Forcing Wan to Generate a 60-Second Single Shot on a 16GB GPU: It Finished, Barely
Wonderful_Sample6291 · reddit · 2026-09-19
A Reddit user deliberately removed the usual safety rails (generating 2–5 second clips and stitching them) and forced the open-source video model Wan to produce one continuous 60-second shot on an RTX 5060 Ti 16GB.
- Specs: 768×448, 24 FPS, 1,441 frames, 30/30 sampling steps, runtime 2h 38m 20s, peak VRAM 15.48GB, VAE decode and MP4 save both completed.
- Quality: The first few seconds looked fine, then it collapsed into some kind of "brown nightmare tunnel" and maintained that for nearly the entire remaining minute.
- Verdict: compute success and 60-second generation: yes; temporal coherence: no; prompt adherence: dead; mysterious brown-dimension consistency: surprisingly excellent.
The result is a vivid demonstration of current open-source video models' limits on long-horizon temporal coherence.
More from Fun
- Satire: The Comfort of 'It's Just a Classifier' — How AI Skeptics Relive 2022 — cephaloform · 2026-09-19
- LLMs aren't stochastic parrots anymore, but AI haters' laments are statistically predictable — Merzmensch · 2026-09-19
- Asking a $100/Month User If They Ever Buy Credits: Tibo Called Out of Touch — SeaTurtleManOG · 2026-09-19
- Grady Booch jokingly proposes 'Existential AI Anxiety Disorder' for the DSM — Grady_Booch · 2026-09-19
- Chinese AI community mocks the Codex quota reset drama: "no tokens, no godfather" — dotey · 2026-09-19
- Polymarket bets on an Anthropic wet-lab pathogen leak: 7% odds by end of 2026 — Polymarket · 2026-09-19