CivArchive
    Colorfixed NoobAI-XL (NAI-XL) V-pred - v3.0
    NSFW
    Preview 90068482
    Preview 90081052
    Preview 90081048
    Preview 90081064
    Preview 90081084
    Preview 90081072
    Preview 90081056
    Preview 90081069

    No styles, no detailers, raw and fixed.

    This is base NoobAI-XL v-pred 1.0 with my colorfixes merged in at a specific weights.

    This is v-pred SDXL model. Read guides on what UIs support it and how to use it.

    Making it was a tug of war between desaturation and oversaturation. Sliders trainer modified to run v-pred, all versions of torch and FA from 2.3 installed, 70+ loras baked, constant back and forth and stuff not working as I thought it would.

    But now everyone can see what I found in this particular model and why I spent all this time fixing it.

    This model inherits original's license.

    This description will probably be updated later.

    Description

    Fixed excess greens, bumped colors and added details a bit.

    FAQ

    Comments (10)

    darkspaceinfinityOct 17, 2025
    CivitAI

    Would this be better to train on then original noobaixl vpred?

    Volnovik
    Author
    Oct 17, 2025· 1 reaction

    No idea. I thought someone would test it, but no one cares. I'll probably try finetuning it to fit small bits further, but I'm fully occupied on irl stuff right now and probably wont be able to start till december.

    darkspaceinfinityOct 18, 2025· 1 reaction

    Well I tried it with a lora and the results were okay, good prompt adherence but oddly enough more washed out colors then when trained on base noobaixl vpred. This was on a mixed dataset of greyscale and colored pictures so maybe that was the problem or my generation settings. Is DDPM the only sampler you recommend? What about schedulers? Would steps or cfg need to change when implementing a lora?

    Volnovik
    Author
    Oct 18, 2025· 1 reaction

    @darkspaceinfinity what ui do you use? Any specific settings?

    darkspaceinfinityOct 18, 2025· 1 reaction

    Wow, feel like a idiot, just read your article on it. Thanks for your hard work.

    Volnovik
    Author
    Oct 18, 2025· 1 reaction

    @darkspaceinfinity just remove Combat Cfg Drift, I stopped using that with this model, maybe that one is desaturating. Also tag dataset properly, monochrome should be mentioned specifically. This can mess up results, vpred seems to be a lot more sensitive. And don't forget vpred settings for training.

    darkspaceinfinityOct 19, 2025

    Any opinion on these two photos? I feel like Euler looks better, but maybe I'm wrong or messed up somewhere.

    https://civitai.com/images/106640844

    https://civitai.com/images/106640213

    Volnovik
    Author
    Oct 19, 2025

    @darkspaceinfinity First one is not found. Other one is cfg++ which is the route that I dont use.

    darkspaceinfinityOct 19, 2025

    @Volnovik Sorry, hopefully this one works. https://civitai.com/images/106651409

    Also updated cfg one. I used resharpen on the DDPM one, but still looks worse then CFG, am I missing something?

    https://civitai.com/images/106651409

    Volnovik
    Author
    Oct 19, 2025

    @darkspaceinfinity Now all other images are deleted. Just do a comparison using prompt x/y/z like this one https://civitai.com/images/62286290

    Also I recommend to stop using hires fix

    Checkpoint
    NoobAI

    Details

    Downloads
    790
    Platform
    CivitAI
    Platform Status
    Available
    Created
    7/23/2025
    Updated
    5/3/2026
    Deleted
    -

    Available On (1 platform)

    Same model published on other platforms. May have additional downloads or version variants.