Wan 2.1 and 2.2: which model needs what
| Model | Weights (BF16) | Minimum VRAM (with offload) | Comfortable | Best for |
|---|---|---|---|---|
| Wan 2.1 T2V 1.3B | ~2.6 GB | 8 GB | 12 GB | 480p tests, fast iteration |
| Wan 2.2 TI2V 5B | ~10 GB | 12 GB | 24 GB | 720p on consumer cards, image-to-video |
| Wan 2.1 / 2.2 14B (T2V, I2V) | ~28 GB | 24 GB (FP8 + block swap) | 48–80 GB | Production quality, 720p, LoRAs |
| Wan 2.2 14B MoE (high + low noise) | ~56 GB total | 24 GB (FP8, swapped) | 80–141 GB | Best motion quality; two experts loaded in turn |
The text encoder (umT5, ~11 GB in BF16) and the VAE add to those numbers unless offloaded to CPU — which is exactly what ComfyUI's Wan workflows and Wan2GP do on 24 GB cards, at a speed cost.
Minutes and dollars per clip, card by card
Indicative times for a 5-second, 81-frame clip with the 14B model at default steps (community benchmarks, single clip, no batching). Cost = today's on-demand rate × minutes; interruptible is half.
| Card | Setup | 480p | 720p | $ / 480p clip | $ / 720p clip | Note |
|---|---|---|---|---|---|---|
| RTX 3090 24 GB · $0.104/hr | Wan 2.1 1.3B / Wan 2.2 5B only | 4 min | — | $0.007 | — | Too little VRAM for 14B without heavy offload |
| RTX 4090 24 GB · $0.262/hr | FP8 + offloading | 7 min | 14 min | $0.031 | $0.061 | The community default; 24 GB with block swapping |
| RTX 5090 32 GB · $0.318/hr | FP8, light offload | 4.5 min | 9 min | $0.024 | $0.048 | 32 GB and 1.79 TB/s — noticeably faster |
| L40S 48 GB · $0.466/hr | BF16, no offload at 480p | 5 min | 9.5 min | $0.039 | $0.074 | 48 GB; datacenter card for batch queues |
| RTX PRO 6000 WS 96 GB · $1.097/hr | BF16, no offload | 3.5 min | 7 min | $0.064 | $0.128 | 96 GB — 720p without any offload tricks |
| H100 PCIE 80 GB · $2.147/hr | BF16, no offload | 2.5 min | 5 min | $0.089 | $0.179 | 80 GB HBM; the throughput pick |
| H200 141 GB · $3.058/hr | BF16, big batches | 2.2 min | 4.2 min | $0.112 | $0.214 | 141 GB — longest clips, highest resolutions |
Read it as two tiers. Consumer cards make each clip cheap but slow; 80 GB+ cards make each clip fast and, because the clock runs shorter, often cheaper too once you count wall-clock. The H100 PCIE at $0.179 per 720p clip beats the RTX 4090 at $0.061 outright.
A batch night, budgeted
Sixty 720p clips for a storyboard, queued overnight on interruptible capacity:
| Card | Wall-clock | GPU cost (interruptible) | Model volume 200 GB | Total |
|---|---|---|---|---|
| RTX 4090 | 14.0 h | $1.83 | $16.00/mo | $17.83 |
| RTX 5090 | 9.0 h | $1.43 | $16.00/mo | $17.43 |
| H100 PCIE | 5.0 h | $5.37 | $16.00/mo | $21.37 |
| H200 | 4.2 h | $6.42 | $16.00/mo | $22.42 |
Each clip is an independent queue item, so an interruption re-runs one clip from the volume, not the night.
Setup: ComfyUI or Wan2GP
powergpu launch --gpu rtx-5090 --template comfyui --disk 80 --volume video-models:/workspace/ComfyUI/models
# ✓ instance i-77e0c9d3 running (29.1s) · $0.318/hr
# https://i-77e0c9d3.powergpu.io:8188 (ComfyUI — load the Wan 2.2 template workflow)ComfyUI carries the official Wan workflows, LoRA support and the video-helper nodes; Wan2GP is the simpler UI with the most aggressive memory optimisations for 12–24 GB cards. Download the 14B weights (20–60 GB) once onto the volume and they mount on every future instance in seconds.
What makes video cheap
- Prototype at 480p on a 5090, final-render at 720p on 80 GB — same workflow JSON, different card.
- Queue clips, not sessions. Per-clip jobs on interruptible capacity pay half and survive interruptions.
- Watch VRAM, not GPU utilisation. Video pipelines run out of memory long before they saturate compute; offloading is a speed tax you can buy out with a bigger card.
- FP8 weights on Ada/Blackwell halve memory with little visible loss; keep BF16 for the final pass if you can afford the card.
- Destroy the instance, keep the volume. Models stay warm at $0.08/GB/month while the GPU bills nothing.
Adjacent reading: the video generation playbook and RTX 5090 vs RTX 4090.
Put the numbers to work
Every price in this guide is our live rate — fixed, ≥30% under the market median, billed per second. Deploy the exact setup above from the console in about 30 seconds, paid in crypto, no card and no KYC.
