Hyper has a buttchin problem but can generate proper text at low resolution, like 512. Even if it still generates improper text, it can be fixed with clip attention multiply node, but the quality is lower than FP8. Turbo doesn't have this buttchin problem but messes up the text at lower resolution. However, the plus is that it has much better quality, almost as good as FP8.
i personally take the turbo 8 model, because you always can increase the steps to get better results. i made good results with 12 or sometimes even 16 steps and with wavespeed it's even faster. 3-4 Seconds for like bit more than 720p image and 25 seconds for a 10 second video with LTX, this is nuts! But i don't use the turbo model, i use the original dev 23gb model with the 8-step turbo Lora. not sure if this is the same you are using here?
I've grown to love the turbo model and use it quite a bit now. For the demo's it was using the Q8 as I can't really run the full dev on my lowly 3060Ti. 😁
All the Flux model links can be found here docs.google.com/document/d/1OuIwvcaTvBM2gJukWdeSuV0Ln_TPVaPQH7xX7FlwULU/edit?usp=sharing And I suggest you to watch this video to make sense of all of it. th-cam.com/video/qt-TeQVpvxg/w-d-xo.htmlsi=FiWXtjHFFEKzU0Pj
Turbo 8 quality is very close to Dev-Q8, with generation times 2 to 3 times faster, which is a good point.
Absolutely! Most images I've tested the slight loss in quality is hardly noticeable.
That’s a really great comparison, thanks for info and the video 🙌🙌
Appreciate it bro! Have fun wit hit! 👍
Let me know which one you prefer to use Turbo or Hyper? And what has your experience been like using them?
Thank you for the vid, good as always. Hyper 8 seems smoother so I think I would use that but haven't tried it yet. Someday I'll try it out.
You’re welcome! Both are fun to use and being faster with less steps is amazing! 😊
Hyper has a buttchin problem but can generate proper text at low resolution, like 512. Even if it still generates improper text, it can be fixed with clip attention multiply node, but the quality is lower than FP8. Turbo doesn't have this buttchin problem but messes up the text at lower resolution. However, the plus is that it has much better quality, almost as good as FP8.
The clef chin issue is with Flux in general but yes it seems like with hyper it's more noticeable. Turbo has it as well.
i personally take the turbo 8 model, because you always can increase the steps to get better results. i made good results with 12 or sometimes even 16 steps and with wavespeed it's even faster. 3-4 Seconds for like bit more than 720p image and 25 seconds for a 10 second video with LTX, this is nuts!
But i don't use the turbo model, i use the original dev 23gb model with the 8-step turbo Lora. not sure if this is the same you are using here?
I've grown to love the turbo model and use it quite a bit now. For the demo's it was using the Q8 as I can't really run the full dev on my lowly 3060Ti. 😁
good stuff thank you
You're welcome! 👍
Can I use it with bnb nf4 v2 model as well ?
Because I tried with similar settings but the output is not that great
As far as I know it doesn't. But the Q8 GGUF is the same size why not run that instead? A bit slower but better than NF4.
@@MonzonMediaThanks man I will give it a try. Can you suggest any hf links if possible to download that particular model version?
All the Flux model links can be found here docs.google.com/document/d/1OuIwvcaTvBM2gJukWdeSuV0Ln_TPVaPQH7xX7FlwULU/edit?usp=sharing And I suggest you to watch this video to make sense of all of it. th-cam.com/video/qt-TeQVpvxg/w-d-xo.htmlsi=FiWXtjHFFEKzU0Pj
What Ui are you using? Is this on Google Collab? Any possibility on tutorial on how you got your Ui set up?
I don't have a high end PC.
This is Forge th-cam.com/video/FKzvHFtc8N0/w-d-xo.htmlsi=poChkq4MgkgUA4RH What are your PC specs and GPU?
I don't understand this is it a checkpoint or a lora?
It’s a Lora.
@@MonzonMedia So it's a lora to speed up image generation time?
@@petertremblay3725 Not necessarily speed up, more like less steps so in turn you can generate faster.
@@MonzonMedia OK thank now i get it, i am 60 so it take me longer to get it! loll
@@petertremblay3725 Take all the time you need and I'd be happy to answer any questions you may have. 👍
Turbo 8 is the best, test it with a 2Mpx image, the quality is much better than the lora Hyper
It's a pretty nice lora especially for people like me with lower end GPU ☺