<h1 id="heading-49456">AstolfoMix</h1><p>An exotic merge of 20 models. See <a target="_blank" rel="ugc" href="https://civarchive.com/articles/3409/diagram-of-unet-and-exotic-merging-methods">this article</a> for description. Ignore the content below <s>will update soon</s>.</p><p></p><h2 id="heading-49457">Abstract</h2><p>I present AstolfoMix, a merge model focusing on <a target="_blank" rel="ugc" href="https://www.urbandictionary.com/define.php?term=absurdres">absurdres</a>, and well as drawing most content out of the box, disregarding the art style, <s>as an attempt to the </s><a target="_blank" rel="ugc" href="https://en.wikipedia.org/wiki/Bias%E2%80%93variance_tradeoff"><s>bias-variance tradeoff</s></a>. Currently it is in anime style, and <em>Astolfo is so cute!</em></p><p></p><h2 id="heading-49458">Introduction</h2><p><a target="_blank" rel="ugc" href="https://huggingface.co/6DammK9/AstolfoMix">See the HuggingFace model page</a>. I am too little time to repeat the contents there. Also <em>all the intermediate / variants</em> are available there. <a target="_blank" rel="ugc" href="https://huggingface.co/6DammK9/bpmodel-sd14-merge">Also check out this model page for more models</a>.</p><p></p><h2 id="heading-49459">Related Works</h2><p>I am not sure, seems that "uniform merge" is an original idea. <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/merge.md">However I did some information gathering about merges</a>. Check out <a target="_blank" rel="ugc" href="https://github.com/Xerxemi/auto-MBW-rt">auto-MBW-rt</a> for the future "boosting" approach, instead of my weaken "bagging" approach. Sorry for <a target="_blank" rel="ugc" href="https://github.com/hako-mikan/sd-webui-supermerger/blob/main/elemental_en.md">manual MBW</a> fans, I don't have any perference, or the "absurdres" is never covered in such exploration.</p><p></p><h2 id="heading-49460">Methodology</h2><p><em>Disclaimer: it is trolling, unless you already know what I'm writing. I won't explain anything here. Also it may be duplicated in Experiments, because I need to be tested independently.</em></p><p></p><h3 id="heading-49461">CS-related theory</h3><p><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/tree/main/ch05">Try your best to read my article</a>. Seek for any PHD students / Profs if that is BS to them (Maybe I will make a paper formally). Also I still need time to <s>conduct test</s> write Python codes which is almost impossible in a <a target="_blank" rel="ugc" href="https://en.wikipedia.org/wiki/996_working_hour_system">996 life</a>.</p><p></p><h3 id="heading-49462">Receipe</h3><p><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch05/receipe-20a.json">Full receipe here</a>. Remember, use <a target="_blank" rel="ugc" href="https://github.com/arenasys/stable-diffusion-webui-model-toolkit">this toolkit</a> to<a target="_blank" rel="ugc" href="https://huggingface.co/CompVis/stable-diffusion-v-1-4-original"> replace with original SD's CLIP</a>. <em>Whole procedure should be carried in WebUI only, without any lines of coding</em>.</p><h3 id="heading-49463"></h3><h2 id="heading-49466">Experiements</h2><p></p><h3 id="heading-6011">Artworks</h3><p>Included in <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/tree/main/ch05">my article</a>, also <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/tags/AstolfoMix/artworks">my artworks</a>, and my usual <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/my_procedure.md">routine (no manual image editing)</a>. For the last part of "correlation", I'll complete it "soon". No score metric is involved, because <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2309.14859">it has been discussed</a>, and I prefer basic metrics on differentiation <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/aesthetic.md">before attempting to correlate aesthetic with logical justification</a>. For my personal feeling, it <em>"draw" anime stuffs with realistic scale, which is unique across all the models available in internet</em>.</p><p></p><h3 id="heading-49463">Prompts</h3><p>This is a base model. <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch02/animevae_pt.md">Use what you are familiar with, without thought of "trigger words"</a>. <strong>PNG info in CivitAI is incomplete</strong>. <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/tags/AstolfoMix/artworks">See my artworks instead</a>. <em>Ignore prompt weight</em> if you're using ComfyUI, <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/prompt.md">or you don't know how it works</a>. "Who What Where" is enough.</p><p></p><h3 id="heading-49464">Samplers</h3><p><strong>Use what you like</strong>. I use default sampler because I am simple, <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch98/k_euler.md">Also it is not bad at all</a>.</p><p></p><h3 id="heading-49465">CFG / STEPS / ADDONS</h3><p><a target="_blank" rel="ugc" href="https://civarchive.com/images/2899281?period=AllTime&amp;sort=Newest&amp;view=categories&amp;modelVersionId=174551&amp;modelId=155604&amp;postId=677828">Thanks for the very first review, any "default range" (CFG 7.5) works (obviously).