Originally posted on Huggingface.
AuraFlow is the fully open-sourced largest flow-based text-to-image generation model.
This model achieves state-of-the-art results on GenEval. Read our blog post for more technical details.
The model is currently in beta. We are working on improving it and the community's feedback is important. Join fal's Discord to give us feedback and stay in touch with the model development.
Credits: A huge thank you to @cloneofsimo and @isidentical for bringing this project to life. It's incredible what two cracked engineers can achieve in such a short period of time. We also extend our gratitude to the incredible researchers whose prior work laid the foundation for our efforts.
Description
FAQ
Comments (50)
lets go to try
Amazing.One great first step into the right direction. :-)
Congratulation~!
16G, wow, a bit large. So this till a SDXL model ? what is the figures of this one ?
It's a new foundational model with its own architecture - not SDXL based. It's also not my model; please see the details in the description. Thanks!
We need much more information about environment,stable diffusion user-interface,workflows etc.At the moment the given information is kinda nebulous,at least for some aspects .
The description,well,at the moment its more marketing than anything else.
A new model is always good news and I see the focus atm is to push the model and to get support.
I am not a friend of the fact people are presenting stuff here only for presenting and leave everything else to the community here;without the ability to moderate the content.
Especially if its a new model,and maybe in this special case.
Its strongly adviced to let the developers handle this,or any related person that has a deeper insight.
No offense at all dear theally <3 I hope you get the point.
@theally Yeah and just after my posting I noticed that you are OP (overpowered?) and that this posting needs more balancing...............
@theally I read the article and thank for your and the team's hard work to open-source. But recently, Hunyuan, SD3, Kolors, and this one comes out independently and individually, which makes StableDiffusion more and more complex and hard to use, each model has their [self] method to use. I do think if we were willing to contribute to open-source, 【if no fanancial problem】, make things easy is one of the important tenders.
Bootleg SD3?
more like what SD3 should of been, it works, doesn't have a batshit bad license and wasn't artificially hyped up so much that if it fails expectations people won't grab pitchforks and torches.
@TheP3NGU1N The problem with the license is no longer relevant! because Stability Ai changed the SD3 license not so long ago, and the fact that people did not meet expectations is their problem! not the SD3 model
@prgfrg23 Then the only problem is the "it works" part.
@prgfrg23 sd3 model is the problem
It is a good new but it is useless for GPU less than 16GB.
It can most probably be quantized to 16 bits
so lucky im receiving soon my 4090... have been linked to my 2070 super for long time
I'm using an RTX3060 with 12GB and it works fine. Yes, it is a little bit slow but still good enough to experiment with it.
How does it compare to SD3? I mean as far as parameters, VAE, etc...?
Could it be further tuned and tweaked and later have controlnets, ipadapter, PAG, adetailer, SegMoE etc?
Can we train LoRAs on it?
Apparently you can train loras
https://github.com/bghira/simpletuner
Is it possible to offload the Text encoder onto the CPU to save on memory? I don't see any way to do that in this workflow.
(Congratulations to you and your team for being this community's savior btw!)
As someone with a 8gb vram GPU, PLEASE optimize the model 🥺🥺🥺
Ignore MAC User?
Sure a good model, I can't confirm because I'm Mac user! CPU takes forever and I don't have a CUDA system. I get an error message because MPS is not supported. Is there still a revised version?
Which Mac model are you using?
@migmag
MacBook Pro M3 Max
How long until PonyFlow?
This model is too big can u please make a pruned version that's 7gb please I can't load this on my phone
Bro I managed to load it in smartwatch, its not that much large.
hands down the best model you can play with open source
It is large, takes almost 1 minute on a 4060 ti 16GB per image. Can't generate anything better than sd1.
skill issue, use some llms to enhance prompt.
Cons: It is a little slow, I've got a 16 GB VRAM and it takes a while.
Pro: Makes some pretty great logos and seems to handle text a lot better than many other models. I was getting not so great results and checked the workflow from huggingface and the KSampler there was set as follows: Sampler-Uni_Pc, Scheduler: Normal, Steps: 25+ and CFG: 3.5. Duplicating this in my ComfyUI workflow has given much clearer and more pleasing results.
I did a bit more testing this evening with v.3, and tried a few samplers.
The default Uni_PC worked fine, as did Euler, and DPM++, however Euler ancestral was a garbled mess of blue pixels.
I hope we can soon train our Loras directly on Civitai
it's here, some time already https://civitai.com/models/train
@menegosm I can find it under custom models. But I can't select a version and therefore can't press the select button. Is it working for you? Unfortunately I can't post a screenshot.
@menegosm or is it aura render xl?
sorry, haven't try it. just know its existence
Laptop: 16gb ram 3070ti 8gb vram, is working fine with 1024x1024 (maybe a bit slow)
looking forward to a more advanced version. Is there a way to inpaint with this model?
Runs on Omen16, rtx4060m 8GB dedicated and 8 shared. 8 images of 1024x512 res and first batch took about 24 minutes, second about 18 minutes.
Though, i do store the checkpoints externally and load over usb.
768x1280 at 30 steps takes me about 50s so something must be very wrong with your setup
See https://civitai.com/articles/6364/auraflow-fp16-diy-checkpoint-with-comfyui for how to get an fp16 AuraFlow model up and running. I was able to generate models on 8gb of ram at 8s/it without --lowvram or --disable-smart-memory
It's interesting to see how quickly comfy was able to respond even though it's not Stable Diffusion.
On the other hand, it's unfortunate that it's unclear whether the webUI of more common tools will be compatible.
The release of version 0.2 is now available. It would be great to include it here.
Nice! Another 17gb model that makes my gpu produce x-rays! 😂🎉
It's up! Thanks
@theally THANKS 😊
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.