The project is closed. Thanks to the 10 people who supported the last two models, 6,500 buzz's. This was enough to train three LoRa models, which is, of course, impossible for further model training.
Description
FAQ
Comments (14)
Please add Q4-8 GGUF or FP8 version!
Not all of us can run big models bf/fp16 :(
Hello, I mainly run on a 4070 laptop (8 GB), during work hours on a laptop with DDR5 48 GB, I get 2.5 s/it at a resolution of 1400x1800.
I'll try to make a pruned FP8, but if the quality suffers too much, I won't add it.
Hi, I tested the FP8 model, the anatomy suffers, the quality does not improve in any way compared to the regular ZTurbo, there is no point in releasing it, sorry.
@AKDesigns i made an open source program that makes it quick and easy to make your own GGUF's, even if you have low RAM https://github.com/qskousen/ggufy
This is great, I have a 5090 card, could you do the full 20 gig version at some point? I have 32gig Vram.
It follows instruction so well at just 11gig, I can only imagine what a full size 20 gig could do.
Nice work and thank you for making this.
Thanks for the feedback. I'll see what I can do at home on my 5080. I hope the system can handle such a large model.
Hi! I tested the full FP32 and there's absolutely no difference with the FP16, not even by 1 pixel. The noise is identical.
@Viennar, @pursuit_of_beauty You would have to train it in FP32 to get any benefit right? Otherwise it's just representing the exact same weights with more bits
@psychologicau Thank you, that makes sense.
If you have a 5090, you should learn how to train models... gg
Brilliant checkpoint. Great results, good variety and plays nice with Lora's
Thank you very much, I'm trying, there's still some anatomy to correct, I hope something else will come to light along the way.
Hi, can you share the text encoder and vae in your workflow?
Hello, VAE from the most common FLUX, encoder Qwen3_4b.