• RotaryKeyboard@lemmy.ninja
    link
    fedilink
    English
    arrow-up
    7
    ·
    1 year ago

    Text generation seems to leave a little to be desired. Five separate generations and “Hello World” always came out as “Hello Word.”

  • Scrubbles@poptalk.scrubbles.tech
    link
    fedilink
    English
    arrow-up
    2
    ·
    1 year ago

    Still kinda new to this, do we download the whole huggingface directory? Just the safetensors? Do we need the refiner too? I guess, what are the steps for “optimal installation”?

    • Sleo@lemmy.caOP
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      So from what I understand, all you need is the base for A1111. If you’re using comfyUI, might want to download the refiner too.

      • Sleo@lemmy.caOP
        link
        fedilink
        English
        arrow-up
        2
        ·
        1 year ago

        Just learned that you can generate image using text to image using base model then send it to img2img and use refiner model with around 0.33 to 0.25 strength.

  • Sabata11792@kbin.social
    link
    fedilink
    arrow-up
    2
    ·
    edit-2
    1 year ago

    Seems to be the upscaler. Using an older one looks great.

    Anyone have any luck getting the model to work? Running it with an anime model an the outputs are either blurry or latent deepfried.

    rompt: cute anime girl crying, face
    Negative prompt: watermark, text
    Steps: 20 | Seed: 2436787329 | Sampler: DPM++ 2S a Karras | CFG scale: 6 | Size: 512x512 | Parser: Full parser | Model: test15 | Model hash: d01b8e6877 | Refiner: Refiners_sd_xl_refiner_1.0 | VAE: Anything-V3.0.vae | Latent sampler: Euler a | Image CFG scale: 6 | Denoising strength: 0.3 | Refiner start: 0.8 | Secondary steps: 20 | Version: 18279db | Pipeline: Original | Operations: “hires | txt2img” | Hires upscale: 2 | Hires steps: 20 | Hires upscaler: Latent