Uploading right now! Thanks! :)

Apparently Pony is doing a model based on this.

@buiscuit_monster Yup thats mostly why I'm getting acquainted with it because pony6 is my most used model and want to prepare for it.

@buiscuit_monster yeah, but is it also usable in forge/a1111? im afraid it will only work in comfy, i dont wanna be forced to use that

So how's in comparison with v2?, didn't the prompt adherence decreased?

@P_Universe I did not test v.2, only v.1 and v.3

I'm using euler + sgm_uniform.

So what sampler and steps do you recommend?

@P_Universe any of the others were rendering okay images and up to personal preference as usual. just for now, stay away from ancestral ones. you'll get a completely warped image that looks like a failing blue screen of death if you know what i mean.

Have you tried to get Flux Schnell working?

@LemonSparkle Haven't tried, and don't think I would have any chance: My card would need to upcast to fp16 and then we're looking at 30G for the pruned model. Compare with SDXL which runs fine-ish (as in spending more time computing instead of doing memory transfers), 6.4G at fp16.

@not_a_number Maybe you could do one of the NF4 versions? I think I saw someone say it was like ~7.6G in NF4.

@LemonSparkle ROCm doesn't support NF4 at all, least of all the ancient version I need to run to get it to work on NAVI14 (RX5500). I'm actually still in awe that AuraFlow even technically worked (though it's unusable) and SDXL is not fast, but absolutely usable. The card is barely not a potato and it's not exactly new.

Use ComfyUI

@Wizer Cool, thanks for the info! I got another reason to download and get into using ComfyUI.

@Wizer Comfy is terrible though.

@StecFX If you think so, you can try using diffusers directly )

Usage example at the link below

https://huggingface.co/fal/AuraFlow-v0.3

@BattleRabbitAIart  I guess you can use SwarmUI, it's basically a more user-friendly UI over comfy.

@dextermorgan I was actually just talking about Swarm with one of my co-workers. It definitely stands out to me and I'll probably give it a try at some point.

AF works with rtx 2060 6gb, the lack of vram is compensated* by RAM (or swap file). You can use ComfyUI to run this locally.
*(NVIDIA Control Panel -> Manage 3D settings -> CUDA - Sysmem Fallback Policy -> Prefer Sysmem Fallback)

Still nothing on the site hours later.

@Wizer Can I run AF on my GTX 1650 4GB?

@Stinkek 
I think it will work, but the speed will be very slow.

You can try running the model in Google Colab (may have to create a swap file)

https://colab.research.google.com/github/comfyanonymous/ComfyUI/blob/master/notebooks/comfyui_colab.ipynb

We're working to Bring AuraFlow to the Civitai Generator - watch this space!

Currently trying to generate something locally, in ComfyUI (I grabbed the wokrflow from here and set it to use v0.3 instead of v0.1 because that's what I went with). I enabled Prefer System Fallback.
My GPU has heated up to 80-82 Celsius, CPU (32 GB RAM) is using ~50% of its power. I wish this was like Fooocus where you can watch generation process, I don't even know if whatever I'm doing is worth it.

Edit: it's gone on for so long my GPU now has a major drop in performance (30% and less) while its temperature stays the same.

Edit 2: it finished! I hit Queue at 8:36 PM and I got the result at 9:20 PM, so the process took slightly under an hour. That's considering I didn't change any settings (kept 25 steps and such).
For comparison, generation on SDXL in Fooocus with Speed preset (30 steps) took me ~6 minutes. I don't use any other preset: Quality is too slow and isn't worth it for my oddly specific goals, the faster options are too poor in quality and don't support negative prompts.

@Wizer I can generate using AuraFlow on my PC, but it's not viable as it takes almost an hour to generate a single image. Even if I double-check my prompt the result may not be worth the time.

I'm now looking at the Colab page and I don't know what to do with it at all, I only worked with RVC Colab before, and that was with step-by-step guidance. Also is it pay-to-use?

Maybe I should just wait for Civitai generation.

@Stinkek Google Colab can be used for free. I have not been able to reconfigure CUDA to use RAM, so there are some limitations in working with AuraFlow. There is enough VRAM to run the base model, but not enough to use VAE at the same time.

You can save the latent using the SaveLatent node and then restart comfyui, or use locally installed comfyui to retrieve images from the latent. To load the latent, you need to use the LoadLatent node, first moving comfyUI_00001_.latent from “comfyUI/output/latents” to “comfyUI/input”.

To use Comfyui in google colab you need to run:

1. Environment Setup - this will install all the necessary components

2. Checkpoints - this will download the necessary models, by default it is sd1.5, you can remove the “#” symbols under the necessary links, and also add the necessary links, in our case link to AF

!wget -c https://huggingface.co/fal/AuraFlow-v0.3/resolve/main/aura_flow_0.3.safetensors -P ./models/checkpoints/

3. Run ComfyUI with cloudflared (Recommended Way) (to restart the program run step 3 again).

In the console you will get the following text:

“This is the URL to access ComfyUI: https:// ... .trycloudflare.com.”

Important: the runtime environment must use T4 GPU (should be set by default, but double-check just in case).

(And, of course, we look forward to Civitai generator support for the AF.)

@Stinkek not sure if you know this now but efficient nodes whatever they are called has an "efficient ksampler which shows the generation as it happens. workflows always try to be as basic as possible. i notice even with professionals that they have a ridiculous amount of steps that a efficient ksampler would show half are doing nothing. see it all the time