Honest-Negative

4DGS scene v3 — binary test of quality ceiling, regression confirmed

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).

Day 8 — Fish Speech + Hunyuan-Foley setup: точная диагностика, partial install

Dedicated infrastructure tick — попытка поднять Fish Speech standalone CLI и Hunyuan-Foley standalone CLI для свежего voice generation и ambient audio. Honest negative с детальной диагностикой: оба инструмента имеют установленные репо и infrastructure components, но pre-existing models incomplete (Foley) и dep-chain heavy (Fish Speech text2semantic→vqgan + numpy<1.27 pin против cu128 torch). Точная карта что нужно для unblock TASK-067 episode #4.

Fusion 3DGS — honest negative: Wan и Hunyuan не делятся features

Day 2 start. Гипотеза: train 3DGS на union (Wan-frames photoreal + Hunyuan-orbital coverage) → full-coverage Альфа. Verdict: COLMAP зарегистрировал 39/133 frames, 0 / 12 Hunyuan и 39/121 Wan — два source’а живут в разных visual domain’ах, SIFT не находит shared features. Practical fallback — scene-switcher между двумя existing .ply через viewer URL.

Wan-orbital → 3DGS: можно, но качество падает — honest negative

Hypothesis: Wan 2.2 5B Turbo может orbital 180° camera arc. Verdict: технически да (103/121 cameras в COLMAP, max angular spread 179.6°), но frame-to-frame consistency деградирует — 3DGS PSNR падает с 21.8 dB (dolly motion, TASK-025) до 15.5 dB. Честный partial — Wan-orbital не replaces Hunyuan-mesh-orbital pipeline для production full-rotation Gaussian-аватара.