As you can probably expect, I'm currently porting my model to FLUX.
As a small teaser and appetizer for what is to come let's enjoy some ATJ in the mean time!đ
Use trigger word "ohwx woman"
ComfyUI Flow i made the images with
User lora via Lora Loader Node with a strength between 1.0 and 1.5
Training FLUX is a little bit more ressource intensive than SDXL is, so any kind of support is appreciated
â¤ď¸â¤ď¸ Support Here â¤ď¸â¤ď¸
Don't forget to âââââ my model if you like it!
Discord
Description
Init
FAQ
Comments (35)
Quality is insane
FLUX is amazing. Can't wait how the big finetune will turn out!
@pyros_sd_models Is so good,i can't imagine The noise your works Will dođ
Do you know if its possible to run a lora using the NF4 Flux checkpoint? I'm seeing in some places that it isn't possible yet, but that is more in regards to Forge's side of things, while I'm on ComfyUI.
When I learnt something the past 20 years is that the words "impossible", "never" and "always" don't exist in the realm of computer science research.
Is it CURRENTLY not possible? seems so, but someone will figure it out I'm pretty sure.
@pyros_sd_models You...have a point there. Clever response. There will be some type of breakthrough sooner or later.
Thank you for posting your workflow, your gens are beautiful.
Hi, how much vram is needed to train flux lora?
I rented 48GB VRAM, and it used 80% of it.
With quantising you SHOULD be able to train on a 4090 too, but I didn't try it so far
I'm training at 20GB with quantization on my Tesla P40. --lowvram is not enabled. I'm using the ostris/ai-toolkit script
Thank you for sharing this awesome Lora! I'm stuck running the comfyui, where can I download the t5xxl_enconly.safetensors encoder?
t5xxl_enconly.safetensors is just the 8bit version of t5 renamed
just use the whatever t5 you have, and your computer can handle
@pyros_sd_models Thank you for your answer :)
had the same problem, thanks
several questions immediately arise - how many photos were in the dataset, the resolution of the photos, what technique the training took place on, and how long it all took! and yes, and most importantly, what exactly was the training?
According to the reddit post on this he used the first 50 images he found on google images. No duplicate filtering, no size filtering, no quality filtering. It was trained on an A40 48GB on runpod with SimpleTuner. Took 4 hours (2 bucks) to train. Dataset was 50 images.
Copied and pasted from the reddit post.
https://www.reddit.com/r/StableDiffusion/comments/1eqrb22/pyro_flux_atj_lora_pyros_nsfw_model_flux_version/
what the other guy said. also you can find my SimpleTuner cfg files on the discord in the "support" channel
@pyros_sd_models hey bro the mods are idiots wtf why did they remove your post
@pyros_sd_models i cant find you in discord bro nor the files
Likeness goes out the window with full body shots.
yeah the dataset was mostly potraits, so potraits is what it's good as. FLUX has general issues with full body shots tho and is basically a resolution issue
I will look into it and try to fix full body. In the meantime please use my ComfyUI workflow, because the upscaler will restore some of the likeness
Use the lora in the face detailer, that'll help.
What? No Emma Watson first? Disappointed!
(I'm joking obviously)
and if you don't get the joke it's because there are probably more Emma Watson loras than stars in sky...
What about Taylor ? LOL
ATJ > Watson any day of the week. One is actually a good actress, the other one will always be Hermione and no amount of fragrance advertisement is going to change that.
Hey man why don't you compile that dataset for us all to enjoy bro!! Share it back here <3
Man, I just keep getting "Lora Key Not Loaded" within my command window as I'm trying to use this. The end result is nice, but not ATJ.
I just fixed it, I find out that the lora is not loaded, so I uninstall and re-install the entire, python, pytorch, and cuda, not because they are out dated, it is because loading the lora needs an older version
@aiamonlylove I didn't realize that the Lora would require an older version of Python, Pytorch, and Cuda. What versions are you running?
I can produce a similar picture, everything works fine, but the face doesn't work, what could be the problem?
I just fixed it, I find out that the lora is not loaded, so I uninstall and re-install the entire, python, pytorch, and cuda, not because they are out dated, it is because loading the lora needs an older version
What am i doing wrong? My Flux isn't using any Lora at all?
I just fixed it, I find out that the lora is not loaded, so I uninstall and re-install the entire, python, pytorch, and cuda, not because they are out dated, it is because loading the lora needs an older version
Yeah, the same thing happened to me and I needed to do a full reinstall of everything. I kept the downloaded models, clips, lora files, and VAE -- Just had to reinstall the ComfyUI, Pytorch, and Dependencies.