WAN2.2 VACE - Multimodal control-KJ

Continue the “unified editing/control” paradigm on the 2.2 backbone. The 2.2 backbone adopts a Mixture‑of‑Experts (MoE) design—high‑noise and low‑noise experts operating at different denoising stages—to improve quality and detail while keeping inference costs manageable. A representative controllable variant is Wan2.2‑VACE‑Fun‑A14B, which supports multi‑modal control conditions (Canny, Depth, OpenPose, MLSD, Trajectory, etc.). A typical workflow is: provide a reference image (to preserve identity/appearance) plus a driving video or its parsed control signals (e.g., pose sequence, trajectory, time‑varying depth/edges) to generate a video driven by that reference image. The VACE/Fun family provides these temporal control interfaces and the unified task support.