Usage

With the Fluid Art LoRA for Flux you can generate many different kinds of fluid art including text, clothing changes, wall art, textures and painting that resembles acrylic pour.

With this model it can be tricky to find the right strength, so play around until you find a value that works. It's best to keep the seed static when experimenting, but values ranging from 1.0 to 2.0 can produce nice results. It's also worth checking out the 1800 step model available at the same link above, this model works well at higher strengths.

The trigger words "fluidart style" are typically required for any good result, although chaining multiple copies of the LoRA in series can also produce stronger effects. In many tests I've used (fluidart style) to increase the strength of the trigger word on the output. If you find something that works well for you while using this model, please share it! 🙂

The ComfyUI workflow is attached to this article called flux-fluidart-lora.json. Drag this into ComfyUI to load the workflow and edit the prompt as needed. This will replicate the cover image of this article.

Training

This model was trained using 54 fluid art images from pexels. Captions were set to the trigger word only. The tool used to train the LoRA model was https://github.com/ostris/ai-toolkit by ostris referencing this guide by alvdansen: https://civitai.com/articles/6656/how-to-train-flux-dev-on-a-3090

Most defaults were kept from the ai-toolkit example config. I trained locally using a 4090 at 3000 steps total. I saved steps 1800 and 2000 which have slightly different results. The total training time took about 3 hours to complete.