</a></p><p>From experience, <strong>CFG <s>4.5</s> 4.0</strong> will be the minimum optimal value. <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/cfg_step.md">Plot if you doubt</a>. Feel free to use <a target="_blank" rel="ugc" href="https://github.com/Bing-su/adetailer">adetailer for local latent upsampling</a>, or <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/hires_fix.md">built-in HiRes fix for global</a> (2.0x for 1024x1024, or 2.5x for 768x768, lower if image breaks). <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/dynamic_cfg.md">Default Dynamic CFG (mimic 1! Stick with the paper!)</a>, with <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/freeu.md">default FreeU (</a><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch01/dynamic_cfg.md">Stick with the paper!)</a>. <em>It is just a base model</em>.</p><p></p><h3 id="heading-6012">Embeddings / LoRAs</h3><p><a target="_blank" rel="ugc" href="https://civarchive.com/images/2899281?period=AllTime&amp;sort=Newest&amp;view=categories&amp;modelVersionId=174551&amp;modelId=155604&amp;postId=677828">Thanks for the very first review, any generatl embedding (trainned from NAI / SD or any models "close" to them) will be fine.</a></p><p>Check out <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/artworks/112661463">this post</a> and <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/artworks/112744854">this post</a> for reference. (They may be appeared as cross posts). In the worst case it loss the ability to draw with Hi-Res upsampling. Moreover (it may be triggering), <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch02/f59359c175.md">it supports VBP's original 273 style embeddings</a>. Mirror will be activated by request.</p><p></p><h3 id="heading-6013">Diversity of contents / Art styles</h3><p>"Yes". But keep in mind, it is just a "good" base model. "Base as most contents". It cannot replace embeddings / LoRAs, or any other tools. You will need prompting skills to do minor stuffs (e.g. non human, non binary stuffs etc.)</p><p></p><h2 id="heading-49467">Discussion</h2><p><em>I think this is not my discussion. </em><strong>Just try to use my model like a base model</strong>: mix it further, LoRA / Finetune, everything is welcome. I may "beat" most models in such hires, and reaching SDXL's level, but I'm sure it comes with price (e.g. image layers as "v-pred", and the "bias" which must be larger then SDXL with 1024 ARB training with LAION dataset out of the box), and that is way out of my capability.</p><p></p><h2 id="heading-49468">Conclusion</h2><p>Just try my model! AstolfoMix only represent my personal "model selection", everybody can use the "uniform merge" to make you own base model, before great SDXL model comes. Also <em>do not waste your time on the risky MBW / finetuning if you don't have abundant resources.</em></p><p></p><h2 id="heading-6014">Appendix</h2><p></p><h3 id="heading-6015">Model used</h3><p><em>It should be same as the </em><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch05/receipe-20a.json"><em>full receipe.</em></a></p><ol><li><p><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch02/f59359c175.md">VBP</a></p></li><li><p><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch02/ae2b38ac14.md">CBP</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/ashen-sensored/mzpikas_tmnd_enhanced">MzPikas TMND Enhanced</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/4384/dreamshaper">DreamShaperV8</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/StereoBartender/CoffeeWithLiquor">CoffeeWithLiquor</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/117192/breakdomainowners-model-i-have-no-rights-if-he-want-it-will-be-taken-down">BreakDomain</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/74165?modelVersionId=78888">AIWMix</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/17427/ether-blu-mix">Ether Blu Mix</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/43331?modelVersionId=94640">majicMIX realistic</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Xynon/SD-Silicon">Silicon29</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Crosstyan/BPModel">BP</a></p></li><li><p><a target="_blank" rel="ugc" href="https://t.me/StableDiffusion_CN/1170018">CGA9</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/153081?modelVersionId=23379">LimeREmix</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/71185/cyberrealistic-classic">CyberRealistic Classic</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/reroti/ORCHIDHEART">ORCHIDHEART</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/17649/bb95-furry-mix">BB95 Furry Mix</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/34469?modelVersionId=167882">Indigo Furry mix</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/15509/aoaoko-pvc-style-model">AOAOKO [PVC Style Model]</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/10415/3-guofeng3">GuoFeng3</a></p></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/3671?modelVersionId=189192">YiffyMix</a></p></li></ol>

AstolfoMix

<p>RL (autombw, BO + ImageReward) of 20 models.</p>

Files with SHA256: aab8357cdcee982754dec3815bc4e271c0908e18d4640313761424e5dac1f892