Observation about upscale

#3
by theo77186 - opened

In the description of the upscale, I've seen this config:

slices:
  - sources:
      - model: zerofata/MS3.2-PaintedFantasy-v2-24B
        layer_range: [0, 29] # layers 0 to 28
  - sources:
      - model: zerofata/MS3.2-PaintedFantasy-v2-24B
        layer_range: [10, 39] # layers 10 to 38

But the layer_range is upper bound excluded, generating a 58 layers model, with the layer 39 from the original being never represented. Is that intentional or an oversight?
I know it's not clearly documented in mergekit, but I still feel like pointing it out.
This also affects the v1.

Hey,

Good spot, it was a counting fail I'd done the first time around. I'd noticed the layer count was one off what I'd expected, so this is helpful to know the reason why. Admittedly forgot to troubleshoot it before doing the v2, where I just reused the same layer config.

Hey,

Good spot, it was a counting fail I'd done the first time around. I'd noticed the layer count was one off what I'd expected, so this is helpful to know the reason why. Admittedly forgot to troubleshoot it before doing the v2, where I just reused the same layer config.

We'll Get a remake YAY!!
I was wondering why there's gap in quality even though upscale models always perform better.

Yup, I had some time over the weekend so redid it with the missing layer.

https://huggingface.co/zerofata/MS3.2-PaintedFantasy-Visage-v3-34B

This feels a lot more like the 24B did, but smarter and more consistent. Which make sense. This unintentionally turned into an interesting experiment on how the last layer impacts MS3.2.

Thanks alot man.

Sign up or log in to comment