This Lora is made for a semi-realistic Niji style with high detail and textured fur The dataset is mostly furry art, but it handles humans just fine too<h3 id="v3:">V3:</h3><ul><li>The dataset has been slightly adjusted</li><li>Network dim has been increased</li><li>The style effect is much stronger</li></ul>This is probably the final version, since I’ve achieved the exact style I was aiming for.I recomend using hiresFIX to get much more refined textures (the key advantage of v3 is the textures, they’re significantly more detailed).I recommend using it together with other style LoRAs (especially the ones listed in the suggested resources).Example of this LoRA’s impact on style (V3, same seed, prompt, etc.):<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/797b0d03-f91a-41d7-94ce-f8acfb52d27d/width=525/797b0d03-f91a-41d7-94ce-f8acfb52d27d.jpeg" /><h3 id="v2.5:">V2.5:</h3><ul><li>The dataset has been expanded 2.5× compared to v2.</li><li>Improved generation stability and higher flexibility (thanks to the new dataset, v2 was a bit overfitted).</li><li>Around 50% of the dataset consists of humans and backgrounds, the rest is furry content.</li></ul><h3 id="v2:">V2:</h3><ul><li>The dataset has become more diverse.</li><li>The dataset now consists of approximately 40% humans, 60% furry.</li><li>The style has become more stable, but I recommend using this LoRA with other niji LoRAs for better results (see Suggested Resources). </li></ul>Works well with Euler a + Exponential/Karras.For stable results, strength 0.8-1 is recommended<a target="_blank" rel="ugc" href="https://tensor.art/models/911051824981614385?source_id=kT61oVzqkkm6o_ooYH7y8hMj">Tensor.art_link</a><span style="background-color:rgb(21, 21, 23);color:rgb(249, 250, 251);font-family:quote-cjk-patch, Inter, system-ui, -apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Oxygen, Ubuntu, Cantarell, &quot;Open Sans&quot;, &quot;Helvetica Neue&quot;, sans-serif;font-size:16px">for V1: If you want softer shades and "fluffy fur", based on my experience, it's best to use NoobAI V-pred models (example - the 3rd image in the post)

Files with SHA256: 49f364e4c37e04265e914df37689cf67158dd399852ff6a99d01bc6c0c712e13