Read the official row
| VAE | T×H×W | Feature dim | Info rate | PSNR↑ | SSIM↑ | LPIPS↓ |
|---|---|---|---|---|---|---|
| SVD | 1×8×8 | 4 | 48 | 28.307 | 0.816 | 0.067 |
| Cosmos | 4×8×8 | 16 | 48 | 29.632 | 0.863 | 0.132 |
| Hunyuan | 4×8×8 | 16 | 48 | 33.298 | 0.922 | 0.023 |
| CogVideoX1.5 | 4×8×8 | 16 | 48 | 32.955 | 0.922 | 0.040 |
| Wan2.1-VAE | 4×8×8 | 16 | 48 | 32.222 | 0.911 | 0.026 |
| Wan2.2-VAE | 4×16×16 | 48 | 64 | 33.223 | 0.922 | 0.022 |
The 2.2 VAE is more compressed spatially (16 vs 8) with a wider latent channel (48 vs 16). Reconstruction still matches or beats the 4×8×8 pack on SSIM and LPIPS in this table. That is the claim to repeat — not a homemade “64× better” slogan.
Patchify on TI2V-5B
README: with an extra patchification layer, TI2V-5B’s total compression reaches 4×32×32. Combined with a 5B dense DiT, a 5-second 720P clip is quoted as under 9 minutes on a single consumer GPU without extra compiler tricks.
A14B still uses the earlier VAE family
The high-compression VAE is called out for the 5B hybrid. The 27B MoE T2V/I2V models keep a compatible latent with the 14B-scale stack (Wan2.1-style 4×8×8). Mixing VAE weights across those families will decode garbage.
FAQ
- Why 4 in time?
- Causal 3D VAE: four frames (plus a boundary) collapse to one temporal latent. That is why video lengths are 4n+1 frames.
- Can I decode 2.2 latents with a 2.1 VAE?
- No. Channel count and spatial stride differ (48 vs 16, 16 vs 8).
Primary sources
Checked against Wan-Video/Wan2.2 README · Aug 28, 2026
