Automatic Per-Scene Hyperparameter Optimization

9/14/2025 - 10/12/2025

Build a system that automatically finds optimal hyperparameters for each scene.

Combine both densification approaches

9/9/2025 - open

After bounty #2 there were two successful submissions. This bounty is more a direct programming task to combine the both submissions.

Make densification obsolete

8/10/2025 - 9/7/2025

This bounty aims to improve training quality without using densification by providing a stronger initialization.

Speed up training

6/23/2025 - 8/2/2025

This bounty challenges you to speed up training by at least 100% (halfing the runtime), without compromising results.

Winner: Florian Hahlbohm

Speed up training

Help us cut training time in half and earn $1500!
This bounty challenges you to speed up training by at least 100% (halfing the runtime), without compromising results.

Update: https://github.com/vincentwoo adds another $300 => $600 in total! Update: https://github.com/mazy1998 adds another $300 => $900 in total! Update: https://github.com/toshas adds $200 => $1100 in total! Update: https://x.com/ChrisAtKIRI adds $200 => $1300 in total! Update: https://github.com/julien-blanchon adds $100 => $1400 in total! Update: Drew Moffitt adds $100 => $1500

🔁 The winning implementation will be merged, and the repository will switch to a GPLv3 license.

🧾 Rules

Fork the repository, including all branches, starting from:
👉 bounty_001 branch
Use this script to benchmark performance:
🕒 timing_mipnerf360.sh
Apply your speed-up and submit a pull request.
Do not degrade quality – final metrics (e.g., PSNR/SSIM) must remain consistent with the baseline.
Target a minimum of 100% speed-up compared to the current implementation.
Use the official MipNeRF360 dataset:
📦 360_v2.zip
Deadline: August 2, 2025 at 11:59 PM PST (midnight) ⏳ All results will be reviewed and the winner announced shortly after.
If multiple entries achieve similar speed-ups:
- The cleanest and earliest implementation will win.
- Final decision rests with the repo maintainer.
Your pull request must include a clear summary of:
- What you optimized
- How you achieved the speed-up
You must not use any code that is licensed under a non-permissive license
By participating, you agree to release your submission under the GPLv3 license.
You must not alter the data loading to pull all data on gpu. Datasets with thousands of images must still work!
Loss logging and saving must still work. You must not strip the functionality. It must be possible to merge the final pr directly without adding back losses, savings, etc.

🏆 Prize

💸 $1500 to the author of the winning pull request.
🧠 Bonus: Your name (or alias) will be featured in the repository’s README.

📣 Discussion

Join our Discord server to discuss ideas, ask questions, or get help:
👉 https://discord.gg/6FaYg29MN7

Good luck, and happy hacking! 🚀

Benchmark (RTX 4090)

Scene	Time
garden	6m 18s
bicycle	5m 44s
stump	5m 44s
bonsai	7m 43s
counter	8m 37s
kitchen	8m 15s
room	7m 29s
Total	49m 50s

Scene	Iteration	PSNR	SSIM	LPIPS	Num Gaussians
garden	30000	27.174416	0.857002	0.157627	1000000
bicycle	30000	25.398046	0.777291	0.255703	1000000
stump	30000	26.797558	0.794485	0.258573	1000000
bonsai	30000	32.632896	0.948878	0.248194	1000000
counter	30000	29.357792	0.917621	0.242679	1000000
kitchen	30000	31.866880	0.934161	0.155137	1000000
room	30000	32.075516	0.930377	0.276833	1000000
mean	30000	29.329015	0.879974	0.227821	1000000