Still working on it?
Hi,
thank you for all your efforts.
Are you still working on an INT4 solution?
Unfortunately, I can't test whether the FP4 variant yields usable results.
The Nunchaku team seems to have been inactive for weeks, with many issues with other models and their ComfyUI-nodes.
Anyway.
I'd like to know if you're still looking for a "usable" INT4 solution?
I would be very grateful.
Thank you! :)
Hello, I sorta hit a wall with the INT4 version. For some reason the quality just falls off a cliff. I can’t spot anything obvious, I was hoping I would have figured it out by now, but sadly not.
I took the “easy” path with adding Chroma to nunchaku and did everything in Python, the official implementation would have written it in C++ but that’s above my pay grade.
The code and everything is on GitHub, maybe it’ll help someone in the future. For the moment don’t see a way forward without starting over.
Thx! :)
Do you have feedback about the FP4-version?
Is it giving better results?
Hello, I sorta hit a wall with the INT4 version. For some reason the quality just falls off a cliff. I can’t spot anything obvious, I was hoping I would have figured it out by now, but sadly not.
I took the “easy” path with adding Chroma to nunchaku and did everything in Python, the official implementation would have written it in C++ but that’s above my pay grade.
The code and everything is on GitHub, maybe it’ll help someone in the future. For the moment don’t see a way forward without starting over.
Unfortunate, as this is the only attempt at a real boost in speed for ChromaHD. I think if lora's were functioning, it would offset the quality loss quite a bit. Ah well, thanks for trying at least. Many Chroma users out there with the hardware, know how, and resources who won't even lift a finger because it runs quick for them on their high end hardware.