• vegeta@lemmy.dbzer0.comOP
    link
    fedilink
    English
    arrow-up
    1
    ·
    6 months ago

    Copying here cause my server is ipv6 only

    The decartunizer is a workflow for Comfy-UI that transform 2d images into more realistic images, that was posted on reddit by user mr-asa.

    At first I thought it would be achievable with img2img on stable diffusion, but no, this workflow has 133 nodes. That is a lot.

    After installing it with a clean Comfy-UI, several kinds of nodes were missing. At first I was hunting for the add-ons to install them, but then I found Comfy-Manager, that has a repository and a button to intall all missing nodes. It’s easily googleable, but here is the link for it: https://github.com/ltdrdata/ComfyUI-Manager I had to install control net and a few checkpoints and loras too. Not the easiest tool to install if you’re not familiar with Comfy-UI, but eventually I got it working.

    Then, my first attempt was to decartunize one of my kid’s drawings:

    First result wasn’t good, without modifying the workflow, I get things like this:

    That’s a consequence that most fine tunes and things done by the stable diffusion community are heavily biased towards drawing girls, for some reason.

    I examined the workflow and saw that 2 “Joytag” nodes, that try to guess a prompt to generate a similar image was generating bad results. I replaced them with a fixed text that is closer to what I wanted: “rectangular creature, masterpiece, best quality, high quality, extremely detailed CG, intricate details. dark, contrasting background”.

    It got weird:

    Then I changed all checkpoints to the base stable diffusion 1.5, and got my final result:

    decartunized

    These were simple changes compared to the orginal workflow, but if anyone wants to look at it, it’s this one