Experiments fine-tuning Flux
I finally had time to try out fine-tuning on product images in the new Flux image generation model. Wow. 😳 I benchmarked it against tests we did with SDXL in November, with the same dataset. It's now pretty close to perfect at matching the real product photos. There are some subtle give-aways if you squint.
These images were generated with a fine-tuned model with about 20 images of the Porsche 911 Dakar. This car is not well known in the base Flux model, so it's easy to tell that the fine-tune is working.
An example prompt "A white car on a very white concrete floor. White background and white walls. Studio photo. Front 3/4 angle showing the front and left side, from low down."
Below are some comparisons against real photos, tests from November with Stable Diffusion XL (SDXL) and tests from today with Flux.1 Pro (fine-tuned).
A bit more experimentation with Flux and the penny dropped that not only can I have a fine-tuned model (to generate realistic Porsche 911 Dakar images), but I can also prompt it to add text.
Meaning... product poster designs just from a text prompt!
The prompt was "A advertisement features the text "We scienced the sh*t out of this". Below, an image of the car is displayed, driving on mars, vast landscape, with red dust being kicked up. Dark sky, stars, small blue earth in the background."
Disclaimer: for the fine-tune to work you need a good dataset of product images. Not a lot of retail products have a good enough variety of photos as you need to make this work well.