<h1 id="astolforf-(3ep)-astolfovl-(2.5ep)-astolfoxl-(2ep)">AstolfoRF (3EP) / AstolfoVL (2.5EP) / AstolfoXL (2EP)</h1>Probably the first (and the only) individual <a target="_blank" rel="ugc" href="https://www.ibm.com/think/topics/parameter-efficient-fine-tuning">Full Finetuning</a> with <a target="_blank" rel="ugc" href="https://huggingface.co/blog/accelerate-deepspeed">multi-GPU </a>and... as open-source as it can (may not be copyleft).<a target="_blank" rel="ugc" href="https://civarchive.com/models/690694?modelVersionId=773022">LoKR</a> <a target="_blank" rel="ugc" href="https://civarchive.com/models/690694?modelVersionId=773022">works</a>, but no thanks. 我不做人了，<a target="_blank" rel="ugc" href="https://civarchive.com/models/647664?modelVersionId=859577">早苗</a>！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>Base model (3EP): <a target="_blank" rel="ugc" href="https://civarchive.com/models/1489964">AstolfoVL</a>, version <a target="_blank" rel="ugc" href="https://civarchive.com/models/1489964?modelVersionId=2712014">2.5EP (VPRED)</a> RF</li><li>Base model (2.5EP): <a target="_blank" rel="ugc" href="https://civarchive.com/models/1671685">AstolfoKarMix-XL</a>, version <a target="_blank" rel="ugc" href="https://civarchive.com/models/1671685?modelVersionId=2167201">Evo-2EP </a>VPRED</li><li>Base model (2EP): <a target="_blank" rel="ugc" href="https://civarchive.com/models/1671685">AstolfoKarMix-XL</a>, version <a target="_blank" rel="ugc" href="https://civarchive.com/models/1671685?modelVersionId=2167201">NIL1.5 v1.2</a></li><li>Base model (1EP): <a target="_blank" rel="ugc" href="https://civarchive.com/models/309514">AstolfoMix-XL</a>, version <a target="_blank" rel="ugc" href="https://civarchive.com/models/309514?modelVersionId=1612266">255c</a></li><li>Tech report: <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/tree/main/ch06">ch06</a></li><li>Training metrics (tensorboard): <a target="_blank" rel="ugc" href="https://huggingface.co/6DammK9/AstolfoXL/tensorboard">HF</a></li><li>Dataset (images &gt; latents): <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/6DammK9/danbooru2024-latents-sdxl-1ktar">danbooru2024</a>, <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/6DammK9/e621_2024-latents-sdxl-1ktar">e621_2024</a></li><li>Dataset (tags + captions): <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/6DammK9/danbooru2024-latents-sdxl-1ktar/blob/main/meta_lat_v3.tar.gz">meta_lat.json</a></li><li>1 step = 16 images, <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch04/4x3090_v2.md">4x RTX 3090 24G</a>.</li><li>778k steps for 1EP, 8.0 + 4.6 = 12.6M images</li><li>Tag + NLP caption with <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch06/sd-scripts-runtime/kohyas.md">A1111 token trick</a></li><li>Trainer codes: <a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-scripts/pull/1950">The PR won't be merged</a></li><li><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch06/sd-scripts-runtime/readme.md#exploring-hardware-platforms-cpu-side">Train parameter (1EP, 2EP)</a>: adamW8bit, UNET 1.5e-6, TE 1.2e-5, BS4 (4 GPU) grad accu 4, <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch06/sd-scripts-runtime/sdxl_original_unet_71.py">71% UNET </a>(Speed + must <a target="_blank" rel="ugc" href="https://www.ibm.com/think/topics/underfitting">underfit</a>)</li><li><a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch06/sd-scripts-runtime/readme.md#exploring-hardware-platforms-cpu-side">Train parameter (2.5EP, 3EP)</a>: adamW8bit, UNET 1.5e-5, TE OFF, BS4 (4 GPU) grad accu 4, <a target="_blank" rel="ugc" href="https://github.com/6DammK9/nai-anime-pure-negative-prompt/blob/main/ch06/sd-scripts-runtime/sdxl_original_unet_71.py">100% UNET </a>(finetune to different <a target="_blank" rel="ugc" href="https://valentinpratz.de/posts/2024-11-14-trigflow/">objective parameters</a>)</li><li>75-100+ days for 1EP. Train 1 EP only. Save per 10k steps.</li><li>Train result and loss curve: <a target="_blank" rel="ugc" href="https://huggingface.co/tensorboard/6DammK9/AstolfoXL/?smoothing=0.999#timeseries&amp;_smoothingWeight=0.999&amp;runSelectionState=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">Tensorboard in HF</a></li><li>Core concept: <a target="_blank" rel="ugc" href="https://en.wikipedia.org/wiki/Unsupervised_learning">Unsupervised learning</a></li><li>Expectation: MID (100% no filter no quality tag) <s>or "reality" "</s><a target="_blank" rel="ugc" href="https://en.wikipedia.org/wiki/Golden_mean_(philosophy)"><s>golden mean</s></a><s>",</s> proven in <a target="_blank" rel="ugc" href="https://huggingface.co/spaces/yoinked/weeb-arena-results">Public Arena (ELO 1500)</a>. Currently in <a target="_blank" rel="ugc" href="https://arena.sharkgone.com/">s3.</a></li><li>Actual: <s>Need TIPO</s> <a target="_blank" rel="ugc" href="https://civarchive.com/articles/23650">non empty negative prompt</a></li></ul><h2 id="how-to-use">How to use</h2><ul><li>(Reference only, unchanged since 2025) CFG4, Euler, Shift 3.0. </li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/618179?modelVersionId=1671158">Train LoRA</a> / merge on top of this model. Compatability should still close to <s>215c</s> base model. <a target="_blank" rel="ugc" href="https://civarchive.com/models/176449?modelVersionId=1649411">Realistic human content</a> is still supported. "Trust me bro".</li><li>Artist tags may not work, but I did trained. Just dump your "NAI" prompts here.</li><li>Use <a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf/z-tipo-extension">TIPO</a> to expand tag based prompts with NLP.</li><li>Short tags will suffer from background latent noise. Tags can be observed from E621 or danbooru.</li><li>All images are just seen once. There is no task or KPI to chase, or that omniscient state has been archived. The loss curve is as flat as it neither converge or diverge.</li></ul>Full docuementation will be published, which is as long as the AstolfoMix series.

