Question About “Tuned Connector Data” and Face Drift Reduction
Thanks for the great work!
I'm curious about what exactly is meant by the following statement:
"Leveraged tuned connector data to reduce face drift and aid long prompts/director."
Does this mean you specifically tuned the Embeddings1DConnector (or another cross-modal connector component) to achieve better facial consistency and reduce face drift?
I find this particularly interesting because I would normally expect improvements in face consistency to come from the diffusion model, transformer, or training data itself. If tuning the cross-modal connector can noticeably improve visual quality, identity preservation, and prompt adherence over long generations, that seems like a very promising direction.
Could you elaborate on what "tuned connector data" refers to and how it contributes to reducing face drift and improving long-prompt/director performance?