Why DiT instead of U-Net
Classic image diffusion used convolutional U-Nets. DiT (Peebles & Xie) replaces the backbone with transformer blocks on patched latents. Video extends the patch grid into time. Wan follows that line (shared with many 2024–2025 video models) so attention can model long-range motion and camera moves.
Flow matching, not epsilon-prediction only
The official code uses flow-matching / rectified-flow style solvers (UniPC, Flow DPM). The network predicts a velocity field from noise to data rather than only ε. Sampling step counts in the README (and ComfyUI templates) assume that scheduler family.
Conditioning
- Text: UMT5-XXL embeddings, cross-attention in each block.
- Image (I2V/FLF2V): CLIP-vision + VAE tokens, extra MLP proj (FLF2V adds a first/last positional embedding of 257×2 tokens in Wan2.1).
- Audio (S2V): a dedicated injection path (see the official S2V audio figure).
- Timestep / SNR: AdaLN; on A14B this also selects which expert is live.
FAQ
- Is Wan an autoregressive video LLM?
- No. It is an iterative denoiser on a latent grid. Token time is diffusion time, not left-to-right words.
- Where is the U-Net?
- There isn’t one in Wan2.2. If a ComfyUI graph shows a UNetLoader, that is a loader name — the weights are still DiT blocks.
Primary sources
Checked against Wan-Video/Wan2.2 README · Aug 28, 2026
