🔔 Bounty #004 - Automatic Per-Scene Hyperparameter Optimization Build a system for gplv3 LichtFeld Studio that automatically finds optimal hyperparams per scene during training. Must use MCMC densification (Num Gaussians tunable). Target: +0.15 dB PSNR over MipNeRF360 baseline. Prize pool: $2,430. Details and link in the thread below!
🧾 Core Rules (Brief): • Execute during training without manual per-scene tuning. • Adjust learning rates for position, scale, rotation, opacity, and Spherical Harmonics. • Modify densification thresholds/intervals, number of iterations, number of Gaussians, and other quality/convergence parameters. • Fork from the bounty_004 branch. • Ensure runs are reproducible.
💡 Approaches you can use: • RL controllers (RLGS / policy to adjust schedules) • Bayesian optimization / SMAC (model-based HPO) • Meta-learning / per-scene adaptation (fast fine-tune) • Gradient-based hypergrad methods (learnable LR schedules) • Population-Based Training / schedule-free optimizers • Any novel combo — but MCMC densification + tunable Gaussians is required.
📦 Submit: • PR to bounty_004 with a runnable entrypoint • Results table (all MipNeRF360 scenes) • Visuals and tech brief • Dependencies + GPLv3-compatible licenses C++ preferred (Python = 20% award reduction). Deadline: Oct 12, 2025, 11:59 PM PST (Oct 12, 2025, 11:59 AM PST). Good luck!
Sponsors of total prize $2,430: @Auki $1000 @fulligin $500 @janusch_patas $300 @YeheLiu $280 @kennethlynne $200 @fhahlbohm $100 @mazy1998 $50 github issue: github.com/MrNeRF/LichtFe… Discord: discord.gg/NqwTqVYVmj