Krea2 FinalCut
INT8 is recommended for most users, unless you notice a massive slowdown when using lora's
Official FP8 or INT8 Convrot scaling/quantization
100's of hours training.
Description
FAQ
Comments (24)
hi ho and thanks. do you olan on releasing an int8 version?
What does "Official" mean in this context? It looks like a case of falsely claiming to be official, which is prohibited by the license.
It is bracketed with the fp8 scaling - meaning the model matches the official fp8 scaled version - in some cases I have used my own scaling such as blockwise
Well done - its precision is unmatched by any other checkpoint I’ve tried so far. Thx for sharing.
Thank you
Is it possible to download bf16 of your wonderful model somewhere? Unfortunately I can't use quantized versions on my setup yet.
Currently no, I may release that version at some point
@Felldude I couldn't run your FP8 Scale version on my Mac. QWEN AI says this file isn't a standalone base model. It's a Delta checkpoint (differential/residual model) or LoRA saved in full-scale format.
He's probably wrong, but I can't convert or run this model.
@Unicom Both the scaled and INT versions use comfy metadata to load the model - I am not familiar with mac to know if they are supported
@Felldude So I will wait bf16 of your model
Wow! Looks promising. Any chance to release the INT8 version? Anyway, thank you for making and sharing this.
I don't however at some point likely the next version I will release the BF16 version which is suited to quantizing to all the other formats
Changed my mind and did convrot
@Felldude Downloading now. Thanks a lot.
Can you double check the Int 8 convrot version, I downloaded it and the result are almost identical to the base model. Thank you for your great work!
You are correct, thank you I had uploaded the base.
The FP8 version download seems to also download the int8 version FYI
You will need to press the drop down menu - the int8 version is slightly larger
Wow this super fast. 13 seconds.
But conditioning control node not working with int8 loader. Results censored.
When I first uploaded the convrot model I uploaded the wrong file - if you downloaded it day one you might try downloading it again
Probably one of the most powerful NSFW‑capable models out there.
Thank you