SDXL Trigger: pamcrystal
Flux Trigger: made of crystal, made out of crystal
The best result is achieved with Flux Schnell.
Description
Support for Flux.1 Dev and Flux.1 Schnell
FAQ
Comments (10)
Can you please provide any advice on flux lora training params and optimiser? Mine one with SimpleRuner and prodigy come out way worse than SDXL one...
SimpleTuner, using this tutorial, is the current best solution: https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md.
The best training properties are, right now, "trial and error." But I must say the training of FLUX is more similar to an SD 1.5 LoRA than SDXL.
@pampelmann I tried using this tutorial to make artist style Lora, but did not succeed - prodigy is overcooked lora, dora doesn't correctly load in comfy, and adamw_fp16 lora with 1e-5 LR probably doesn't learn enoug after 4000 steps on 200 images with batch 2 =( On validation images preview become always worse and worse or almost didn't change after 1000 steps, and when I try to use - it's somehow change output, but not even close to style, even if push lora to 4.0 strenght =(
SD 1.5 was much easier to train =)
I've tried all your prompts, settings and seeds and can't get anything close to your sample images.. You're look amazing but mine are really bad.. I've tried a few different version of Flux, I've adjust the prompt weight from .50 all the way up to 2.5, different step amounts and just don't understand why the lora doesn't work for me.. I've used it in both ComfyUI and SwarmUI.. What's your secrets?? Now that I look what other users posted, none of those images look like your samples either... Very strange..
You can drag one of my images into your Comfy and use it for the LoRA workflow. I hope this helps you.
@Pampelmusi Thank you for your help! It works wonderfully!!! I posted images!
@SouthbayJay you're welcome
The Invoke Community Edition doesn't seem to recognize this model.
It looks like that the SDXL version of this LoRA modifies the CLIP text encoder in a way conflicting with the fine tuning of Illustrious XL. When using it with an Illustrious-derived base model, you will need to keep the CLIP strength very low (e.g. 0.2 thru 0.4) while make U-NET strength relatively high (> 1.2). If your generation tool doesn't allow separate LoRA application strengths, well, it may be better to avoid using this LoRA with an Illustrious model. JFYI.
Nice lora! Good work! ^^
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.