Summary
I noticed that the recently added MotifVideo pipelines currently do not expose pipeline-level LoRA loading support, even though the underlying transformer appears to already be PEFT-ready.
Before working on a PR, I'd like to check whether adding LoRA support for the MotifVideo pipelines would be considered in scope for Diffusers.
What I found
During a local inspection of the implementation:
MotifVideoTransformer3DModel inherits from PeftAdapterMixin.- The MotifVideo pipelines (
MotifVideoPipeline and MotifVideoImage2VideoPipeline) do not inherit from a *LoraLoaderMixin.
- I could not find a
MotifVideoLoraLoaderMixin in src/diffusers/loaders/lora_pipeline.py, while many other recently added model families (e.g. Wan, Cosmos, Flux2, LTX2, Ideogram4, etc.) provide dedicated pipeline-level LoRA loader mixins.
This suggests that the model already contains the PEFT infrastructure, but the pipeline API does not currently expose the standard Diffusers LoRA loading interface.
Proposed feature
If this is considered in scope, I would be interested in contributing support for loading LoRA adapters in the MotifVideo pipelines by adding a dedicated MotifVideoLoraLoaderMixin and wiring it into the relevant pipelines, following the existing patterns used by other model families.
Question
Would the maintainers be open to a PR adding pipeline-level LoRA support for MotifVideo?
If LoRA support was intentionally omitted (for example because of architectural limitations or lack of compatibility with existing LoRA checkpoints), I'd also appreciate understanding the reasoning before spending time on an implementation.
Summary
I noticed that the recently added MotifVideo pipelines currently do not expose pipeline-level LoRA loading support, even though the underlying transformer appears to already be PEFT-ready.
Before working on a PR, I'd like to check whether adding LoRA support for the MotifVideo pipelines would be considered in scope for Diffusers.
What I found
During a local inspection of the implementation:
MotifVideoTransformer3DModelinherits fromPeftAdapterMixin.MotifVideoPipelineandMotifVideoImage2VideoPipeline) do not inherit from a*LoraLoaderMixin.MotifVideoLoraLoaderMixininsrc/diffusers/loaders/lora_pipeline.py, while many other recently added model families (e.g. Wan, Cosmos, Flux2, LTX2, Ideogram4, etc.) provide dedicated pipeline-level LoRA loader mixins.This suggests that the model already contains the PEFT infrastructure, but the pipeline API does not currently expose the standard Diffusers LoRA loading interface.
Proposed feature
If this is considered in scope, I would be interested in contributing support for loading LoRA adapters in the MotifVideo pipelines by adding a dedicated
MotifVideoLoraLoaderMixinand wiring it into the relevant pipelines, following the existing patterns used by other model families.Question
Would the maintainers be open to a PR adding pipeline-level LoRA support for MotifVideo?
If LoRA support was intentionally omitted (for example because of architectural limitations or lack of compatibility with existing LoRA checkpoints), I'd also appreciate understanding the reasoning before spending time on an implementation.