ComfyUI now natively supports masking and scheduling of LoRA and model weights. This works for all LoRAs and models - as long as the LoRAs and models are not of conflicting architectures. With these features, you can: Apply LoRAs via masks to specific areas of images, or to specific areas of specific frames for video models - loaded via Create Hook LoRA and CreateHook LoRA (MO) nodes. Apply models via masks to specific areas of images, or to specific areas of specific frames for video models - loaded via Create Hook Model as LoRA and Create Hook Model as LoRA (MO) nodes. Schedule LoRA and model strengths for different steps in the sampling process, allowing to mix the style and compositional benefits of separate LoRAs/models. Use the Create Hook LoRA/Model as LoRA nodes mentioned above for more simple LoRA/model mixing without needing to chain model/clip inputs and outputs.

The update also revamps the ModelPatcher system + adds wrappers/callbacks to make the lives of developers easier going forward, but that will be explained in a future blogpost and expanded even further with future updates.

LoRA - Low-Rank Adaptation. A technique for efficiently finetuning machine learning models