Aesthetic
Lora for FLUX
Update 2: Tried using the kohya trainer. The result is the 10R 2E CDO-0.5 version. Previews use my impressionism LoRA. It isn't needed but I liked the results so I stuck with it. Workflow and prompts attached to previews. Just open them in ComfyUI.
Update 1: Added the stronger 10k step trained LoRA model with direct comparison images (left is LoRA strength 0, right is LoRA strength 1.2). Removed misleading (anime) from name, since the LoRA doesn't really make images more anime-like.
Recommended strength to use really depends on your workflow!
Purpose: Imparts a more painterly and wild look at the cost of some coherency. It was trained on painterly vibrant anime art that didn't concern itself with fine details.
This is a proof-of-concept style LoRA for Flux. And only a proof of concept!
NOTE: For some prompts it will have a dramatic effect, for others… nothing, probably needs a bigger and a more diverse dataset then the one I used.
Using the following workflow: [Simple+Advanced] FLUX.1 ComfyUI Workflows - LoRA Support v2 (No CLIP) for ComfyUI
Description
FAQ
Comments (11)
Can you share what tool and parameters you used? This is really good... and what GPU did you train on?
so it has begun...
i think you used diffusers model for training , i'm in middle of using SimpleTuner which it very close x-flux but for more general usage , theres an option to use local model , i want to give atry to use this merged model https://civitai.com/models/629858 , please need your advice
With my workflows I am getting far more anime results, the guidance makes semi 3D stuff at 3.5, and more anime at 5 to 10
The strength needs to be lower, about 0.6 to 0.7
Thanks for the LORA!
Sounds like we need to prioritize the ability to run Flux with LoRAs :P
Will it work with Flux.S as well? Seems more reasonable licence-wise for Civit?
Why suggest strength 1.2?
In my (limited) testing I've so far come to the conclusion that it does its best work at 0.6 to 1.0 strength and the results grow erratic above 1.0
add furry lora>????
can use with forge ui?
can't we just quickly convert all the SDXL loras to FLUX?