WanGP v18 adds Kandinsky 6.0 and much faster video VAEs
WanGP v18.00 rewrites the LTX 2, Wan and HunyuanVideo VAEs to be 2 to 3.5 times faster and lighter on VRAM. It also adds the Kandinsky 6.0 audio-video model, Wan 2.2 Prism Video and Wan 2.1 camera control.
What it does
The LTX 2 VAE is about 3x faster, and the Wan 2.1 VAE up to 3.5x faster. Decoding 1080p Wan drops from 28 s and 10.5 GB to 11 s and 4.7 GB, with no more tiling seams on Wan.
New models include Kandinsky 6.0 (Pro and Lite, distilled, with synced speech, music and effects), Wan 2.2 Prism Video (image-to-video with audio up to 2K), and Wan 2.1 camera-control models that fly the camera along a path from a start image.
Results and limits
The developer reports 20 seconds of 480p LTX 2 video in about 28 seconds on an RTX 5090 with Profile 4, and says CUDA graph acceleration gives about 0% gain on H3. Kandinsky clips are still only 5 seconds. These are the developer's own figures.
Worth knowing
On RTX 50xx cards, tick Config > Performance > WanGP Compute Cache to avoid a 10-15 second delay at each first generation. RTX 20xx and 30xx cards now get the full v17 VRAM savings.
Try it yourself
Update WanGP, pick Wan 2.1 Camera Control, load [your start image], draw a camera path ([slow dolly-in]), and generate. Then try Kandinsky 6.0 Lite with: "[Character] says '[line]' in [setting]."