training status loading
v6.1 Pico
waiting for metrics.json
OpenSuperSampling
v6.1 Pico is training right now
OpenSuperSampling is training a unified game reconstruction pipeline for super-resolution and frame extrapolation from one temporal model.
v5-pixel-temporal: measured
25.703 dB
↑ higher is better
v5-pixel-temporal, LPIPS 0.1666, temporal ratio 0.337x
LPIPS ↓ lower is better / temporal ratio ↓ lower is better (smaller frame-to-frame delta)
GPU usage
loading GPU status
Held-out PSNR / LPIPS by version
OSS measured points against conservative DLSS / FSR / XeSS reference lines.
PSNR — higher is better
LPIPS — lower is better
Reference lines are from competitor-published or third-party-benchmarked values on heterogeneous content (Cyberpunk 2077, Death Stranding, Hitman 3, etc.) — not direct head-to-head on the TartanAir oldtown batch OSS uses. They establish ROM order-of-magnitude context, not apples-to-apples comparison. OSS aims to publish head-to-head numbers on game-engine footage as the v6 cross-game-engine eval matures. v4 has a trajectory plus a distribution-shift cliff : the solid v4 line is the SRGD held-out trajectory across training, while the hollow marker at 300K is the same model family on the TartanAir oldtown batch v5+v6.1 use — the "rainbow mess" failure mode that motivated the v5+v6 architectural redesign. v6 is built to NOT distribution-shift like this; v6.1 cross-game-engine eval (UE5 / Unity / Source 2) is the ongoing test of whether that architectural design actually generalizes.
Sources: bicubic is measured on the OSS TartanAir oldtown batch; FSR references use AMD GPUOpen/FidelityFX documentation plus Digital Foundry/Insider Gaming comparisons; XeSS references use Intel XeSS paper and game-benchmark coverage; DLSS references use NVIDIA DLSS publications plus Digital Foundry/TechPowerUp comparisons. Values are favorable competitor-side reference anchors, not OSS-measured head-to-head scores.
Inspect training lineage
Open a run to compare training, measurement, and superseded references.
loading dashboard data...
Read canonical v6 design
The canonical v6 design is the source of truth for the persistent Gaussian canvas, covariance-resampled rasterizer, cross-attention fusion path, and OSS-FX frame extrapolation plan.
Open canonical memo