krea2真实微调版本使用10000张图像对于krea2 raw模型进行训练后再从turbo模型恢复出turbo能力得出来的turbo大模型训练后的raw模型之后等有空了放出。目前看来lokr任然是效果与开销最低的玩法。有兴趣的可以看一下我的付费Lokr模型以下是链接：<a target="_blank" rel="ugc" href="https://krea2.whitebay.click/">https://krea2.whitebay.click/</a><a target="_blank" rel="ugc" href="https://t.me/+EwMu9USA3hg5Y2E9">可以进入交流群获取偶尔掉落的免费模型：<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">https://t.me/+EwMu9USA3hg5Y2E9</a><span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">针对网上的一些人的说法，给你们一个答复。<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">针对早期的免费模型，是出于学习目的。并且在成本可控的情况下进行免费分享。<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">在sdxl之后的模型，微调的成本已经爆炸，个人已经无力负担模型训练的成本。所以转向小规模付费群体来平衡模型训练支出。<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">这次这个模型是出于一个实验性质的微调krea2，我也写了只有10000张图的微调。对于krea2这种模型，属于微乎其微了。更多是出于科研性质，把turbo和raw模型一起免费放出，为的是起一个抛砖引玉的作用。<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">你们所说的其他作者效果好免费分享，不过只是训练了一个lora之后融合进入大模型中，成本几乎为0.<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">这一点我不做自证也不做解释。10000张图的微调对我来说就是100w步的训练过程，哪怕我用h200*8，也需要三天时间的训练。然而我任何一个lokr模型的训练量都远超这个，我也不愿意做融合模型来欺骗自己欺骗大众。然而真实的费用是3天的h200的算力租金是几乎2000美金。<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">我付出了时间付出了金钱，然而你只出了一张嘴。<span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">微调krea2模型必然存在raw模型和turbo模型（能懂的自然懂，不懂的也没必要解释）After training the Krea2 RAW model on 10,000 images, then restoring the Turbo capability from the Turbo model to derive the Turbo model.The trained RAW model will be released later when I have time.For now, LoKR still appears to be the approach with the best results at the lowest cost. Those interested can check out my paid lokr model.<a target="_blank" rel="ugc" href="https://eec5643c88844ec8a27f98893e73b0ac.region1.waas.aigate.cc/launch.html">https://eec5643c88844ec8a27f98893e73b0ac.region1.waas.aigate.cc/launch.html</a><span style="background-color:rgb(34, 34, 34);color:rgb(221, 221, 221);font-family:&quot;Helvetica Neue&quot;, Helvetica, Arial, &quot;Hiragino Sans GB&quot;, &quot;Hiragino Sans GB W3&quot;, &quot;Microsoft YaHei UI&quot;, &quot;Microsoft YaHei&quot;, sans-serif;font-size:13px">You can join the chat group to get the occasional free models:<a target="_blank" rel="ugc" href="https://t.me/+EwMu9USA3hg5Y2E9"><span style="background-color:rgb(38, 38, 38);color:rgb(45, 119, 229);font-family:&quot;Color Emoji&quot;, system-ui, &quot;PingFang SC&quot;, PingFangSC-Regular, &quot;Microsoft YaHei&quot;, &quot;Hiragino Sans GB&quot;, &quot;Heiti SC&quot;, &quot;WenQuanYi Micro Hei&quot;, Arial, Helvetica, sans-serif, &quot;Apple Braille&quot;, &quot;Color Emoji Fix&quot;;font-size:14px">https://t.me/+EwMu9USA3hg5Y2E9</a><hr />A reply to some of the things people have been saying online.As for my early free models — those were released for learning purposes, and I shared them for free while the cost was still manageable.For models after SDXL, the cost of fine-tuning has exploded. As an individual, I can no longer afford the cost of model training. So I moved toward a small paying group to offset the training expenses.This particular model was an experimental fine-tune of Krea2. I stated clearly that it's a fine-tune on only 10,000 images — which, for a model like Krea2, is negligible. It was more of a research exercise. Putting the turbo and raw models out for free together was meant to throw out a brick and see what jade comes back.As for the "other authors get great results and share for free" argument — that's just training a single LoRA and merging it into the base model. It costs essentially nothing.I won't defend myself on this, and I won't explain it further. For me, a 10,000-image fine-tune is a 1,000,000-step training run. Even with 8×H200, that's three days of training. And every one of my LoKr models is trained far beyond that. Nor am I willing to produce merged models to fool myself and fool everyone else. The real number: three days of H200 compute rental runs close to $2,000.I put in the time and I put in the money. All you put in was your mouth.Fine-tuning Krea2 necessarily involves both a raw model and a turbo model. (Those who get it, get it — no need to explain to those who don't.) <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3cbeaf4e-841f-489d-973b-9bf3d3194b09/original=true/3cbeaf4e-841f-489d-973b-9bf3d3194b09.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c1a5c1d2-434f-474f-b999-5c289bd7ac76/original=true/c1a5c1d2-434f-474f-b999-5c289bd7ac76.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/70722c2a-3792-4586-9a1b-571d129f4391/original=true/70722c2a-3792-4586-9a1b-571d129f4391.jpeg" />

Files with SHA256: ade3d4bf21b6cd16cdff3a1a067b5b62e5cb96f0238da7ccac068d815a5391d4