An important thing to note... this new Turbo lora is just HALF the size of the Hyper loras. So, if the quality is similar, I suppose it's good to save the extra 660MB of VRAM.
is there not a way to generate an idea first in low quality as possible then run that same picture through a lora that handles higher quality + upscaling surelty this would yield the best experience for people without insane machines
Use SDXL to create a base image and then an Img2Img workflow with Flux to get something nicer. Of course, you'll lose the prompt following capability of Flux in that case.
honnestly after many tests this upscaler is garbage. The output is always worst, adding some cartoonish details. I've tested it with all kind of different settings, even at 50 steps + it sucks.
The trouble with Flux models is no matter what jiggerypokery you do - 9/10 times you end up with Fluxina. And all the other effort is rendered moot. Its making me skip through videos that talk about anything Flux
Worked on my GGUF Q4KS unet and also seems to perform better than Hyper Flux with a smaller package. Thanks.
I use it with Q6 gguf model , with 3 step and results are accurate
what is the difference with the flux hyper lora? is the quality better? or is there anything better? it also allow to use in 8 steps
This is my real question. I've been using the hyper loras for quite awhile already. And with those, I can choose any flux checkpoint I want. 🤷🏼
An important thing to note... this new Turbo lora is just HALF the size of the Hyper loras. So, if the quality is similar, I suppose it's good to save the extra 660MB of VRAM.
does it work with gguf Q8?
It's crazy you don't have a discord group.
On the first example, it looked like the turbo lora just ignored the CR2 in the prompt.
is there not a way to generate an idea first in low quality as possible then run that same picture through a lora that handles higher quality + upscaling surelty this would yield the best experience for people without insane machines
Use SDXL to create a base image and then an Img2Img workflow with Flux to get something nicer. Of course, you'll lose the prompt following capability of Flux in that case.
honnestly after many tests this upscaler is garbage. The output is always worst, adding some cartoonish details. I've tested it with all kind of different settings, even at 50 steps + it sucks.
Someone has tried generating text with it? It looks way worse on that.
The trouble with Flux models is no matter what jiggerypokery you do - 9/10 times you end up with Fluxina. And all the other effort is rendered moot. Its making me skip through videos that talk about anything Flux
the jasper model is too stooooooopid , the image u were showing in upscale looks awful, sorry.
but i use krita! i dont need this!