Cog-SDXL

This is an implementation of Stability AI's SDXL as a Cog model.

Development

Follow the model pushing guide to push your own fork of SDXL to Replicate.

for prediction,

cog predict -i prompt="a photo of TOK"

cog train -i input_images=@example_datasets/__data.zip -i use_face_detection_instead=True

cog run -p 5000 python -m cog.server.http

2023-08-17

2023-08-12

Input types are inferred from input name extensions, or from the input_images_filetype argument
Preprocssing are now done with fp16, and if no mask is found, the model will use the whole image

2023-08-11

Default to 768x768 resolution training
Rank as argument now, default to 32
Now uses Swin2SR caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr as default, and will upscale + downscale to 768x768

Name		Name	Last commit message	Last commit date
Latest commit History 107 Commits
example_datasets		example_datasets
feature-extractor		feature-extractor
script		script
.dockerignore		.dockerignore
.gitignore		.gitignore
README.md		README.md
clip_embedder.py		clip_embedder.py
cog.yaml		cog.yaml
create_hyperparam_sweep.py		create_hyperparam_sweep.py
dataset_and_utils.py		dataset_and_utils.py
instructions_README.md		instructions_README.md
io_utils.py		io_utils.py
make_image_data.py		make_image_data.py
predict.py		predict.py
predict_backup.py		predict_backup.py
predict_deprecated.py		predict_deprecated.py
prep_model_dir.py		prep_model_dir.py
preprocess.py		preprocess.py
samples.py		samples.py
test_lora.py		test_lora.py
test_train.sh		test_train.sh
trainer_pti.py		trainer_pti.py
training_exp.sh		training_exp.sh
training_exp_n_tokens.sh		training_exp_n_tokens.sh
zip_args.sh		zip_args.sh