Learning Goals
The Lightning model and LCM model can be used as needed.
Plain Text
복사
Models for Shorter Generation Time (Turbo, Lightning, LCM)
Lightning, LCM, and Turbo are technologies with the same goal: faster generation times and high-quality images.
They have different names based on the specific learning methods used.
The advantage is speed, but the drawback is that the image quality is lower compared to traditional LoRA and checkpoint models.
"Turbo+LCM" means that both methods were used together for training the model.
Turbo
The Turbo model is released by Stability AI.
It is fine-tuned to produce sharp images in just one sampling step.
However, commercial use of the Turbo model is not permitted without a Stability AI membership.
As a result, the alternative Lightning model is widely used.
Lightning
Lightning uses a technique called "Progressive Adversarial Diffusion Distillation" to generate images with fewer sampling steps.
It is commonly used as an alternative to Turbo.
Since the recommended Ksampler settings for Lightning can vary depending on the checkpoint, it is advised to refer to the checkpoint pages on sites like Civitai for specific recommendations.
LCM
LCM is a new type of diffusion model trained to generate images in just one step. It is primarily used as LCM-LoRA.
LCM-LoRA
LCM-LoRA is the LCM technology trained as a LoRA model. It can be combined with any checkpoint, which means using LCM-LoRA can reduce the generation time for all checkpoints.
When using LCM-LoRA, the Sampling Method must be set to LCM.
Recommended KSampler settings:
steps : 4 - 10
cfg : 1 - 2
sampler_name : lcm
scheduler : sgm_uniform
