2 steps always ignore many details, 8 steps will just same time as LCM.
So, why not keep balance ?
4-5 steps are fine to generate details which not loose so much and speed up the generation.
The original version was 300MB, it's a big one isn't it?
Don't worry, PAseer's version has been "slimmed weight down" while maintaining quality.
PAseer-Lighting only has 12MB
I've always suffered because the lora always try to change the meaning of my prompts while testing the original version.
That made me realize it has too much influence on the tokens.
PAseer's version reduces the influence of the Lighting-LORA on the prompts, and now you can generate the image more rely on your Prompts.
How to USE
1: Download the LORA.
2: Put PAseer-SDXL-AcceleratorLighting lora into .models/loras.
3: Refresh your UI.
4: Load it in your LORA-Loader.
5: Set steps 【4 to 5】
6: Set sampler 【as you like but LCM】
7: Set CFG 【1 to 2】
8: Start your generation.
Description
V1
FAQ
Comments (6)
In the description you recommend 4-5 steps and LCM sampler, while all your examples have 10 steps and Euler a. Would be good to see results with the settings you recommended or updated description to avoid confusion.
[But LCM]== do NOT use LCM.
Because Civitai Online Generation can not use less than 10 steps.
8 steps also avaiable but waste time.
If you want to use it on local, 4 steps are fine.
@Aseer My bad about the LCM, missed that. I've never generated here, didn't know about the 10 steps limitation. I always generate locally, will give it a try. Thanks for the clarification!
Lite-ning! I'll show myself the door
It really works well with ponyFaetality and everclearPNYByZovya and most realistic checkpoints. it's cool that you can generate with just 4-5 steps with almost all checkpoints and all in high quality. Thank you.
Perfect for my potato PC and best of all, it follows my prompts even at 1.5 CFG.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.