Furry/anthro artist <a target="_blank" rel="ugc" href="https://e621.net/artists/18891">Dredjir.</a> Made for Anima.<hr /><h3 id="recommended-weights">Recommended Weights</h3>V2.8 and V2.1: Weigh between 1-1.3V2.5: Weigh between 0.7-1V2 and V1.8: Weigh between 0.8-1 for best results.<s>V1.5 and V1: </s><s>Weigh between</s><s> </s><s>0.6-1</s><s> </s><s>depending on how deformed or deep fried your images are.</s><hr /><h3 id="helpful-tips">Helpful Tips</h3><ul><li>When generating feet in general, it is usually partially obscured, either hidden by an object or one feet is hidden by the other feet, so adding <code>partially obscured feet</code> in negative prompt typically helps.<ul><li>Note: V2.1 and later versions don't have as much problems with this, but you can still try it if it still persists.</li></ul></li><li>If watermarks and dialogue appear, even though you didn't want it, you can try adding <code>watermark, speech bubble</code> in negative prompt (I already removed all artist watermarks from my dataset, however, your checkpoint and other Loras may be forcing unrelated watermarks onto the image anyways).</li></ul><hr /><h3 id="version-info">Version Info</h3><code>V2.8</code> was retrained once again with different settings. Using my experience and knowledge, it is more flexible and saves space at the same time. Quality is similar as last version.<code>V2.5</code> is retrained again, but with different settings. It is much larger in file size, but way more accurate to original style.<code>V2.1</code> is retrained locally and is smaller in file size compared to previous versions. It has a better balance between prompt and style adherence.<code>V2</code> resembles a lot of the artist's own style and it follows prompts decently but it can still struggle with words or concepts.<code>V1.8</code><s> is one of the versions similar to the shading that you can get.</s><code>V1.5 More Shaded</code><s> is still the same as V1.5 but it has more of a 2.5D look. It does tend to look more like generic furry AI slop.</s><code>V1.5</code><s> </s><s>contains smaller dataset and still rough.</s><code>V1</code><s> </s><s>represents the overall rough artstyle, but it does seem to struggle with artist watermarks and signatures.</s><hr />If needed, DM me and I will take this model down upon request.Have fun!

Dredjir Style (Anima)

Retrained again with different settings using experienced knowledge. Much more flexible than previous versions. It is also smaller in file size.Weigh between 1-1.3

Files with SHA256: d8242a34c557eda8f4e8b3740fb2c2216af716b0b55a3967ee815e58b51ca95b