you again for your understanding.