<h3 id="hassaku-moss">Hassaku MOSS</h3><hr />The model was trained on a small but dense dataset of 120k images, it still produces realistic-looking results, especially when no characters are involved or the prompt uses natural language. Adding “nose” as a negative tag helps. This is a test to see if MOSS can be trained at all (it can be trained). Look at <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">About this version for more details -&gt;<hr />Join my <a target="_blank" rel="ugc" href="https://discord.gg/zSR5FcYWWE">Discord</a> for everything related to anime models and AI art. If you’d like to support my work, you can do so on <a target="_new" rel="ugc" href="https://subscribestar.adult/citrus-models">SubscribeStar</a> Images for training captioned by my own <a target="_blank" rel="ugc" href="https://github.com/Ikena1992/VLCaptioner">captioner</a> <hr /><h3 id="supporters">Supporters</h3>Special thanks to my supporters: Estryark, SETI, Kodokuna and dataset captioners mariotheknight and tamashiicolle<hr /><h3 id="recommended-resolutions-for-moss">Recommended Resolutions for MOSS</h3>Here are some resolution suggestions:<ul><li>1536 × 640</li><li>1344 × 768</li><li>1216 × 832</li><li>1152 × 896</li><li>1024 × 1024</li><li>896 × 1152</li><li>832 × 1216 — most recommended</li><li>768 × 1344</li><li>640 × 1536</li></ul>Larger resolutions are possible. Note: It is a Diffusion Transformer (DiT) model, it is less restrictive on sizes in compare to a Unet like Illustrious, you can use bigger resolutions. Steps： 30–50 CFG Scale： 3-5 Negative:<pre><code>worst quality, watermark, jpeg artifact, blurry</code></pre>Add "nose", if needed

Hassaku MOSS

The model was trained on a small but dense dataset of 120k images, it still produces realistic-looking results, especially when no characters are involved or the prompt uses natural language. Adding “nose” as a negative tag helps. This is a test to see whether MOSS can be trained at all. The style isn’t stable and I don’t want to push it further, since backgrounds and especially text degrade quickly. MOSS is large and expensive to train, so I’m waiting for the art/anime variant to be released and will focus on that instead. Trained on my own custom trainer with a modified muon optimizer, based on <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2608.20818">Periodic Row-wise Muon</a> but intead of adamW in MuonW, with adafactor. Captions consisting of long NL, short NL, tags with only normal quality tags like "masterpiece". Note: Those 120k images made MOSS mostly more acceptible for my style merge. It is slitly more stable at NSFW in compare to base MOSS, so far i see, but it was only one epoch on a relative low and save LR.

Files with SHA256: 0bd6ad5a7c6ed19e364e2269cc579658a7df4968899d9555b53dbbe9313c7237