Derin, Mehmet OguzChristie, MarcHan, Ping-HsuanLin, Shih-SyunPietroni, NicoSchneider, TeseoTsai, Hsin-RueyWang, Yu-ShuenZhang, Eugene2025-10-072025-10-072025978-3-03868-295-0https://doi.org/10.2312/pg.20251307https://diglib.eg.org/handle/10.2312/pg20251307Path-traced sequences at one sample per pixel (1-spp) are attractive for interactive previews but remain severely noisy, particularly under caustics, indirect lighting, and volumetric media. We present SampleMono, a novel approach that performs multi-frame spatiotemporal extrapolation of low-resolution and low-sample Monte Carlo sequences without requiring auxiliary buffers or scene-specific information. We transfer and prune a pre-trained video generation backbone and fine-tune it on SampleMono GYM, a synthetic Monte Carlo dataset, to generate four clean high-resolution frames from a longer window of noisy inputs, thereby decoupling render and presentation timelines. Our experiments demonstrate that by combining a frozen VAE encoder-decoder and training of a video generation model pruned to two transformer layers, our pipeline can both provide spatial upsampling and temporal extrapolation to a long sequence of 16 RGB frames of 50 milliseconds time delta between frames at 256×144 resolution with severe Monte Carlo noise, generating subsequent four RGB frames of 12.5 milliseconds time delta between frames at 1280×720 resolution with substantially reduced noise at varying quality while fitting VRAM budget of 5GB. We plan to publish the code for data GYM, model pruning, pipeline training, and rendering.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Transfer learning; Ray tracing; TrackingComputing methodologies → Transfer learningRay tracingTracking10.2312/pg.202513072 pages