CivArchive
    Preview 852744
    Preview 852733
    Preview 852740
    Preview 852735
    Preview 852737
    Preview 852736
    Preview 852739
    Preview 852732
    Preview 852730
    Preview 852741
    Preview 852743
    Preview 852748
    Preview 852746

    MeinaPastel aims to make illustrations with a 2d feeling to like pastel or colorful images!
    I have a discord where you can share images, discuss prompt and ask for help. https://discord.gg/meinaverse
    我有个可以让你分享图片和参与讨论与询问问题的discord群。

    I also have a ko-fi and Patreon page where you can support me or buy me a coffee <3 , it will be very much appreciated:
    https://ko-fi.com/meina and https://www.patreon.com/MeinaMix

    MeinaPastel is officially hosted for online generation in:
    - Mage.space ( with animate feature )
    - SeaArt
    ---------------------------------------------
    Recommendations:
    Sampler: DPM++ 2M Karras/SDE Karras/euler a: 20 to 40 steps.
    Resolutions: 512x768, 512x1024 for Portrait!
    Resolutions: 768x512, 1024x512, 1536x512 for Landscape!
    CFG Scale: 4 to 11.
    Hires.fix: R-ESRGAN 4x+Anime6b, with 15 steps at 0.4 denoising.
    V6(Pastel) Hires.fix: Latent (Nearest-exact) with 10 to 15 steps at 0.55 denoising.
    Clip Skip: 2.
    Negatives: '(worst quality, low quality:1.4), monochrome, zombie, (interlocked fingers)'
    ---------------------------------------------

    Description

    This version have Baked VAE
    ( all of the images in the samples are upscaled by 2x using SDwebui extras tab.)

    If you have any questions related to the model or its samples, how to get started or just want to share your creations, feel free to join the newly created discord: https://discord.gg/F6yeQtEQ98 <3

    V5 is an update to fix weird behaviors that V4 had, specially in higher resolutions and increase the anime style!

    ---------------------------------------------
    Recommendations:
    Sampler: DPM++ 2M Karras/SDE Karras/euler a: 20 to 40 steps.
    Resolutions: 512x768, 512x1024 for Portrait!
    Resolutions: 768x512, 1024x512, 1536x512 for Landscape!
    CFG Scale: 4 to 11.
    Hires.fix: R-ESRGAN 4x+Anime6b, with 15 steps at 0.4 denoising.
    Clip Skip: 2.
    Negatives: ' (worst quality, low quality:1.4), zombie, (interlocked fingers), [monochrome:0.8]'
    ---------------------------------------------

    FAQ

    Comments (7)

    smoonHackerMay 20, 2023· 16 reactions
    CivitAI

    How in the actual f... Im gonna render these in 4K??! My 12GB cannot make it!! 😫

    Meina
    Author
    May 20, 2023· 2 reactions

    4k? o.o
    the max i can gen is 512x1024 and use hires.fix 2x, then for the samples i use the extras tab of SD webui to 2x again as civitai likes higher resolution more to be sharp even when not click in the image

    smoonHackerMay 20, 2023· 2 reactions

    @Meina Hooo! I Never thought that the HiRes2x combined with the 2x gain of SD could get that sharp! I'll try it and post a review. Thanks!!

    onepiecefanMay 21, 2023· 4 reactions

    I would use tiledvae extension. Lets me do hires multiplier of 3 on an 8gb card. Probably higher if I reduced the tile size. From their you can do tiled-controlnet with ultimate upscale to go as large as you want.

    orx_ibxMay 23, 2023

    For 4k images, I usually do low-res initial image (512x768) > 2x Hi-res upscale > 2x SD Upscale at low denoising strength (0.2-0.4).

    CMDRAKENOMay 30, 2023· 1 reaction

    simple start by making 960x540 then use the ultimate SD upscaler

    SwordjwhJun 9, 2023· 2 reactions

    You can try to use the high-definition repair function, but it will delay the calculation time, and there will be certain side effects (such as drawing strange lines), my 1060 3G can be enlarged from 960*540 to 1920*1080, your 12G video memory may be able to Larger, the algorithm recommends Latent (nearest-exact), SwinIR_4x, or ESRGAN_4x, Latent (nearest-exact) has less pressure on computing power, but the effect is not as good as the other two, which is different from the built-in zoom function , the images they run out are clearer and more natural (not as unnatural as directly zooming in)