Regular lightning, but on Civitai, if you mark a LoRA to be for SDXL it doesn't show up in the '+Add' list when using Pony. Alternatively this LoRA won't show up in the '+Add' list for people who've selected a SDXL model.

@TomLucidor 
https://civitai.com/images/33418468
Works great with anime, when using for anime I would recommend using a higher CFG if you want the image to be more colourful and dynamic.

@TomLucidor Should work with any Pony or SDXL model. From my personal testing Hyper 8step tends to produce more saturated and contrasty images relative to Lightning 8step. If using a CFG from 1 to 3, 8 steps should be enough. Going under 8steps tends to decrease image quality. I would only increase steps above 8 if you're getting weird yellow/green distortion in the images. A CFG of about 1.5 tends to work well with Real Dream 11, CFG of 2.0 with Real Dream 12, CFG of 3 with Pony V6, CFG of 3 with Autism. Exact settings will vary based on model and Lora used.

@nylonticklegal386 I use the Drawthings app on Mac to merge and so can only talk for Drawthings. I go to the model section, select mixing, select a checkpoint, select the lightning lora. I leave all of the other settings at default. So the lora is at 100% weight, model precision is at float 16 and the vae is set to 'same mix'. I then click 'mix' and it combines the lora with the checkpoint.

Either Civitai is glitching or they've changed something on the backend.

They should allow us to use SDXL, Pony and illustrious LoRA's interchangeably.

@Nyaa314 It wasn't working yesterday, but it seems to be working with SDXL, Pony and illustrious again.

Merging the LoRa into the model restores CFG usage because CFG doesn’t work with Hyper or Lightning LoRas when used standalone in DrawThings. When using acceleration LoRas standalone, DrawThings defaults to a CFG of 1, which is too low for models based on Pony or Illustrious. Merging the acceleration LoRa into the model makes CFG control work again. For Hyper LoRa with Illustrious or Pony models, setting the CFG to 2 results in much better image quality.

With SDXL-based models like Juggernaut or Dream Shaper, merging the LoRa still improves speed, but CFG of 1 works well enough in these cases. Using CFG of 1 significantly speeds up generation time on DrawThings compared to higher CFG values. However, for some models, CFG of 1 may not be high enough.

@alexvola My understanding is that Shift is mainly used for models similar to Flux, like Flux, Stable Diffusion 3.5, and possibly Huyuan. It doesn’t seem to make a difference for SDXL-based models. For SDXL models, I would recommend leaving Shift on the default setting, which I believe is 1.

@alexvola 
This model is really good: WAI v11.
It puts out better images than personal merge. Its default image output quality is really good. When used with a Hi-res image size of 1280x1856, with a Hi-res first pass size of 832x1216 and a 2nd pass strength of 45%. It puts out great images with great anatomy, that usually don't need 'Single Detailer' aka A-detailer or Image to image face-fix.

I use the positive prompt: (photorealistic, realistic, masterpiece, best quality, amazing quality, very aesthetic, absurd res, scenery, glossy image, shiny image, dramatic lighting, volumetric lighting, dramatic shadows).

I use the negative prompt: glitched, distorted, blurry, low res, low quality, bad quality, worst quality, low detail, worst detail, cropped, censor, censored, text, signature, watermark, username, artist name, chibi, child, young, short, ugly, bad eyes, closed eyes, deformed, disfigured, bad anatomy, bad hands, monochrome, grayscale, sketch, simple background,

Note: You can remove the tags 'photorealistic, realistic' from the positive prompt if you want a more 2D look.

I use it with a DMD2 LoRA at a CFG of 1 and 8 steps, but if you're using a Hyper LoRA you could probably generate the images at a CFG of 2 or 2.5 at 8 or about 10 steps.

If you want to use DMD2, you can get it from within the DrawThings app it's in the community LoRA section. I use it as it's more powerful than Hyper and allows you to generate images at a CFG of 1. Generating images at a CFG of 1 is a lot faster than anything above 1. I wouldn't use it with a base SDXL model thou as it's too powerful and cooks the image.

So use:
WAI v11 merged with DMD2 at 100%.
In model menu convert this model to 8-bit to halve RAM usage.
Use a CFG of 1 at 8 steps with Eular A Trailing.

For extra high quality you can also add these settings:
Image size of 1280x1856 in the basic menu.
In the advanced section activate Hi-res fix, set Hi-res size to 832x1216 and its strength to 45%.
Turn off Tiled decoding and Tiled diffusion.