AstolfoRF - SDXL Rectified Flow

Switched trining objective, base model is <a target="_blank" rel="ugc" href="https://civitai.com/models/1671685?modelVersionId=2586876">Evo 2EP</a>.Make sure negative prompt in A1111 / ReForge must have some arbitary value (recommended a single ",") to avoid the undesired "zero uncond" issue. ComfyUI has no such issue. It is the inversion of <a target="_blank" rel="ugc" href="https://civitai.com/models/1898715?modelVersionId=2705799">Evo 2.5EP</a>.Since the furry femboy is here, I guess I no longer need merging afterward.

astolforfSDXL_25EPVpred.yaml

astolforfSDXL_25EPVpred.safetensors

2.5EP (vpred)

3EP (RF)

395k (1.5EP)

400k

<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

<a target="_blank" rel="ugc" href="https://civitai.com/models/1671685?modelVersionId=2167201">AK-NIL1.5</a> merged with a special variant of <a target="_blank" rel="ugc" href="https://civitai.com/models/536954?modelVersionId=2173556">Hina's LyCORIS finetune</a>.Not useable with full weight (1.0), but it works at ratio 0.55 (hyperparameter).Technically it is no longer a NoobAI Vpred model. It is a more like a IL Vpred model.<s>Karmix... Carmix... Whatever.</s>

astolfocarmixVpredxl_akNIL15BasedV01.yaml

astolfocarmixVpredxl_akNIL15BasedV01.safetensors

AK-NIL1.5 Based (v0.1)

AC-Evo 2.5EP

NoobAI based

Switched trining objective again (RF), base model is <a target="_blank" rel="ugc" href="https://civitai.com/models/1489964?modelVersionId=2712014">2.5EP (Vpred)</a>.Make sure negative prompt in A1111 / ReForge must have some arbitary value (recommended a single ",") to avoid the undesired "zero uncond" issue. ComfyUI has no such issue.Since the furry femboy is here, I guess I no longer need merging afterward.For A1111, since I don't have idea to make it support RF yet, there is a "<a target="_blank" rel="ugc" href="https://valentinpratz.de/posts/2024-11-14-trigflow/">partial trigflow</a>" config file... yep, it is the vpred yaml file once again. It looks a bit weird but at least you can see the contents.

astolforfSDXL_3epRF.safetensors

astolforfSDXL_3epRF.yaml

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.

astolfocarmixVpredxl_acEvo25EP.safetensors

astolfocarmixVpredxl_acEvo25EP.yaml

Sanity check: Merging 11 NoobAI VPred models should yield another NoobAI VPred model.

Files with SHA256: e8874310904a33f12d1c45f0ee61165067efc1d67293ab6347f878105eb8c68a