<h1 id="astolfokarmix:-astolfomix-xl-+-karmix-xl">AstolfoKarmix: <a target="_blank" rel="ugc" href="https://civarchive.com/models/309514">AstolfoMix-XL</a> + <a target="_blank" rel="ugc" href="https://civarchive.com/models/2194383/karmix-xl-v0">Karmix-XL</a></h1>An experimental merge for mating 2 different base model.Discord: "<a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch00/red_pill.md">Good luck</a>".<h2 id="specification">Specification</h2><ul><li>Merge log / receipe: <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch05/astolfo_mix_e2e.ipynb">e2e script</a>, <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch05/README_AK.md">log</a>. 2 x 3 models + 1 tie breaker.</li></ul><ul><li>Algorithm used: <a target="_blank" rel="ugc" href="https://civarchive.com/articles/3409/astolfomix-experiments-on-merging-algorithms-v141">"Read from bottom"</a>. "Modded DELLA". Parameter in e2e script.</li></ul><ul><li>(Measured in L2 norm) 87.5% NoobAI eps + 12.5% "everything else".</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/articles/15866">Tech report</a>.</li></ul><h2 id="how-to-use">How to use</h2><ul><li>Karmix: "Another NoobAI / IL merge". You know the drill.</li><li>AstolfoMix: Train LoRA / merge on top of it. Tested for realistic model already. Introduced forgotten general contents.</li><li>NoobAI eps: Recovered some artist style and NSFW content. It replaced many "conflicted vectors".</li></ul>

AstolfoKarmix-XL

NIL1.5: <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2203.05482">Average weighting</a> as (<a rel="ugc" href="https://civitai.com/models/1232765/illustrious-xl-10">IL1</a> + <a target="_blank" rel="ugc" href="https://civitai.com/models/1369089/illustrious-xl-20">IL2</a> + <a target="_blank" rel="ugc" href="https://civitai.com/models/833294?modelVersionId=1116447">NoobAI V1.1 Eps</a>) / 3Did not apply Rebasin merge because of <code>inf</code> / <code>nan</code> issue.NIL1.5 alone can output reasonable images already. NIL1.5 is merged via A1111, remaining AK procedure via sd-mecha.

Files with SHA256: 5fc43074a7d41a1e8760fc82dec1fcc90a7ba9af16c80a2bddfbeb0088ab7345