Wan 2.2 on one RTX 5090: frame count doesn't touch VRAM, but resolution drops it by 10GB
Realistic-Fennel-190 · reddit · 2026-09-15
The author ran seven controlled experiments with the official ComfyUI Wan 2.2 5B template on a single RTX 5090 (121-frame, 5s, 24fps baseline, cfg 5.0, fixed seed) and the data overturned a common assumption.
Key findings
- Frame count doesn't move VRAM peak: 61/121/181 frames all peaked at exactly 25,579 MiB—a 2.5s clip costs the same memory as a 7.5s one
- Resolution dominates: dropping from 1280x704 to 832x480 (44% of pixels) cut peak VRAM to 19,083 MiB, a 10GB saving
- So when you hit OOM, shortening the video won't help; lower the resolution instead
Timing and cost (rental at $0.46/hr)
- Time ≈ 24.3s + 6.54s × steps; the 24s fixed cost (text encode + VAE decode + file write) means going below 4 steps buys almost nothing
- 1280x704/121f/20 steps: 155.3s, $0.020; at 480p it drops to $0.007
- Sampling time scales at roughly token count^1.85 for both resolution and frame reductions—matching attention theory—though the law breaks from 121 to 181 frames
Other notes: the anomalous 29,195 MiB run was likely the 6.3GB text encoder loading for a fresh prompt (cached runs settle at 25,579); and a wrong template silently produced a still image instead of video (watch for SD3 latent nodes, Save Image endings, and a CLIP loader set to lumina2). Full parameters, seed, and prompt are in the post for reproducibility.
More from Infra
- Tuning SGLang on a single 5090 for Qwen3.8-27B: 100 tok/s but only 82k context — ni1by2thetrue · 2026-09-15
- Four dev boards hooked to the internet: test AI-written firmware on real silicon via HTTPS — SelfishlyWandering · 2026-09-15
- Grouped Value Attention shrinks KV cache by reconstructing keys on demand — Vishesh Tripathi · 2026-09-15
- jinfer brings native AI inference to the JVM, matching llama.cpp on CPU with zero Python — mukel90 · 2026-09-15
- Stateless LLM failover preserves ~0% context; ContinuityBench proxy hits 99.20% CPR — its_vayishu · 2026-09-15
- 500 trillion tokens in a week: chart points to surging LLM usage — gajesh · 2026-09-15