Meet some problems when reproducing your highly detailed results.

[snowflakewang]

Hello, thank you for your brilliant work! Your results shown on the project page are really highly detailed.
For example, your 3D CGI Pixar Lionel Messi artfully kicking paint-filled bottles has a detailed back view of Messi. But after I ran NeRF & NeuS, Geometry, and Texture boosting, I still got a result with relatively non-detailed texture. I also tried your code in the threestudio-dreamcraft3d repo, which produced a slightly better result than the former one. But it still falls behind your result on the project page.

I wonder whether I ignore some implementation details when reproducing your results (I have used .yaml files given by the repo and 8 * A100 GPUs to train all stages). If you could give me an instruction to obtain better results, I will very appreciate it.

Thanks a lot :)

The first image is produced by threestudio-dreamcraft3d.

The second image is produced by DreamCraft3D.

[MrTornado24]

Hi, thanks for pointing it out. I think the main difference is that we rewrite BSD guidance into a single guidance file for efficiency. We are still tuning some hyperparameters to make the results better and more robust.

[snowflakewang]

Thanks for your quick reply. Do you mean that using 1) hyperparameter tuning and 2) the BSD guidance you rewrote, can achieve similar results compared to the results on the project page?

[MrTornado24]

@snowflakewang Yeah, exactly.

[snowflakewang]

Thank you for your reply! According to other issues, I find that zero123-xl.ckpt is of importance for geometry sculpturing and primary texture modeling. Could you please tell me, if I want to use zero123-xl.ckpt, which guidance should I use, zero123-guidance or zero123-unified-guidance?

Thanks a lot :)