Wan2.1 Image to Video Support for MacOS MLX
When are we getting Wan2.1 Image to Video Support for MacOS MLX. Community please can someone do this miracle? We need full support not CPU only
Are you familiar with DrawThings? The OSX/iOS-optimized app built on Swift, and at this point evolved into a versatile t2i/i2i/t2v/i2v inference/training environment with many clever integrations, and a very effective alternative to the MLX route. Recently, the main developer and curator of DrawThings, LiuLiu, has released (into the app) SVDQuant-method quantized accelerated versions of base Hunyuan Video and of the SkyReels fine-tune. I can personally confirm that with these quants, it takes mere minutes to generate Hunyuan clips at around 512x resolution on an M3 Macbook Air (16gb RAM). So, if making suchlike OSX-optimized SVDquants is possible for Hunyuan, it certainly doesn't seem too implausible in regards to Wan2.1, especially the 1.3, even now. With that said, one potential complication might be Wan2.1's much-trumpeted special new VAE, but it's also possible that it might translate smoothly enough. I would also be curious whether the Wan2.1 uses some kind of a modified/custom-tuned T5XXL, and how inter-operable it might be with the regular old T5XXL used by Flux, SD3, CogX, Mochi, and other models. In any case though, even a modified T5 shouldn't pose too much of an issue beyond forcing one to download more stuff.