Fooocus icon indicating copy to clipboard operation
Fooocus copied to clipboard

Support for SDXL Turbo

Open Szakbal2001 opened this issue 1 year ago • 10 comments

Stability AI's latest 1 step generation model

Szakbal2001 avatar Dec 19 '23 05:12 Szakbal2001

It's supported already.

  1. tick Advanced
  2. Advanced tab, top right corner
  3. tick developer debug mode
  4. Scheduler -> turbo
  5. Forced sampling steps -> 1-4
  6. Forced resolution -> 512 x 512
  7. CFG -> 1.0-1.4

turbo

jepjoo avatar Dec 19 '23 08:12 jepjoo

Is there a sampler that works particularly well or bad with the Turbo scheduler? And is it usable with any SDXL models since it is a scheduler instead of being built into a checkpoint?

IamTirion avatar Dec 19 '23 17:12 IamTirion

Here's my workflow. It's not as difficult as you might think. Hope this helps.

https://mindrenders.com/sdxl-turbo-fooocus/

ItsNoted avatar Dec 20 '23 16:12 ItsNoted

@ItsNoted thank you for the post, i already read it a few days ago and with Dreamshaper XL also XL resolutions work really well.

Making it easier to use by adding SDXL Turbo as a performance preset isn't recommended though in the current state as it has been "released under a non-commercial research license that permits personal, non-commercial use." plus one has to use specific SDXL Turbo models.

see

  • https://stability.ai/news/stability-ai-sdxl-turbo
  • https://huggingface.co/stabilityai/sdxl-turbo#recommendations

mashb1t avatar Dec 22 '23 21:12 mashb1t

Making it easier to use by adding SDXL Turbo as a performance preset isn't recommended though in the current state as it has been "released under a non-commercial research license that permits personal, non-commercial use."

Not arguing against you, but just saying, this will be the case for all the models Stability AI release in the future. They said it here: https://stability.ai/news/introducing-stability-ai-membership "A Membership Tier for Every Application

As we grow, our ethos remains unchanged: Making state-of-the-art, open generative AI accessible to all. To achieve this, we've introduced three membership tiers designed to serve a wide range of users – from individual generative AI enthusiasts to Fortune 500 companies – this membership ensures affordable self-hosted solutions for all through our existing core models, with continual additions over time.

Non-Commercial At no cost for personal use and research, this first membership tier is rooted in our fundamental belief that this technology should be transparent and widely available to everyone. At no cost, this tier offers access to our full suite of core models for personal and research use.

Professional For creators, developers, and startups, currently offered at $20 USD a month, Professional members receive access to self-hosted, commercial usage of our full suite of core models with all the benefits in the Non-Commercial tier and access to our Professional tier Discord community. If you wish to use our core models commercially but find the cost to be a barrier, we have some discounts available to support these circumstances. Please contact us to learn more."

IamTirion avatar Dec 22 '23 21:12 IamTirion

@IamTirion oh, i didn't know about that, thank you for the information! This might lead to us having to mention/reference this information somewhere when we add support for one of their core models by default.

mashb1t avatar Dec 22 '23 22:12 mashb1t

It's supported already.

  1. tick Advanced
  2. Advanced tab, top right corner
  3. tick developer debug mode
  4. Scheduler -> turbo
  5. Forced sampling steps -> 1-4
  6. Forced resolution -> 512 x 512
  7. CFG -> 1.0-1.4

turbo

I followed this but ended up with this error: AttributeError: 'Patcher' object has no attribute 'load_device'

Any idea why? I used sdxl turbo fp16

jandzouana avatar Feb 05 '24 09:02 jandzouana

It's supported already.

  1. tick Advanced
  2. Advanced tab, top right corner
  3. tick developer debug mode
  4. Scheduler -> turbo
  5. Forced sampling steps -> 1-4
  6. Forced resolution -> 512 x 512
  7. CFG -> 1.0-1.4

turbo

I followed this but ended up with this error: AttributeError: 'Patcher' object has no attribute 'load_device'

Any idea why? I used sdxl turbo fp16

The same for me! Same error following all the steps

LiteSoul avatar Feb 25 '24 00:02 LiteSoul

i had this issue to until i put performance to extreme speed then it download the proper file. but it still only works for Extreme not Speed or Quality. I assume this is by design.

Rotoslider avatar Mar 05 '24 02:03 Rotoslider

Sampler appears to be stuck with "euler" regardless of setting in advanced. Using DreamshaperXL Turbo 2.1.

jonshipman avatar Apr 18 '24 20:04 jonshipman