AUTOMATIC1111
[Feature Request]: Please add support for the FLUX model, thank you!
Is there an existing issue for this?
- [X] I have searched the existing issues and checked the recent builds/commits
What would your feature do ?
The FLUX model's hand processing and prompt accuracy are incredibly powerful, and it's been super popular recently!
Proposed workflow
- thank you!
- thank you!
- thank you!
Additional information
No response
Duplicate of https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/16311 You can close it.
TLDR: SD WebUI uses a different diffusion library than what Flux is supported on. It would be a lot of effort to change it.
@FlyWieEinAirplane and that's bad. I don't wanna use that clusterfuck comfyui =_=
@FlyWieEinAirplane and that's bad. I don't wanna use that clusterfuck
comfyui=_=
Then use Forge
Why does A1111 use a different diffusers setup than the standard release?
Why does A1111 use a different diffusers setup than the standard release?
I think there is nothing standard about Flux
Even in Forge only "Euler" and sometimes "LMS" samplers are working on FLUX and this only on the "simple" type (no karras or others). So Forge seems to have problems with FLUX too.
Rendering a prompt from A1111 in Forge with identical settings, generates a different picture. ... at least on my system.
So I think it's very important, to keep care on integrating FLUX, not to disturb the usual behaviour and functionality of A1111 for the older models. So if FLUX comes, it's welcome, but not for every price.
Karras and exponential sampler types do not work on SD3/FLUX type of models regardless of backend.
Karras and exponential sampler types do not work on SD3/FLUX type of models regardless of backend.
... because it's not yet included/added/improved ...
Karras and exponential sampler types do not work on SD3/FLUX type of models regardless of backend.
... because it's not yet included/added/improved ... This is how it was explained to me - "early SD models used stochastic flow, ie they were designed every step to take random noise and remove it until they get an image back. Modern SD models (SD3, Flux) use rectified flow, ie they were designed to take one initial noise blob and step iteratively towards a known trajectory of where the image will be. Rectified flow is the key to how example how turbo models work - once you have a known trajectory, you can distill the model to travel the known path much quicker.
Karras, Exponential, and all the other schedulers n samplers are basically ways of actively fucking up the denoising process, based on the principle that - well, it's random intelligent noise removal, maybe if we give it differenty noise, it'll do betterer. And sometimes it does. These all don't apply to rectified flow because it is now trying to work with a predictable coherent trajectory, so fucking around with the noise just breaks the one thing it's been trained to do."
[lllyasviel]/ [stable-diffusion-webui-forge] 似乎现在支持 Flux
But I can't live without comfyui anymore.