Codex built and runs a local anime pipeline: SD + LoRA → Wan 2.2 on RTX 5060 Ti
Wonderful_Sample6291 · reddit · 2026-09-07
A Redditor gave Codex one objective — build a local pipeline producing a short anime sequence on an RTX 5060 Ti 16GB — then only reviewed outputs and adjusted strategy.
Pipeline: Codex (production supervisor) → ComfyUI → Stable Diffusion + character LoRA → image QC → Wan 2.2 clips → QC/retry → editing.
Key lessons:
- Naively stitching 5s Wan clips failed on continuity: face, clothes, pose and background drifted every generation.
- New division of labor: SD handles quantity and controlled source images; Wan handles motion (mostly 2-3s shots, 4-5s only for simple motion); editing hides bad transitions.
- For hard transitions (walk → run → takeoff → flight → landing), the pipeline pre-generates SD reference images (closeups, poses, rear/distant shots); if continuity holds, inherit the previous frame; if broken, cut camera or regenerate.
- Quantity alone yields '1,000 versions of the same character'; the fix is a master character reference + fixed LoRA/settings + low-variation generation + strict QC.
- 16GB VRAM is treated as a shared resource with sequential stage loading: Wan at 480×704/24fps takes 111s for 3s and 180-200s for 5s of video, nearly saturating VRAM.
Current output is a 30s sequence, visibly imperfect but a reproducible proof of concept for agent-driven video production.
More from coding & agent
- Mechanize CTO Details Agent-First Work: ~1 Week per Task, Running Multiple Agents in Parallel — himanshustwts · 2026-09-07
- Is 'Plan Mode' Obsolete? Devs Debate Whether AI Coding Planners Still Earn Their Keep — JeremyNguyenPhD · 2026-09-07
- NUS, Princeton, Stanford & Oxford unveil Recuris: recursive self-improving agent memory without training — jiqizhixin · 2026-09-07
- Model prompts itself better than I do, so I built an orchestrator for my orchestrator: the ambassador — cephaloform · 2026-09-07
- mobilecode: An Open Source OpenCode Fork for Building iOS and Android Projects — Saboo_Shubham_ · 2026-09-07
- GitHub Launches HydraFusion: Multi-Model Orchestration Matches Opus 5 Quality at Lower Cost — unixterminal · 2026-09-07