Soft Watercolor Style LORA
Flux Guidance used: 4.0
Sampler/scheduler combos used: Euler/Normal or Euler/Beta at 20 steps ; or DDIM/DDIM_Uniform at 30 steps (my favorite).
Model used: Flux Dev.
No trigger word needed.
Strength: 1.0 works fine, 0.5 - 1.5 is also OK if you want to adjust the style strength.
Pro tip: After generating 700+ images, without mentioning the artistic style, only the subject+composition, I've got not a single watermark or signature with Flux Dev!! I've only mentioned once the "watercolor style painting" in the prompt, and boom - I've got an artist signature. So use prompts without describing the artistic style itself. Strength 1.0 works just great!
How this LORA was "born" :
Recently, ChatGPT image generation caused a big boom in the world of AI. So I've decided to collect enough training data to create my own LORA.
I aimed to soft watercolor style, and used ai-toolkit as a training software on my private cloud GPUs. After more than 24 hours of generating soft watercolor images of various subjects i had more than 350 perfect images to train my own LORA. First i've tagged them with my comfyUI workflow (by using LLAVA and some file manipulation techniques to output the same txt file as the image has).
The result is available to download publicly, I am pretty satisfied with it :)
I will appreciate if you will share your creations and thoughts about this LORA.