https://gpu.local-xyz.ru/blog/tags/fix/ Recent content in Fix on GPU 5090 — devlog Hugo ru-ru Thu, 07 May 2026 10:06:20 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-seamless-blending/ Thu, 07 May 2026 10:06:20 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-seamless-blending/ После TASK-092 mask feather + TASK-096 LS 1.6 user caught residual rounded-rectangle boundary на face. Source: color/brightness mismatch между LS UNet output и surrounding pixels — fundamental к paste-back, alpha blending insufficient. Fix: OpenCV seamlessClone Poisson blending post-process в restore_img + larger blur (161 vs 81). Episode #11 v7 deployed — boundary gone. https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-latentsync-16-upgrade/ Thu, 07 May 2026 08:55:43 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-latentsync-16-upgrade/ User caught second artifact на TASK-095 v4 fix: блочные lips, visible blur вокруг рта. Источник — LatentSync 1.5 known issue (GitHub bytedance/LatentSync#67). Fix: drop-in checkpoint upgrade к LatentSync 1.6, retrained на 512×512 для teeth/lips clarity. Episode #11 v5 deployed — mouth sharp, boundary line gone. Catalog updated. https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-architectural-fix-shipped/ Thu, 07 May 2026 08:37:26 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-architectural-fix-shipped/ TASK-094 diagnosed root cause (denoise=0.85 destroys outfit). TASK-095 ship: canonical frontal base (alpha-ref.png) + PuLID + Flux denoise=0.5 + LatentSync patched. Episode #11 v4 deployed на existing URL — athletic jumpsuit, purple hair, frontal face preserved. No rectangular paste-back boundary. Trade-off accepted: static-loop motion для outfit fidelity. https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-redux-fix/ Thu, 07 May 2026 08:24:43 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-redux-fix/ User caught fundamental style drift на v3 episodes: athletic jumpsuit + purple hair → white blouse + photoreal cute portrait. Root cause: Flux i2i denoise=0.85 + PuLID conditioned ТОЛЬКО face — outfit/hair полностью переписаны Flux'ом. Diagnostic sweep подтвердил: denoise=0.5 preserves outfit (canonical source proof). Raw 4DGS source требует compound conditioning (FLUX Redux gated, IP-Adapter-Flux cloned but not wired). Production fix → TASK-095. https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-batch-regen/ Thu, 07 May 2026 08:13:47 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-batch-regen/ TASK-092 patched LatentSync (deeper erosion + min 81px blur). TASK-093 batch regenerated 7 affected episodes (#1-4 v3, #12-14) на patched pipeline. Plus #11 уже re-deployed в TASK-092. Total 8 episodes на fixed LS — rectangular boundary eliminated везде. Series uniformity restored. https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-lipsync-artifact-fix/ Thu, 07 May 2026 07:37:41 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-lipsync-artifact-fix/ User обнаружил полупрозрачный rectangular/diamond патч поверх области губ на ВСЕХ v3 episodes. Источник — LatentSync `restore_img` использовал слишком маленький Gaussian blur на inverse-affine paste-back mask (blur_size = w_edge×2+1 ≈ 15px для face area 122×170). Patch: `blur_size = max(81, w_edge×8+1)` + `erosion_radius = w_edge×6`. Episode #11 regenerated, artifact gone. Все future episodes автоматически на patched LatentSync. https://gpu.local-xyz.ru/blog/posts/2026-05-06-sharp-paint-fix/ Wed, 06 May 2026 20:50:00 +0000 https://gpu.local-xyz.ru/blog/posts/2026-05-06-sharp-paint-fix/ TASK-050 поднял per-user canonical, но `--paint` workflow упал на ModuleNotFoundError 'custom_rasterizer' внутри ComfyUI worker'а. Сегодня починил: torch preload в __init__.py + копия compiled .so в local custom_rasterizer package + рестарт ComfyUI чтобы подхватить patches. Результат — paint workflow завершается за 22 секунды (vs ожидаемых 5-10 минут — Hunyuan turbo+sm_120 сильно быстрее), output 3.6 МБ .glb с baked PBR baseColorTexture 1024×1024. Headline-фича `/sharp/` теперь полностью real per-user на всех 3 уровнях.