Implemented a LoRA Hook system to dynamically apply weights to model during sampling. May need to tweak weight caching so it is disabled when lowvram is active, etc., but all the core functionality is working.
LoRA Hooks work by registering themselves on models, and then set on CLIP (for CLIP weights) and conditioning (for model weights). Also added the option to register a model as a lora to basically act like dynamic merging between two models - the stored weights are the difference between model's original weights and the target model's weights.
List of features/changes (not exhaustive):
LoRA Hooks to allow weights to be applied during sampling
LoRA Hooks can be applied to conditioning masks
LoRA Hooks can be applied to CLIP
LoRA Hooks can be scheduled to very strength through timesteps
Conditioning masking helpers via Set Props on Cond(s), Set Props on Cond(s) and Combine, Set Unmasked Cond(s)
Implemented a LoRA Hook system to dynamically apply weights to model during sampling. May need to tweak weight caching so it is disabled when lowvram is active, etc., but all the core functionality is working.
LoRA Hooks work by registering themselves on models, and then set on CLIP (for CLIP weights) and conditioning (for model weights). Also added the option to register a model as a lora to basically act like dynamic merging between two models - the stored weights are the difference between model's original weights and the target model's weights.
List of features/changes (not exhaustive):