I merged SDXL Turbo with a lot of other top rated models on this site using a technique I implemented here from "Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch".
https://github.com/martyn/safetensors-merge-supermario
Supports LCM, SD Turbo, SD XL, 1024 res. The technique says it's for language models but it's generic.
No refiner necessary.
Description
Initial release
FAQ
Comments (11)
So the merging wars have started, again.
https://github.com/yule-BUAA/MergeLM
For anyone who want to read the origina repo. I think it's way way more complete about this technique than the other. However I have to say that merging without control... I know how it was on SD 1.5, in SDXL/Turbo, unsure. Cheers mate!
Oh... That was an idea I had about 5 months ago when I started developing AlbedoBaseXL, but it seems it's already been published in a paper! The validity of that paper is certainly being proven by the performance of my AlbedoBaseXL. I should take a closer look at the paper. Thank you.
@albedobond Im gonna hit you up on Discord soon asking a bit about this type of merging cause I have some questions. The way DARE operates looks clearly promising but I have to see it outside the LLM's area cause reading the paper I can see what it does and that gainings are there. Txt2img models? Unsure.
DDIM sampler at 6 steps with BasicScheduler / simple option works - no LCM.
Nothing was trained like that. Cool.
Doesn't work at all with the comfyUI turbo workflow. Neither with 512x512 nor 1024x1024. In which case, I think it's misleading to call this a "turbo" model. https://i.imgur.com/ePR0BwT.png https://i.imgur.com/paRLpAq.png With the actual turbo model you get better results than this with 1 step (this is 2 steps).
The title is descriptive - this is a merge of the turbo model with LCM / SDXL models using the attached code.
It's not meant to be a one-step turbo model.
@ntc What's the point of merging with turbo then?
@wowlad reduces the number of steps for 1024x1024 and enables 512x512.
I never claimed 1 step generation. Normal XL models don't work at all with the turbo scheduler. See the github.
Mamma mia! Excellent model!
可以解答一下，为什么模型加载不了呢
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.