Valheol

Can confirm, this fixes the issue for chroma when seeing the faketextencoder error after enabling caching of text embeddings.

> https://github.com/AbstractEyes/comfyui-lycoris > > I have made a correct implementation for precomputation since the hook prepare system does not function correctly with the intended goals of lycoris. The primary (OFFICIAL)...

Here is a Wan2.2 ComfyUI workflow with links to working fp8 scaled models that also supports LoRAs: https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/video_wan2_2_14B_t2v.json The fp8 scaled Wan2.2 models are 1/3rd the size of the originals....

The Wan2.2 MOE implementation feeds the high noise model latent outputs directly into the low noise model. The low noise model latent outputs are fed into VAE. Only one model...

Bonus: The existing WAN pipeline does support dual transformers.