How To Convert 24 GB FP16 Fine Tuning / DreamBooth FLUX Checkpoints into 12 GB FP8 Checkpoints Tutorial (Patreon)

As you already know, FLUX Fine Tuning / DreamBooth is yielding the very best results compared to FLUX LoRA training. Here a detailed public post about it : https://www.patreon.com/posts/113970485

When you look the above post, you will see Fine Tuning / DreamBooth always yields way better results than any LoRA training.

Perhaps the biggest disadvantage of Fine Tuning is that its saved checkpoints are FP16 thus each one is taking 23.8 GB HDD space

But with using Kohya GUI, you can convert FP16 models into FP8 and there is no visible quality loss - the size gets down to half 11.9 GB

The first image above shows how to convert into FP8 and the second grid image shows there is no visible difference

Check all the images

And here how to do full proper Fine Tuning / DreamBooth as low as with 6 GB GPU : https://www.patreon.com/posts/112099700

Left one is FP16, middle is FP8 made with CUDA, right is FP made with CPU - no diff between CUDA and CPU