furusu / ControlNet

Can be used in sd-webui-controlnet . Put the yaml file with model because it is sd2.1 v_prediction model.

Model is the difference from the original checkpoint created by extract_controlnet_diff.py .

These are not the final version and need to be adjusted to be more accurate. Someone please do this.

wd15_controlnet_edge_depth_pose.ckpt

• target model: https://huggingface.co/waifu-diffusion/wd-1-5-beta2
• dataset: 20k 1girl images
• epoch: 5
• batch size: 8
• lr: 5e-5
• amp:bfloat16

diff_control_wd15beta2_canny.safetensors

updated in 2023/03/05.

• target model: https://huggingface.co/waifu-diffusion/wd-1-5-beta2
• dataset: 11k 1girl images
• epoch: 2
• batch size: 16
• lr: 5e-5
• amp:bfloat16

and resumed on:

• dataset: 60k 1girl images
• epoch: 2
• batch size: 16
• lr: 1e-5
• amp:bfloat16

diff_control_wd15beta2_pose.safetensors

updated in 2023/03/06.

• target model: https://huggingface.co/waifu-diffusion/wd-1-5-beta2
• dataset: 14k umamusume images. Therefore, the model may have a bias
• epoch: 5 (The “sudden convergence phenomenon” occurred at the fourth epoch)
• batch size: 16
• lr: 5e-5
• amp:bfloat16

and resumed on:

• dataset: 52k 1girl images
• epoch: 2
• batch size: 16
• lr: 1e-5
• amp:bfloat16

diff_control_wd15beta2_depth.safetensors

updated in 2023/03/06.

• target model: https://huggingface.co/waifu-diffusion/wd-1-5-beta2
• dataset: 11k 1girl images
• epoch: 2
• batch size: 16
• lr: 5e-5
• amp:bfloat16

and resumed on:

• dataset: 60k 1girl images
• epoch: 1
• batch size: 16
• lr: 1e-5
• amp:bfloat16