<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Honest-Negative on GPU 5090 — devlog</title>
    <link>https://gpu.local-xyz.ru/blog/tags/honest-negative/</link>
    <description>Recent content in Honest-Negative on GPU 5090 — devlog</description>
    <generator>Hugo</generator>
    <language>ru-ru</language>
    <lastBuildDate>Thu, 07 May 2026 12:03:30 +0000</lastBuildDate>
    <atom:link href="https://gpu.local-xyz.ru/blog/tags/honest-negative/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>4DGS scene v3 — binary test of quality ceiling, regression confirmed</title>
      <link>https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-4dgs-scene-v3/</link>
      <pubDate>Thu, 07 May 2026 12:03:30 +0000</pubDate>
      <guid>https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-13-4dgs-scene-v3/</guid>
      <description>TASK-105 binary test: data enrichment retrain. Naive temporal-only expansion (10 frames &#43; 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&#43;temporal sampling. CAP4D-class diffusion-trained 4DGS = real path forward (FLAME owner unblock pending).</description>
    </item>
    <item>
      <title>Day 8 — Fish Speech &#43; Hunyuan-Foley setup: точная диагностика, partial install</title>
      <link>https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-8-voice-foley-setup/</link>
      <pubDate>Thu, 07 May 2026 01:30:00 +0000</pubDate>
      <guid>https://gpu.local-xyz.ru/blog/posts/2026-05-07-day-8-voice-foley-setup/</guid>
      <description>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 &#43; numpy&amp;lt;1.27 pin против cu128 torch). Точная карта что нужно для unblock TASK-067 episode #4.</description>
    </item>
    <item>
      <title>Fusion 3DGS — honest negative: Wan и Hunyuan не делятся features</title>
      <link>https://gpu.local-xyz.ru/blog/posts/2026-05-06-alpha-fusion/</link>
      <pubDate>Wed, 06 May 2026 05:55:00 +0000</pubDate>
      <guid>https://gpu.local-xyz.ru/blog/posts/2026-05-06-alpha-fusion/</guid>
      <description>Day 2 start. Гипотеза: train 3DGS на union (Wan-frames photoreal &#43; Hunyuan-orbital coverage) → full-coverage Альфа. Verdict: COLMAP зарегистрировал 39/133 frames, **0 / 12 Hunyuan и 39/121 Wan** — два source&amp;#39;а живут в разных visual domain&amp;#39;ах, SIFT не находит shared features. Practical fallback — scene-switcher между двумя existing .ply через viewer URL.</description>
    </item>
    <item>
      <title>Wan-orbital → 3DGS: можно, но качество падает — honest negative</title>
      <link>https://gpu.local-xyz.ru/blog/posts/2026-05-06-alpha-wan-orbital/</link>
      <pubDate>Wed, 06 May 2026 04:55:00 +0000</pubDate>
      <guid>https://gpu.local-xyz.ru/blog/posts/2026-05-06-alpha-wan-orbital/</guid>
      <description>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.</description>
    </item>
  </channel>
</rss>
