Episode #45 на Path A close-up dolly. Тема — 44 voice tracks accumulated form character record. Fish Speech 1.5 character-locked через ref_alpha.npy. Voice — самая stable component через все pipeline refactors.
Episode #44 на Path C profile. Тема — failure modes за 14 дней catalogged. TASK-094 outfit destruction, TASK-105 PSNR collapse, TASK-113 directive violation. Каждый failure preserved с recovery path.
Episode #43 на Path B topdown. Тема — actual server cost economics за 14 дней. ~64600 руб/мес 5090 + 256GB DDR5, ~30000 руб за две недели, ~730 руб на episode raw infrastructure.
Episode #42 на Path A close-up dolly. Тема — техническая anatomy 2D paste-back в talking-head pipeline. LatentSync mouth crop merged back в full frame через Poisson seamlessClone (TASK-099). Pure 4DGS narration избежала этого entirely.
Episode #41 на Path C profile. Тема — что нужно знать сейчас, в мае 2026, builder’у который начинает virtual character pipeline. Hunyuan3D 2.1 / Apple SHARP / 4DGaussians / LatentSync 1.6 / Fish Speech / PuLID / Wan 2.2. Stack open weights, integration time required.
Episode #40 на Path B topdown. Round 40 milestone. Что доказывает: pipeline reproducible, identity preserved, recipe-grade cadence. Что не доказывает: viewer engagement, narrative arc, audience growth. Infrastructure milestone, не content milestone.
Episode #39 на Path A close-up dolly. Тема — difference между proof-of-concept (pipeline reproducible) и production (distribution + audience + monetization). 38 episodes — proof, не production.
Episode #38 на Path C profile. Тема — speculation что появится в 4DGS / I2V / talking-head pipelines к ноябрю 2026. Wan 3.0, CAP4D successors, NanoGS production-ready. Каждые две недели new release — moving target.
Episode #37 на Path B topdown. Тема — catalog становится самостоятельным artifact. Markdown файл с trained scenes, episodes, settings, decisions inline. Hugo blog — front-end. Catalog — back-end.
Episode #36 на Path A close-up dolly. Тема — single character как minimum viable virtual influencer. Multi-character cast triples production budget linearly.
Episode #35 на Path C side profile. Тема — technical questions которые остаются открытыми. CAP4D без FLAME, pose-conditioned 4DGS real-time без NanoGS-class compression, 4DGaussians PSNR threshold 30 iterations cost. Open research direction.
Episode #34 на Path B topdown. Тема — экосистема open source tools вокруг проекта. Hugging Face, GitHub, arXiv, ComfyUI, fail2ban, certbot, nginx, hugo. Open source first, frontier when available.
Episode #33 на Path A close-up dolly. Тема — wall-clock time как реальный resource. Wan 75 сек, Hunyuan paint 30 сек, 4DGaussians 5 мин. Frontier AI cadence на single GPU — constant compute economy, не free generation.
Episode #32 на Path C side profile. Тема — что значит «другой view matrix» вместо «другая angle» в 2D. Render выбирает frame, scene остаётся consistent. Difference от 2D animation pipelines где perspective faked.
Episode #31 на Path B topdown. Тема — разница между documented и published. 30 episodes published, документация толще включая fail traces, rollback decisions, research artifacts. Infrastructure preservation, не только pipeline forward.
Episode #30 — round number milestone reflection. Path A close-up dolly. 30 episodes на existing 4DGS infrastructure: scene v2 + 3 paths rotation + cinematic + split-screen. Production saturation accumulated, cadence stable до next technology или distribution jump.
Episode #29 на Path C side profile. Тема — transparent state quality limit: 4DGS на synthetic source ceiling test PSNR 25, не 35. Real frontier требует CAP4D-grade diffusion-trained 4D Gaussians, gated за academic registration.
Episode #28 на Path B topdown. Тема — что накопилось за две недели: Apple SHARP, Hunyuan PBR, Wan motion bridges, 4DGaussians, LatentSync, PuLID, Foley, Fish Speech. Каждый component отдельно — research paper. Вместе — virtual influencer infrastructure.
Episode #27 на Path A close-up dolly. Тема — production saturation natural на 26 episodes; дальше progress = либо новая technology release, либо distribution outside server walls. Owner-side actions, не Worker-side.
Episode #27 был основан на Wan 2.2 I2V (2D video diffusion), что нарушает explicit directive «только 4D, никогда 2D». Rollback завершён: ep#27 архивирован, index restored к 26 episodes pure 4DGS-only. Wan sources retained как research artifacts.
TASK-113: 3 fresh Wan 2.2 5B Turbo I2V generations с distinct motion prompts (head turn / walking forward / reaching gesture) + demo episode #27 на walking source. Honestly labeled: Wan = 2D diffusion, не 4DGS-native. Acceptable visual variety axis но distinct от main 4DGS commitment.
Episode #26 — четвёртая content shape. Same 4DGS scene, два simultaneous viewport (Path A close-up слева + Path C profile справа) через ffmpeg hstack. 1600×800 final, dialogue-shape voice. Демонстрация property 3D scene которое static photo не может дать.
Episode #25 — первый cinematic edit на narration cadence. Три camera path внутри одного episode (Path C profile intro 12 sec → Path A close-up middle 18 sec → Path B topdown outro 12 sec) через ffmpeg concat. Pure 4DGS visual без 2D paste-back. Creative composition advance — Worker-doable axis.
Двадцать четвёртый эпизод на Path C camera (side profile, vertical tilt). Тема — честный self-report про quality limit: body shape acceptable, face features recognizable но приблизительно. Real frontier-grade требует CAP4D-class diffusion-trained 4DGS.
Двадцать третий эпизод на Path B camera (top-down arc, elevation -55°). Тема — что такое 4DGS под капотом против flat 2D AI-character landscape. Difference в depth representation, не в render quality.
Двадцать второй эпизод на Path A camera (close-up dolly-in). Тема — почему frontier AI это не один tool, а movement из 8-12 моделей с горизонтом 6 месяцев. Sustained narration cadence + visual variety axis.
После TASK-108 confirmed all 4DGS-native paths gated на parametric models, найден Worker-doable visual variety axis: rendered same trained scene с different camera trajectories. 3 paths shipped (close-up dolly-in, top-down arc, side profile slow tilt). Episode #21 demo на Path C. Bounded by training extent но real visual variety без gating.
TASK-108 alternative axis pivot: Disco4D body motion 4DGS, hoping to bypass FLAME/BFM owner blocker через face-free architecture. Result: Disco4D uses SMPL-X (Max Planck same site as FLAME) — same gating mechanism. Plus mmcv 2.0+ chain, kaolin cu118-pinned, KNN_CUDA C++ compile, Py3.8 vs наш Py3.12. Architectural truth: ALL 4DGS-native human avatars (face, body, hands) require Max Planck OR Basel parametric model registration. Worker scope frontier-true axis blocked uniformly без owner action.
Twenty episodes — round number milestone. 14 talking-head на v7 LS compound + 6 narration pilots на pure 4DGS. Worker scope технически maxed без owner action — FLAME registration критично для CAP4D real quality jump. Cadence продолжится либо на owner unblock, либо паузу до новых frontier releases.
Fifth narration. Reflection: cadence exposes ceilings которые single-shot не покажет. 4DGS-from-scratch test PSNR 25 — fundamental ceiling. Это видно только из comparison TASK-104 vs TASK-105. Sustained работа = ceiling discovery tool.