<h3 id="vpred-models-are-trained-with-v-parameterization-so-they-will-only-work-on-vpred-models-like-noob-ai-xl-vpred-r3mix-or-obsession-etc.-cecy9inmw">VPred models are trained with V-Parameterization so they will only work on VPred models like NOOB-AI-XL-vpred, r3mix, or Obsession, etc.</h3>This is a DoRA so make sure to have your stuff somewhat updated in the last year<h2 id="remember-to-check-out-the-artist!!!-without-their-art-these-loras-wouldn't-exist!!!-4tosc9nkn">Remember to <a target="_blank" rel="ugc" href="https://www.patreon.com/sulcate">check out the artist</a>!!! Without their art, these LoRAs wouldn't exist!!!</h2>NoobAI and Obsession could replicate their style on their own, however, the style is clearer when using a separate model like this one.<h3 id="hires.fix-usage-is-heavily-advised.-don't-use-quality-tags-to-get-more-of-the-artist-style-from-the-dora-and-not-from-the-main-model.-activation-tag-is-mandatory.-zmeefcahz">Hires.fix usage is heavily advised. Don't use quality tags to get more of the artist style from the DoRA and not from the main model. Activation tag is mandatory.</h3>Model works pretty well in my testing with character LoRAs. It does what it says in the title. In my testing, found that Diffusion in Low Bits: float8-e4m3fn the best result though the difference between auto, and bnb-nf4 was minimalHad to create a model with this banger artist since there isn't anything for NOOB/Illustrious. There was one model but I felt like I had to make an attempt myself to make another one. During training that model was removed (I guess perfect opportunity for me to upload this one). Also, did quite a lot of reading on optimizers and samplers to realize what the problem in my training software was, that didn't let me use Adafactor. Turns out it was the sampler CAWR which was broken in an update so had to switch to RAWR but the result seems the same (I think?). Thank you for your attention, Mr. Nobody! Who even reads this far?

Sulcate (Artist Style) | Illustrious/NAI-XL

<ul><li>Trained 480 images with noobaiXLNAIXL_epsilonPred11Version</li><li>Finally got Adafactor sampler to work... Turns out I had to swap CAWR with RAWR... I hated training with Prodigy</li><li>Note that the model doesn't spam speech bubbles. I simply used original tags used for training and was too lazy to modify them</li></ul>

Files with SHA256: 8b46f44426af7fca5409985db2c266fd5fedce4c8bb6d192ba97b04ab9c2f39d