Cog SDXL Outpainting with LoRA support

This is an implementation of Stability AI's SDXL as a Cog model with ControlNet and Replicate's LoRA support.

Basic Usage

For prediction:

cog predict -i prompt="shot in the style of sksfer, ..." -i image=@image.png -i lora_weights="https://pbxt.replicate.delivery/mwN3AFyYZyouOB03Uhw8ubKW9rpqMgdtL9zYV9GF2WGDiwbE/trained_model.tar"

You may need to increase the default lora_scale value to big values such as 0.8 ... 0.95.

Limitations

• lora_weights only accepts models trained in Replicate and is a mandatory parameter.
• expert_ensemble_refiner is currently not supported, you can use base_image_refiner instead.
• For the moment, the model only uses canny as the conditional image.

For further improvements of this project, feel free to fork and PR!