training

Invertible Consistency Distillation | Training example on MS-COCO

The training code for Stable Diffusion 1.5 and XL. This code is based on latent-consistency-model.

Environment

conda create -n icd-train python=3.10 -y 
conda activate icd-train

pip3 install -r requirements.txt

Dataset

We provide the training example on the MS-COCO dataset: train2014.tar.gz - contains the original COCO2014 train set.

Run training

1. Download the train dataset:
      bash data/download_coco_train2014.sh
2. Download the validation dataset:
      bash data/download_coco_val2014.sh
3. Download the files for FID calculation:
      bash stats/download_fid_files.sh
4. Download the pretrained CFG distilled teacher:
      bash pretrained/download_cfg_distill_sd15.sh or bash pretrained/download_cfg_distill_sdxl.sh
5. Run the training:
      bash sh_scripts/run_sd15_lora.sh or bash sh_scripts/run_sdxl_lora.sh

Training settings for the released models

Model	Batch size	LR	Max train steps	Preserve loss coefs
iCD-SD1.5	512	8e-6	6000	1.5, 1.5
iCD-XL	128	8e-6	6000	1.5, 1.5