Wan-AI/Wan2.2-I2V-A14B · image-to-video
| Condition · Value | |
|---|---|
| GPU | RTX 3090 24GB |
| Power cap (W) | 300 |
| Resolution | 1280x720 |
| Frames | 49 |
| Steps | 30 |
| CFG | 4.0 |
| Shift | 5.0 |
| Seed | 42 |
| Precision profile | fp8 weight offload, PERSIST=0 (active expert VRAM-resident), FAST_OFFLOAD=0 (no pinned RAM -> wedge-safe) |
| Metric · Value | |
|---|---|
| Wall time (s) | 2596.4 |
| Denoise time (s) | 2262.9 |
| Per-step time (s) | 75.4 |
| Peak host RAM (GB) | 80.7 |
| Repeat count | 2 |
| Render stability | STABLE — visual QA: commercial-grade sharp start-to-finish (strawberry). Laplacian sharp_ratio unreliable for i2v (flags subject-exits-frame as degradation). |
COMPLETED, 2 runs (supersedes the earlier 848x480/40-step attempt that did not finish). Wan2.2-I2V-A14B (14B MoE, high+low noise DiT) renders a 1280x720/49f/30-step image-to-video on a single 24GB RTX 3090 in 2741s (~46min) at the STABLE GPU1=300W cap: 80.7GB peak host RAM, quality STABLE (sharp_ratio 1.083, motion 1.897), 8.59MB mp4. Three defenses made it complete where it previously wedged/crashed: (1) WSL disk expanded 47GB->879GB free (diskpart expand vdisk + resize2fs) so mmap model-load stops thrashing; (2) .wslconfig memory=190GB (< physical 251GB) so the VM can't starve Windows; (3) WAN_FAST_OFFLOAD=0 avoids pinned-RAM. Power: 420W on GPU1 crashes the box on a sustained i2v denoise (nvlddmkm 14/153 -> Kernel-Power 41, ~4min in), so 300W is the sustainable render condition. peak VRAM needs a Windows-side re-measurement. Evidence: veizik_demo/bench/i2v_strawberry.log.
Run the same fixed configuration on your card and open a result PR — see Run it yourself. Raw run manifests are attached to the internal evidence record and are released alongside external reproduction tooling.