<h1 id="astolfocarmix-vpredxl-:-astolfokarmix-karmix-xl-+-vpred-models">AstolfoCarmix-VPredXL : <a target="_blank" rel="ugc" href="https://civarchive.com/models/1671685/astolfokarmix-xl">AstolfoKarmix</a> - <a rel="ugc" href="https://rentry.co/-introducing-karcher-mean-experimental-model-v0">Karmix-XL</a> +<a target="_blank" rel="ugc" href="https://rentry.co/-introducing-karcher-mean-experimental-model-v0"> </a>VPred Models</h1>An experimental merge for mating <s>2 different base model</s> VPred Models.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>. 11 models + 1 selecetd for VAE.</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) <s>87.5%</s> 100% NoobAI vpred and finetunes / merges..</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>NoobAI vpred: You know the drill.</li><li>SDXL Vpred in A1111: Download <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch06/sd-scripts-runtime/sd_xl_base_vpred.yaml">this yaml</a> and rename as <code>*.yaml</code>.</li></ul>

AstolfoCarmix-VPredXL

Based from <a target="_blank" rel="ugc" href="https://civitai.com/models/1671685?modelVersionId=2586876">AK-Evo 2EP</a>, first 57EP of 6k Astolfo Dataset, then 0.5EP of full 12.4M dataset, finally merge base model with ratio 0.82.The idea is similar to the <a target="_blank" rel="ugc" href="https://civitai.com/models/1898715?modelVersionId=2389469">previous version</a> (0.55), but the <a target="_blank" rel="ugc" href="https://civitai.com/articles/23650/">dataset alignment</a> is confirmed.Planned for 0.2EP only, but eventually it hits another 0.5EP. Next will be RF.For ComfyUI users, use <a target="_blank" rel="ugc" href="https://civitai.com/models/2412096?modelVersionId=2711984">this workflow</a> for workaround. Otherwise just head to my non Evo 2.5EP (ratio 1.00) with full contents.

Files with SHA256: ef263d8e8eb272f759371c064d4e8426c0ee8b3cafd813b8f0911c330d37c292