• CavendishOPM
      link
      English
      23 months ago

      Agreed. Still looking for the right mix of anatomy finetuning models.

  • @[email protected]
    link
    fedilink
    13 months ago

    Are you using Lazy Mix Real Amateur or a different checkpoint? It looks like one of the better tuned smaller SD1.5 models, at least I assume it is small if you are doing some kind of branding.

    I’ve been playing with Flux, which makes old SD stuff look dated, but it is so freaking slow with only 16 GB of a GPU. The best I have at the ~2GB mark is Lazy Mix Real Amateur V3b at 2.1. I’m curious if you’ve undercut that one with a smaller model.

    • CavendishOPM
      link
      English
      23 months ago

      Not using a small model, this is the full 22 gigabyte Flux.1 Dev fp32, plus a couple of customized lora merges that I’m experimenting with. I have a 3090 with 24GB VRAM, which gets maxed out. Takes about 1:30 to 2:10 to generate each image at 2 megapixel resolutions. It’s a real basic flux workflow in Comfy, I don’t even run them through any upscaling nodes.