Wan2.2 TI2V-5B on a single RTX 4090: ~8.5 min per 5s clip, faces broke in 2 of 3 tests
NoEagle3604 · reddit · 2026-10-28
A hands-on benchmark of Wan2.2-TI2V-5B image-to-video on a single 24GB RTX 4090, with DGX Spark as comparison. Pipeline: stills generated/edited with Qwen-Image-2.1 (nf4, 40 steps, CFG off; 6 edit passes for the jellyfish since repeated edits add noise; earlier stills used as reference to keep character/style consistent), then animated with Wan2.2 (50 steps, guidance 5.0, 121 frames @24fps ≈5s), prompts describing motion and ending with "static camera". Numbers: 518s per clip on the 4090 with 24,008 MiB peak VRAM (nearly maxed out); DGX Spark took 1,338–1,420s at 30,135 MiB median peak. PYTORCHCUDAALLOCCONF=expandablesegments:True was required on the 4090 or VAE decode OOMed. Results: faces distorted in 2 of 3 clips (the clean jellyfish clip has the girl facing away, proving nothing); the train clip also overexposed. Caveats: stock settings — no Lightning/distill LoRA, no TeaCache, no extra quantization; run via model-compose (one YAML auto-building venv/weights/API; author discloses ties to that team) rather than ComfyUI; only 3 stills × 2 seeds, no comparison to Wan 14B/LTX/MiniMax H3. Full setup in MindrLabs/image-to-video-wan on GitHub; the author asks how others keep faces intact.
More from Infra
- China's STAR50 index plunges 40% in four months, wiping out $480B in AI stock value — Kyrannio · 2026-10-28
- Mixed NVIDIA+AMD multi-GPU LLM setup benchmarks slower than a single card — LMasterGame · 2026-10-28
- 8 open-source repos to self-host a lean AI agent stack instead of stacking $20/month SaaS — SaiTarunRonanki · 2026-10-28
- Corning Optical Q3: sales up 44% to $2.38B, Gen AI revenue doubles y/y — demian_ai · 2026-10-28
- AI-powered formal verification finds real Redis/Valkey bugs, incl. a 23-char freeze — LingmingZhang · 2026-10-28
- 112 TPS on a single RTX 3090: LlamAmpere v0.6 pushes 27B models with 400K context — Brief-Tap-6616 · 2026-10-28