It’s been cooking for a while but we can now officially announce that the Flux.1-schnell is finally available on the AI Horde!
Flux is one of the most exciting Generative AI text2image models to come out this year, from a team of ex-stability.ai developers, and seemingly consumed all the attention of the GenAI enthusiasts overnight. It’s a very powerful model but as a downside it requires a significantly more powerful PC to run than the more popular SDXL models were until now.
The model available on the horde is primarily the fp8 compact version we took from civitAI, as it simplifies the amount of downloads we have to juggle.
I was really eager to offer the flux.1-dev version as well, as it has a lot more LoRas available and is a bit more versatile, but sadly its license contains some requirements which do not appear to allow a service like the AI Horde to provide it, even though it’s a completely free service for everyone. However we have reached to the Black Forest Labs via email to ask for clarification or exception for this and will let you all know if we hear back.
To use it, head over to Artbot or Lucid Creations and simply select the Flux.1-Schnell fp8 (Compact) model for your generation. However keep in mind that this model is quite different from the Stable Diffusion models you’re used to until now, so you need to adjust your request as following to get good results:
- Set sampler to k_euler
- Set steps between 4 and 8 (4 is enough for most images)
- Set cfg to 1
Also keep in mind that the model won’t use the negative prompt. Instead it benefits massively from using native speech to describe what you want to draw instead of a tag-based approach.
If you are running a dreamer worker make sure you check our instructions in our discord channel on the best settings to run flux. This is a big model, so GPU with 16G-24G VRAM are the best for running it at a decent speed and we could use all the help we can get.
If you are making integrations with the AI Horde, make sure you use the flux branch of the image reference repository until it’s merged into main on the end of the month, if you’re using it to retrieve model requirements.
Along with flux, tazlin has done some amazing work on adding the latest version of comfy and improving the stability and speed of the worker. I mean, just look at this changelog! This also greatly improves our support for AMD cards. They might not be as fast as nvidia, but they should work!
Finally we’ve added some improvements on the horde itself to allow slower workers to offer models. If you have an older GPU which often gets timed out and put on maintenance on the Horde due to speed, you can now set yourself as an extra_slow_worker which will extend your TTL and will be used by things like automated bots, or apps like that sweet AI Wallpaper Changer.
Finally, I’ve also extended our deployments ansible collection so that if you use a Linux system, you can easily deploy any number of reGen workers, even multiple in the same server to take advantage of multiple GPUs. It will even deploy the AMD drivers for you if you want it. With this I am continuing to extend the tools to allow more people to run the AI Horde infrastructure on their own.
We hope the existence of flux on the Horde will allow unlimited creativity from people who want access to the model but don’t have the hardware to run it. Now more than ever, people with mid-range GPUs can offer what they can run, such as SDXL or SD 1.5 models, and in turn, benefit from others offering the larger models like flux and we all benefit through mutual aid!
Enjoy!