Model is a LoCon trained on NoobAI ε-pred 1.0 and is untested on base ILXL.Dataset is plucked from a few realistic-looking 3D sources, and then completely autotagged.As a disclaimer, this model is more of an experiment than anything. NoobAI of course has some knowledge of 3D art trained into it, but I have yet to find a way of making it consistent, let alone porduce anything approaching decent results. The aim of this LoRA is to try to skew it to something more usable.<h2 id="v2.0-(latest)-x49eoynai">v2.0 (latest)</h2>Style model based on 3387 assorted images of realistic 3D art.Completely retrained on a new dataset, with each artist now tagged for reproducibility (see list below and/or XYZ comparison in gallery). Artists were selected to give as broad a range of concepts/subjects as possible.<h3 id="usage-2a0dayofo">Usage</h3>BasicWorks well anywhere between a strength of <code>0.8</code> and <code>1.0</code> Base trigger is <code>3d</code>Helper tags are <code>realistic</code> and/or <code>photorealism</code> Subjectively, <code>dramatic lighting</code> adds some nice magic to images.Artist Styles24 different artists have been scooped up and tagged for, these can be accessed with varying degrees of success and watermarking (see gallery for examples):<ul><li><code>apone3d</code></li><li><code>ascaniololololol</code></li><li><code>batesz</code></li><li><code>billyhhyb</code></li><li><code>bulgingsenpai</code></li><li><code>checkpik</code></li><li><code>cosmiclife</code></li><li><code>drdabblur</code></li><li><code>dzat</code></li><li><code>ecksoh</code></li><li><code>eupharrow</code></li><li><code>integlol</code></li><li><code>johncooperbeast3dx</code></li><li><code>leeterr</code></li><li><code>littleslice-sfm</code></li><li><code>m71z30</code></li><li><code>masterkabal</code></li><li><code>nagoonimation</code></li><li><code>notill</code></li><li><code>oscillator</code></li><li><code>radianteld</code></li><li><code>shizzyzzzzzz</code></li><li><code>snow_kitsune</code> (obvious tag conflict with <code>snow</code>)</li><li><code>yuni</code></li></ul>Issues<ul><li>Watermarking: Quite a bit of watermarking comes in with some of the artist tags, and can be difficult to prompt out without turning down the LorA strength and thereby losing the quality.</li></ul><h2 id="v1.0-d7fhs4fbo">v1.0</h2>Style model based on 1787 assorted images of realistic 3D art. Unlike v2.0, artists are not tagged, so the style is what you see is what you get.<h3 id="usage-fu4aoxabv">Usage</h3>BasicWorks well anywhere between a strength of <code>0.8</code> and <code>1.0</code> Base trigger is <code>3d</code>Helper tags are <code>realistic</code> and/or <code>photorealism</code>Issues<ul><li>Character tags upset style: As is a common issue with the ILXL/NAIXL base models in general, character tags tend to be a little overtrained and can override a style LoRA (see the Pharah/Widowmaker images in the initial gallery). You can try turning them down, e.g. <code>(pharah:0.7)</code>, but it can be pretty difficult to work with them.</li><li>Weird Lighting: For some reason, the lighting can go a bit weird at <code>1.0</code> sometimes, so you might want to try lowering it a bit. I think this may be due to the fact that, as good as <code>wd-swinv2-tagger-v3</code> is at detecting the more visible concepts in an image, it doesn't really have a clue about tagging lighting etc. Manual tagging would hopefully improve this, but... it's almost 2,000 images <a target="_blank" rel="ugc" href="https://emojipedia.org/smiling-face-with-tear">🥲</a></li><li>No dataset: As a corollary to the above, I would usually post the dataset for a model I upload, but given the number of images sadly it won't fit on civit.</li></ul>

3D Realistic (24-in-one) [Multi-style] (NoobAI)

<ul><li>trained on 1787 auto-tagged images on NoobAI ε-pred 1.0</li><li>single combined style only</li></ul>

Files with SHA256: 9824747ee2d48cbe1d3c130adc74d5bcc61fd88d030eaafe8930b1815c42f616