Hello!
Some of you may have been wondering where the heck I have been, and when is the next PixelWave model going to drop!! Well I thought I'd give a quick update on what I have been up to.
I have been trying to fine tune FLUX.1-schnell.
I want to make a schnell model that is not de-distilled, looks as good, if not better than PixelWave 03 Dev, to be more creative, to work well with LoRAs, and to be purely Apache 2.0 licensed. 🤯 💀
This means I can't fine tune dev at all and then transfer the learnings to schnell. I must only train on schnell. I have to be very careful not to break down the time distillation, so you can still generate an image in 8 steps.
Well after many, many failed attempts, I have finally started to see some results. Like when I train the model, it actually learns the content of the training data! 😱 Sure, the model is still fragile, and training it feels like guiding a bull through a fine china shop, but I'm making progress. I can confidently say, that yes, you can train schnell without breaking the time distillation or ruining the model. You just have to avoid certain parameters in the model *cough* time_in *cough*
So now I figured out I can actually train the freakin' thing. Hopefully in the next month I can release something I can be happy with and you can enjoy too!
Have a nice day!
humblemikey
