The new FLUX model is yet far from practical reach for lower end GPU. Even though, the community made some insane improvements to get it running in a near proper way. ~20s per image. It is still very impressive, to see the capability of the model to follow the prompt given. Still, it is to mention, that the FLUX model is not very versatile – as the output is very stereotypical, sexist and boring out of the box. With the raise of retrained models and LoRas, we will sure see some new dimensions.

amateur photography of casual long haired man in distressed inflated old oversized black punk leather jacket with holes and oversized damaged pants lying down on the floor sleeping, asymmetrical, wild unwashed dirty fatty hair, undercut, diy tattoos, realistic skin with wrinkles, looking to the ceiling, „ANALOG RITES“ in wild fat drippy wet paint wiggly oversized red hand painted letters on the back of the jacket in white ink, a lot of crazy wild white handdrawn lines are all over the jacket , dirty room with dirty white walls, hundreds of crazy chaotic handdrawn fat black lines painted all over the walls in background, a red damaged sofa, paint on the floor, fluffy white smoke and fog, 19mm lens, vignette

pose variations

Variations for the body pose only in the prompt seem to work pretty good, when stereotypical.

img2img workflow

The img2img workflow is one of the most powerful features of working with generative image tools. A pretty basic workflow without masking for the lowend NF4 models looks like this:

DOWNLOAD WORKFLOW: basic_img2img_flux_nf4

With variation of the denoise value, we can set a distance how far we can vary from the original. Denoise needs to be pretty high to see variations in this workflow 🙁

sources and links

Again Sebastian Kamph did some very nice tutorial video for installation and downloads.