Automatic Per-Scene Hyperparameter Optimization

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

Challenge. Different scenes need different hyperparameters. Create an automatic optimization system (RL-based or other approaches) that discovers the best settings per scene during training without manual tuning and scales to new scenes.

💸 Prize Pool: $2,430

• AukiNetwork: $1,000
• Vincent Woo: $500
• MrNeRF: $300
• Yehe Liu: $280
• Kenneth Lynne: $200
• Florian Hahlbohm: $100
• Mazeyar Moeini: $50

---

🧾 Rules

1. Fork from the bounty_004 branch.

2. Automatic per-scene optimization must run during training (no manual per-scene tuning). Your solution should be able to tune parameters such as:

   • Learning rates (e.g., position, scale, rotation, opacity, SH)
   • Densification thresholds and intervals
   • Number of iterations (treated as a hyperparameter)
   • Number of Gaussians (see densification below)
   • Any other parameters that affect quality or convergence

   => For instance, changing only the number of iterations as main driver does not qualify! It should be clear that the system tries to figure out an optimal configuration over a none trivial subset of parameters.

3. Densification strategy: MCMC. The approach must use MCMC-based densification; the number of Gaussians is a tunable hyperparameter.

4. Target improvement: Achieve an average +0.15 dB PSNR improvement over baseline on the MipNeRF360 dataset (and further scenes). See benchmarks below.

Quality Metrics Summary

Scene	Iteration	PSNR	SSIM	LPIPS	Num Gaussians
garden	30000	27.8539	0.862883	0.107563	4,937,304
bicycle	30000	25.7642	0.785712	0.188216	5,684,053
stump	30000	26.9556	0.810134	0.213571	4,647,623
bonsai	30000	32.5415	0.953107	0.246894	1,120,498
counter	30000	29.2535	0.929803	0.244577	886,049
kitchen	30000	31.5344	0.935762	0.154214	1,129,135
room	30000	32.0918	0.936561	0.272640	1,199,942
mean	30000	29.4278	0.887709	0.203954	2,800,657

5. Dataset: Use the official MipNeRF360 dataset: 📦 360_v2.zip → http://storage.googleapis.com/gresearch/refraw360/360_v2.zip

6. Generalization test: Your method will be evaluated on one or two undisclosed scenes in addition to MipNeRF360. Approaches must scale beyond MipNeRF360 without scene-specific hacks.

7. Licensing: You may only use GPLv3-compatible dependencies (e.g., MIT, Apache-2.0, BSD, GPLv3, etc.). List all third-party deps in your README with licenses.

8. Implementation language:

   • Preferred: C++ implementation committed to bounty_004 branch.
   • Alternative: Python implementations are accepted, but total award is reduced by 20%.

9. Reproducibility: Provide a single command (or script) to reproduce your results per scene, including fixed seeds where relevant.

Important: Results are not only measured via PSNR - they will also be visually inspected for artifacts (ghosting, floaters, oversmoothing, texture shimmering, etc.).

---

💡 Approach Ideas (non-exhaustive)

• Reinforcement learning (e.g., RLGS-style controllers)
• Bayesian optimization (e.g., model-based HPO)
• Meta-learning / per-scene adaptation
• Gradient-based hyperparameter optimization
• Population-based training / schedule-free optimizers
• Your novel approach!

Helpful starting points:

• Schedule-Free Optimizers (Facebook Research): https://github.com/facebookresearch/schedule_free
• SMAC (AutoML / Bayesian optimization): https://github.com/automl/SMAC3
• PyTorch Lightning Tuner: https://lightning.ai/docs/pytorch/stable/api/lightning.pytorch.tuner.tuning.Tuner.html
• (RLGS: Reinforcement Learning-Based Adaptive Hyperparameter Tuning for Gaussian Splatting): https://arxiv.org/pdf/2508.04078

---

📦 Submission Requirements

Your PR must include:

1. Working implementation on bounty_004 branch (C++ preferred; Python accepted with 20% award reduction).
2. Automation entrypoint (script/CLI) to run the optimizer per scene.
3. Results table covering all MipNeRF360 scenes: PSNR, SSIM, LPIPS, training time, #iterations, #Gaussians, and key hyperparameters found.
4. Visuals: A short gallery (or links) with representative renders for at least 3 scenes highlighting improvements.
5. Technical brief describing the optimization strategy, search space, controller/optimizer, and any constraints/priors.
6. License & dependencies section listing all third-party libraries and their licenses (must be GPLv3-compatible).

---

🗓️ Deadline

October 12, 2025 at 11:59 PM PST

---

💰 Prize Distribution

• 70% to the winning PR
• 30% shared among strong qualifying submissions that meet all requirements

Organizers reserve the right to adjust awards for ties or extraordinary contributions.

---

📫 Questions

Open an issue or discuss in the designated thread on discord.

Good luck, and happy hacking! 🚀