Fast Stable Diffusion approaches list
2024-03-14

Here is a list of the current approaches to fast stable diffusion inference as of now. I won't elaborate on them for now (do your own research), since the space is moving so quickly that my information might quickly become obsolete.

Checkpoint scale approaches for fast txt2img:
- SDXL Lightning
- SDXL turbo
- SD2.1 turbo
- SDXL turbo model merge
- SDXL turbo model merge + Lightning LoRA/LCM LoRA
- SD1.5 finetune + LCM LoRA
- SDXL finetune + Turbo LoRA/LCM LoRA/Lightning LoRA

Speed tricks:
- Taesdxl dramatically improves VAE decoding speed.

Quality tricks (variable effectiveness):
- DPO LoRA
- Detail adder LoRA
- Hands/Face/Skin/Eyes LoRA
- Bad quality LoRA
- Negative embeddings