*Important, please read this before using the model because this is very experimental, as I am still trying to fine the optimal settingsYou can download the clip text encoder here: <a target="_blank" rel="ugc" href="https://huggingface.co/suzushi/miso-diffusion-m-beta">https://huggingface.co/suzushi/miso-diffusion-m-beta</a>Miso Diffusion M (Beta) is an attempt to fine tune stable diffusion 3.5 medium on anime dataset. In comfy ui it uses as little as 2.4 gb vram without the t5 text encoder. This version is a step up from previous version (alpha) , trained on 160k image for 3 epoch to see how it adapts to anime.Recommanded setting, euler, cfg:5 , 28-40 steps, though dpm ++ 2m also works but haven't done much testing, prompt: danbooru style tagging. I recommand simply generating with a batch size of 4 to 8 and pick the best one.Quality tag<pre><code>Masterpiece, Perfect Quality, High quality, Normal Quality, Low quality</code></pre>Aesthetic Tag<pre><code>Very Aesthetic, aesthetic</code></pre>Pleasent<pre><code>Very pleasent, pleasent, unpleasent</code></pre>Additional tag: high resolution, elegantTraining is done on gh200. Switched lr scheduler to cosine this timeTraining setting: Adafactor with a batchsize of 40, lr_scheduler: cosineSD3.5 Specific setting:enable_scaled_pos_embed = truepos_emb_random_crop_rate = 0.2weighting_scheme = "flow"Train Clip: true, Train t5xxl: false

Files with SHA256: d7d7f41a1435441472bd830ab221541e783ce29b4745707ab38fc39822fb57f9