v1: 128 dim
v1-64dim: half dimension, similar quality
v1-16dim: 94MB, similar quality(but different results in some prompts... I suggest the 128 dim or 64 dim version, but if you have a gpu with 6 GB vram try this version)
LoRA based on sdxl turbo, you can use the TURBO with any stable diffusion xl checkpoint, few seconds = 1 image(3 seconds with a nvidia rtx 3060 12 GB with 1024x768 resolution)
---
If you want to support me, check out my Patreon: https://www.patreon.com/Shiroppo
Follow Me on Twitter/X: https://x.com/ShiroppoTwit
Pixiv: https://www.pixiv.net/users/96490888
---
Tested on webui 1111 v1.6.0
Tested on ComfyUI v1754 [777f6b15]: workflow --download here--
(new version of workflow November 30, 2023 remove strange violet fog, keep sgm_uniform as scheduler and you can use every sampler)
1-Select your favourite stable diffusion xl checkpoint
2-Download this LoRA, use my workflow for ComfyUI or any workflow with LoRA loader
For webui 1111 write in the prompt <lora:sd_xl_turbo_lora_v1:1>
3-Sampling method on webui 1111: LCM(install animatediff extension if you don't see it in sampling list)
Sampling method on ComfyUI: all, with the workflow of November 30, 2023
4-CFG Scale: from 1 to 2.5
5-Sampling steps: 4+
Hugginface page: https://huggingface.co/shiroppo/sd_xl_turbo_lora
Description
v1-64dim: half dimension, similar quality
FAQ
Comments (18)
How is this different from LCM-lora which is already published https://civitai.com/models/195519?modelVersionId=225222 ?
Is this a lora extract or etc of https://huggingface.co/stabilityai/sdxl-turbo? Just tested the comfy workflow, and at 4 steps, this does seem to be slightly better than LCM-lora ~6.
@halr9000 Yes, this is based on stability ai turbo model... which it was published on November 28, same speed but better quality... see the second image for the quality difference.
@halr9000 Exactly I extract the LoRA few hours ago
different model. This is based on SDXL Turbo, a model that's capable of doing a single step image generation. Latent consistency is a different concept that reduces the necessary number of steps, but is based on a different concept.
So at 4-6 steps you'll see an improvement in quality, but at 1 step, there's going to be no question which method works better
What kind of question is this? you might as well ask what's the difference between SDXL-Turbo and regular LCM?
Not a good idea to use lcm sampler. It degrades quality. Using any other sampler though, you get the same quality as SDXL-Turbo!
You can only use LCM or Euler Ancenstral(on ComfyUI sometimes quite quite quite good results), i tested with the other sampler and the quality is very bad
Edit: new ComfyUI workflow, now good quality with all samplers
Hello, I've been getting trying this with the JuggernautXL models (v6 and v7) using the ComfyUI workflow and every output has a colourful fog over it. I've tried both versions of the LoRA and both are having the same problem for me.
Is there anything I might be missing?
https://civitai.com/api/download/training-data/243508 <-- have you tried this workflow?
Example with juggernautXL_version6: https://civitai.com/images/4009285
@shiroppo I'm using that workflow with the only changes being the SDXL model and the prompt
Example output: https://civitai.com/posts/905540
@2haloes I'm investigating.. basically by removing the SDTurboScheduler node, everything is resolved but by doing this you can only use lcm and euler ancestral as sampler, with the SDTurboScheduler that "fog" is removed only with steps equal to or greater than 6... making everything useless... I need to check that component(understand who created it) and the reason for the "incompatibility" with the lora
@shiroppo Sorry for breaking it! I hope it's not going to be too annoying to fix.
Thank you so much for investigating!
https://civitai.com/api/download/training-data/243508 <-- with this workflow it's all ok(same link but new version)
at the end it was enough to use the sgm_uniform as scheduler and every sampler works now, completely removing the SDTurboScheduler node which was incompatible with this LoRA
Thanks a lot for making me aware of the problem!
The purple haze appears on a lot of samplers for A1111, Otherwise this lora is crazy powerful for SDXL
For A1111 I have to check, but I think that for now only lcm sampler is good, also with sdxl turbo checkpoint. For comfyui with my last workflow(few hours ago) I didn't see purple haze anymore.
I hope A1111 to be as fast as ComfyUI, both in generation speed and implementation of new tech. Thanks for this lora.
4-CFG Scale: from 1 to 2
cfg 2.5 looks better