Spaces:
Running
on
Zero
Testing images - bad results
Feedback:
Fucks up faces most of the time. Check the dragon head or goku. Either it squishes them or washes them up. Also, it loves to flatten the model. Like, Goku starts well at the feet and lower legs, but further up it presses the model into 2D completely. Pikachu worked quite well as a 3D image with shadows, but 2D image went all flat (guess not enough data in training?). Overall 2D images didn't really work out, but if the images has lots of shadows and light reflections, it performs well (unless you look into the face :D).
Don't take my comments to harsh! Love the progress, I just always spit everything blatantly out what I see :D
@silverbeard Thanks for the feedback. Our mesh generation quality highly depends on the quality of multi-view images generated by the multi-view diffusion model. We use Zero123++, which is only trained on Objaverse rendering images and not good at generating human body or faces. If you try the goku image in the official demo of Zero123++, you will find that it generats flat 2D images and distorted face (as shown below), that's why the 3D output is terrible. Maybe we need a more powerful multi-view diffusion model to improve the results.
