gsplat branch?

[antithing]

Hi! the gsplat branch looks interesting! are you seeing any quality improvements or speedups using gsplat?

[ndming]

Hi,

It's hard to say at the moment if the migration would improve the reconstruction quality or training time because we have not fully ported the code to gsplat yet. Nevertheless, as stated in the main paper, we will optimize our implementation of the material module and align everything with gsplat's design.

[antithing]

Thank you! Can you please comment here when the work is completed? Thank you again for this amazing project

[tommybotch]

Hello @ndming -- I was wondering if there were any updates on porting to gsplat? Please let me know! Thank you for your help :)

[ndming]

Hi @tommybotch,

Thanks a lot for your interest. I just pushed new commits with substantial updates. The benchmark on DTU for diffuse-only training now achieves an average CD of 0.47 (the original was 0.51), and the average runtime is 11.8 minutes on our RTX5090 32GB (the original was 30mins on a RTX4090 24GB). There's also a slight increase to the PSNR, sitting at 34.48 in the latest run:

{
  "20260814_180744": {
    "total_scenes": 15,
    "completed_scenes": 15,
    "psnr": 34.48018887837728,
    "ssim": 0.9441045165061951,
    "lpips": 0.10712230255206426,
    "time_mins": 11.75867637793223,
    "vram_gb": 0.4969651540120443,
    "chamfer": 0.47437724462269915
  }
}

To see the exact steps to reproduce the above result, please have look at the benchmarks/dtu.py script.

I moved most of the plane depth rendering logic to CUDA and also incorporated the fast NCC patch match module from RaDe-GS to the new commits. It's quite tricky because gsplat has many more features I need to keep an eye on to prevent the new geometric rendering from breaking them. At the moment, plane depth rendering cannot work with gsplat's packed mode or distributed training (yet).

The next step for this is to finish the material training path, but I'm afraid that would have to wait for a while because I'm having a new project lined up and currently I'm the sole maintainer of the repo. I appreciate your waiting and will try my best to have material training path settled.