Hey everyone, this is Luci! Like many here, I am fascinated by Flux and what it entails so I thought I'd share my findings.
Just to note also that this model is 100% synthetic. I made 400 images with an SD1.5 checkpoint, cut them down to 25 and trained an embedding on that back in the day. This model was made with 21 images upscaled from that embedding with a size of 1024x1024.
I trained this with 2,000 steps with Dev (earlier wrote Schnell here but looked back at the script and despite inputting Schnell it seems it trained on Dev (boo). I run the LORA with Schnell and use 6 steps to get the best outputs. It also works pretty great with Dev but I'm all about those Schnell possibilities. All images shown are with Schnell.
Thought I'd share with y'all as LORA training for Flux is relatively new and by sharing I'm hopeful we can all improve!
Description
FAQ
Comments (11)
Very good! Thanks!
what did you use to train the lora?
trained it on replicate, look for the ai-toolkit. Takes about an hour so $5 or so.
@nudesai There's no need to hide my comment. I was just saying that it didn't work for me. The best thing to do is to ask yourself why it doesn't work for everyone.
@nokai I'm not here to guide you how to use it. You can either do that yourself or not. I know it works hence all the images I have made with it. If you can't get it to work, then I don't think that is worth just reviewing it negatively but I'm not getting into an argument on this. I'd rather hide and move on. I'm not providing a troubleshooting service.
Hey, it works really well. What did you use to train the LoRA?
Welcome back! I liked the ones you made for 1.5, even though they were pulled from CivitAI.
Anyone to clarify?
I take the images as workflows. They try to load "flux1.snell.fp8" as a checkpoint. In the install of that model; they specify to put that safetensor in unet directory. Therefore it does not load as a checkpoint..?
Then I change the checkpoint as any good one and I get images, even very good, but nooone of them look like Luci ...
Any thought?
if you are using the fp8, it just goes into the model loader, u dont need to separate the encoder etc, load that model and where you were putting the t5 etc, rewire just from checkpoint loader and then add this as a lora into the workflow.
Will try that, thanks
Would love to see you add natural pubic hair and more realistic vaginas to this