latent-consistency

/

lcm-lora-sdxl

like 761

Follow

Latent Consistency 191

I am applying this SDXL LCM-LoRA to waiNSFWIllustrious - a fully fine-tuned model on SDXL base model. the final image style is not consistent with waiNSFWIllustrious originally generated one. My thought:

Base SDXL's path: noise → [step1] → [step2] → ... → [step50] → image
LCM LoRA learns: noise → [shortcut learned from above denoising trajectory] → image
waiNSFWIllustrious's path: noise → [step1'] → [step2'] → ... → [step50'] → image'
↑ Different trajectory!
since The entire denoising trajectory is changed, so we have to distill another LCM-lora on the base model waiNSFWIllustrious, even though waiNSFWIllustrious is fine-tuned out of SDXL. please confirm whether my thought is right or not