Has someone implemented this? I want to try out this shift
idkman
What I dislike about lower quantization is quality degradation. In my small experience, i find 7b models dumb (I've only tested Q4KM GGUF), and needed to be provided proper context before moving forward with the constructive conversation (be chat or instruct).
If this issue can be circumvented in lower quantization, I'm all in.
In context of SD, going below fp16 would only make things faster at cost of quality, and I personally like to go in depth with my prompts. For simpler prompts sure, even lighting and turbo are good in that regard.
Haha no. I mean in the context of providing in control, DPO does a better job than FreeU.
https://www.youtube.com/watch?v=mrKk0yYUdnU
The paper mentions it manipulates main net features(B) and skip features(S).
Not sure what the main net features are, but I'm assuming skip features are like clip skip?
Although the paper mentions S didn't do anything, in my tests (sdxl base), it had a tremendous amount of adverse effect on the image generated.
TBH, i enjoy using DPO more than freeU, but I'm curious to learn.
references:
main repo: https://github.com/ChenyangSi/FreeU
https://github.com/lllyasviel/Fooocus/discussions/618
FreeU: Free Lunch in Diffusion U-Net Explained: https://piped.privacydev.net/watch?v=eFmkJ_oEW5s
This official tutorial on HF mentions the unsatisfactory defaults https://huggingface.co/docs/diffusers/v0.22.1/en/using-diffusers/freeu
Forge is wayyy ahead! I got annoyed by poor memory management on 1.8.0 and found Forge. I'm loving it
Then, Proton it is
Thanks! Much appreciated
@even_adder how do you keep up with upcoming progress in machine learning?
Thanks to the pinned thread by db0, which got me started into all of this, and now I'm intrigued what this space has to offer to an individual.
Mullvad or Proton
I had a really hard time installing the tensorRT extension for 1111