Spaces:

---

CompVis

/

EDGS

like 39

Running on Zero

Hello, thank you for your interest in our work. You're right, there are no differences to the core of our method between the demo, the colab notebook or just a python code we provide. The only difference is in the configurations.

1. Usually, for 360 scenes you need a lot of views that describe the scene from each and every viewpoint. But to carefully localize them you need to run COLMAP to determine intrinsics and extrinsics for all the cameras. This is computationally heavy and may take from 30 minutes to 2 hours for 150 photographs in our experiments. So it is very inpractical for the scanario of a Hugging face demo. Hence, we limit this case(HF demo) for the forward-facing scenes which can work efficiently using just a few dozens of images. 2) Moreover, we observed that for this scenario scene converges very fast - in about 200-400 step(maybe 1000 steps for very tricky scenes). So this differs from the conventional 30K steps people usually do for 3DGS.
2. But probably the only substantial difference is the parameters cfg.init_wC.nns_per_ref which determines with how many neighboring images do we compare each of the reference images to extract matched points and use them to initialize splats. In the demo version we set this number to 1 but in the colab version and in our code we set it to 3. The initialization time depends linearly on this parameter but gives just a small boost in quality measured in PSNR, LPIPS, SSIM.

Hope it helps you.

Thanks for getting back to me and providing such a detailed explanation and your awesome research!