Latent consistency distillation

This script is used to train a model with latent consistency distillation.

• test_train_lcm_lora_bucket.py is not memory optimization
• train_lcm_distill_sdxl_lora.py.py with memory optimization

Reference

https://github.com/huggingface/diffusers/blob/main/examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

huggingface/diffusers#5778

https://github.com/NovelAI/novelai-aspect-ratio-bucketing

Midfy&Add:

• Fix the dataset loading issue.
• Fix the LoRA weight saving issue.
• Reduce memory consumption during training.
• Add an aspect-ratio-bucketing option.

📕 Dataset structure

├── train_datasetr
│   ├── 000000.jpg
│   ├── 000000.caption
│   ├── 000001.jpg
│   ├── 000001.caption
│   ├── 000002.jpg
│   ├── 000002.caption
│   │...