Ceiling on GPU 5090 — devlog https://gpu.local-xyz.ru/blog/tags/ceiling/ Recent content in Ceiling on GPU 5090 — devlog Hugo ru-ru Thu, 07 May 2026 12:10:06 +0000 Episode #16 — 4DGS quality reality, gritty https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-alpha-episode-16/ Thu, 07 May 2026 12:10:06 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-alpha-episode-16/ Second narration episode — gritty technical observation о 4DGS-from-scratch ceiling после TASK-105 binary test. 12 spatial + 22 Wan train = test PSNR 25; добавил 121 frame Wan — упал к 13. Diversity > volume. Real 4D avatar требует CAP4D-class diffusion-trained — pending FLAME unblock. Pure 4DGS narration на v2 source, 35 sec, ~15 sec compute. 4DGS scene v3 — binary test of quality ceiling, regression confirmed https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-4dgs-scene-v3/ Thu, 07 May 2026 12:03:30 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-4dgs-scene-v3/ TASK-105 binary test: data enrichment retrain. Naive temporal-only expansion (10 frames + 121 Wan motion frames at fixed camera) caused REGRESSION — train PSNR 30.5 (locked в overfit), test PSNR 13.16 (vs v2 25.4). v3 rolled back, v2 restored as production. Ceiling confirmed: 4DGS-from-scratch на synthetic data cannot generalize beyond training distribution без diverse spatial+temporal sampling. CAP4D-class diffusion-trained 4DGS = real path forward (FLAME owner unblock pending).