Deprecation notice

This model was a research project focused on the effect of fine-tuning the OpenCLIP text encoder.

It has been deprecated in favour of a newer checkpoint that continued training this model.

This model remains accessible as a test comparison and possible base model for fine-tuning.

Training data

Base model: stabilityai/stable-diffusion-2-1

Data: 3300 midjourney 5.1 upscaled images with their captions.

Training parameters

Duration: 4000 steps LR scheduler: polynomial Batch size: 3 Text encoder: Thawed Optimizer: 8bit ADAM No prior loss preservation