MOSS-heal v1.0无触发词（no tigger word）本lora训练于：MOSS-heal <a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/MOSS-Image-heal">https://huggingface.co/KBlueLeaf/MOSS-Image-heal</a>This LoRA training and MOSS-heal: <a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf/MOSS-Image-heal">https://huggingface.co/KBlueLeaf/MOSS-Image-heal</a>关于MOSS-heal:是KohakuBlueleaf 使用2M danbooru + 1M laion coco text heavy進行修復訓練並蒸餾原版模型輸出避免大規模偏移。訓練使用KohakuMossTrainer完成，本次修復訓練使用約200gpu hour的B300完成。修復訓練後可以看到residual stream以及model weight中的極端值跟整體數值大小顯著下降，可以顯著提升fp16以及量化模型的穩定性以及訓練的穩定性。也使得LoRA微調等訓練更加容易。更好训练lora，更适合生成动漫画风，但是文字能力有所退化。感谢KohakuBlueleafAbout MOSS-heal:KohakuBlueleaf used 2M Danbooru + 1M LAION COCO text-heavy data to perform repair training and distilled the original model’s outputs to avoid large-scale drift. The training was done using KohakuMossTrainer. This repair training took about 200 GPU hours on B300. After repair training, the extreme values and overall magnitudes in the residual stream and model weights are significantly reduced, which can significantly improve the stability of fp16 and quantized models as well as training stability. It also makes LoRA fine-tuning and other training easier.better for training LoRAs, more suitable for generating anime-style images, but text capabilities have degraded somewhat.Thanks to KohakuBlueleaf.<hr />krea2 v1.0无触发词（no tigger word）经过多次尝试，还是把krea2版本的lora训练出来了。我对此lora依旧不满意，但是也不得不承认这是在我当前所能达到的极限了。krea2模型训练门槛太高，我的试错机会也不多。等未来有比较成熟的训练方案后再继续尝试吧。After numerous attempts, I finally managed to train the Krea2 version of the LoRA. I'm still not satisfied with it, but I have to admit that this is the best I can achieve with my current resources. The Krea2 model has a very high barrier to training, and I don't have many chances to experiment. I'll wait until a more mature training approach becomes available and give it another try then.<hr /><h2 id="anima-base-v2.0tigger-word@meionanimastyle">anima base v2.0触发词tigger word：@meionANIMAstyle</h2>更新详解：更新了数据集。这次采用了比较激进的训练方法，采用完全nl标的格式，且训练的是lycoris的lokr，在画风上表现会更强但是会损失一定的兼容性，当与其他 LoRa 一起使用时，建议降低权重。Update Details: The dataset has been updated. A more aggressive training approach was adopted this time, using a fully NL-targeted format. The model being trained is the LyCORIS LoKr, which results in stronger stylistic expression but at the cost of some compatibility,so when using it with other LoRa, it is recommended to reduce the weights.以下是anima base v2.0和anima base v1.0的对比图：Here is the comparison chart between Anima Base v2.0 and Anima Base v1.0.：<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/4ab2eaa9-61ab-495c-9806-afd5e04d212c/original=true/4ab2eaa9-61ab-495c-9806-afd5e04d212c.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2b33b5d8-3aab-4bed-9faa-b873145f8e06/original=true/2b33b5d8-3aab-4bed-9faa-b873145f8e06.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6fdf74f4-a3c2-46c0-97b0-759f1b59f02f/original=true/6fdf74f4-a3c2-46c0-97b0-759f1b59f02f.jpeg" /><hr /><h2 id="anima-base-v1.0tigger-word@meionre5style">anima base v1.0触发词tigger word：@meionRE5style</h2>没想到anima base这么快就来了。这次采用的是一半nl格式打标和一半tag格式打标，@meion依旧也可以加，anima base训练了gel，对单个画师的画风还是很还原的。PS：最近发现的问题，在出现green eyes和underwear的条件下，会出现角色被替换成原神中的草神的情况，我排查了一下，是训练集那张画作打标居然只有触发词，其他为空。不介意的话可以继续用，介意的话等我过段时间再训练一版。<hr /><h2 id="anima-pre3-v1.0-tigger-word@meion">anima pre3 v1.0 触发词tigger word：@meion</h2>经过一个星期的努力，总算是摸索出anima的lora要怎么训练了，本lora训练于anima p3。后续anima出正式版会重新训练。特别感谢LuIynx的炉子，让我这个5060笔记本电脑也能较为快速的炼anima的lora。项目地址：<a target="_blank" rel="ugc" href="https://github.com/WhitecrowAurora/lora-rescripts/releases/tag/v1.5.6">https://github.com/WhitecrowAurora/lora-rescripts</a><hr /><h2 id="il-v2.0-tigger-wordmeion_style">il V2.0 触发词tigger word：Meion_style</h2>推荐权重lora weight:0.3-1.0<hr /><h2 id="il-v1.0tigger-wordmeion">il V1.0触发词tigger word：meion</h2>推荐权重lora weight:0.3-0.8

铭音 meion style 【krea2&MOSS&anima&il】

meion_moss_style_heal_lora_rd64_15MP-step_00006000.safetensors

MOSS-heal v1.0

moss base v1.0

krea2 v1.0

anima base v2.0

anima base v1.0 v1.0

anima pre3  v1.0

il v2.0

il v1.0

Files with SHA256: 058296921a6ae3a6a8127e3bc96813a65703a43eab1d1ed0abb2aaa6bca3af31