<h3 id="a-completely-new-milestone-in-model-training-my-initial-desire-was-to-make-not-just-a-retro-model-but-a-full-fledged-retro-tool-that-does-not-imitate-old-anime-but-fully-recreates-them-and-with-this-release-i-am-a-few-more-steps-closer-to-this.-(use-booru-tags)-tq8bkuba4">A completely new milestone in model training, my initial desire was to make not just a retro model, but a full-fledged retro tool that does not imitate old anime, but fully recreates them, and with this release I am a few more steps closer to this. (USE BOORU TAGS)</h3><a target="_blank" rel="ugc" href="https://boosty.to/girlsai/posts/64705f5b-b5cb-4e9e-b806-0613976cd823?share=post_link">NEW ERA 5.0 (NAI) / MODEL</a><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/5f0a90a6-29d5-4ce1-8d6e-0a67bf989558/width=525/5f0a90a6-29d5-4ce1-8d6e-0a67bf989558.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c3cfa30e-f472-43ef-9cf4-98f8541d75b6/width=525/c3cfa30e-f472-43ef-9cf4-98f8541d75b6.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f7c7d9b6-6608-4292-941c-fbb276864881/width=525/f7c7d9b6-6608-4292-941c-fbb276864881.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e20bd1d0-ab59-46e3-9fb7-fada5d8fe635/width=525/e20bd1d0-ab59-46e3-9fb7-fada5d8fe635.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f6d8f3cd-07cd-4aa7-83c1-bf7fe1283bed/width=525/f6d8f3cd-07cd-4aa7-83c1-bf7fe1283bed.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e632e794-0608-43a3-aa26-2b34a97b6061/width=525/e632e794-0608-43a3-aa26-2b34a97b6061.jpeg" />The basis of the model will be the addition of artists from the 80s-00s, as well as a huge number of charactersLet's start in order:the release was a little delayed because I tried to improve the finished lora, improve detail, volume, and I realized that all attempts at the moment are not working and I need to study this topic deeper. But what came out cannot be compared with the first version, which is a model, you ask - how can the lora be better than the model? Easy, firstly, I used all the video memory during training to achieve the best results, according to my calculations, the best lore (or model) needs about 60-70 gigabytes of video memory, my 24 is not enough, but enough for the neural network to finally start remembering characters, and most importantly their style. Including unusual faces, eyes, noses, even non-standard physiques, I still believe that this process can be improved, but at the moment this is the maximum that I could achieve.Lora weight: 0.7-1 (at weight 1 there are problems with anatomy, because Lora trained for a very long time)CFG scale - 3-4 (immediate warning! Use cfg scale 3, I still couldn't solve the problem of art fading with a high cfg scale, so just use a lower one. Don't use less, there may be problems with the anatomy)prompts: retro artstyle, 1girl &lt;lora:NEW ERA_v2.1:0.8&gt;<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/01d12644-a169-4b97-a20a-61de63ac67b5/width=525/01d12644-a169-4b97-a20a-61de63ac67b5.jpeg" /><a target="_blank" rel="ugc" href="https://civarchive.com/models/282341?modelVersionId=629577">4th tail (Hentai model)</a> it is very important to use this model, on which I trained my dataset, on other models (except for the original pony) there are problems with recognizing artists, characters and, in principle, their influence on the lora is too highUse prompts: masterpiece, best qualityNegative prompts: worst quality, low quality, bad hands, fewer digits, extra digits, bad anatomy, english text, engrish textanime screencap, anime coloring - two strong tokens, works great, makes the image look like screencaps from anime, can be used both together to enhance it, and separately retro artstyle - main retro token, present in almost all trained images and produces different results from the 80s-90s    1990s \(style\)- a very strong token that significantly changes the style of the model1980s \(style\)-  currently about the same as the previous one (the reason is that the style of the nineties is a lot of the early nineties, and the early nineties are like the eighties)2000s \(style\)- again it doesn’t work perfectly, but it’s slightly closer to 00s display of multiple characters:these are no longer “twins” who are like two peas in a pod, with the same emotions, skin, hairstyles, clothes, etc.prompt:score_9, source_anime, 3girls, 1980s \(style\), cleavage, by kawarajima kou, smile, open mouth, one eye closed, serious, closed mouth, official art &lt;lora:NEW_ERAv2.1:1&gt;<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6baa1c50-9306-4b68-967f-1c721237e3ad/width=525/6baa1c50-9306-4b68-967f-1c721237e3ad.jpeg" />You can use standard tokens from pony diffusion if you think that the quality does not suit you:score_9, source_anime, score_8_up, score_7_up, best qualityI want to say right away that there is now a lot of discussion in the community about the advisability of using artists in neural networks, I’ll say it frankly and without censorship - I don’t give a fuck. I will not hide artists under any three-letter designations so that it is difficult for people to find them in the model (as the creator of pony diffusion does, this is his right, I do not blame him, but I will not do this myself). I will be adding tons of artists, characters, anime titles under their original names.The following artists are currently implemented (always put "by" at the beginning):by urushihara satoshiby danmakumanby kitazume hiroyuki - need add moreby kawarajima kouby kotobuki tsukasaby hirano toshihiro - need add a little more<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a1da566f-1193-42fd-a18f-6881b21e982a/width=525/a1da566f-1193-42fd-a18f-6881b21e982a.jpeg" />there are other artists in the dataset, but their influence on the lora is still too weak.At the moment I have not compiled a list of added characters, there are so many of them that I’m afraid I’m not ready to spend time on this yet, but most likely I will.if you are interested in the details, in my post <a target="_blank" rel="ugc" href="https://boosty.to/girlsai/posts/a3baad1f-1f76-4fe4-805a-20755608a57c?share=post_link">NEW_ERAv1.0</a> everything is described in detail<h1 id="the-following-information-applies-only-to-older-models-based-on-stable-diffusion-1.5:-fu98w98ku">The following information applies only to older models based on Stable Diffusion 1.5:</h1>A small version comparison (SD 1.5):<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/aa366c00-a2eb-4864-9ed5-ba281b2859f2/width=525/aa366c00-a2eb-4864-9ed5-ba281b2859f2.jpeg" />comparison of models (SD 1.5):<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e07bad42-dee4-4c42-9cca-a2c9c1e9efc9/width=525/e07bad42-dee4-4c42-9cca-a2c9c1e9efc9.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/878f6c70-3e50-4aa6-83c5-bcd8dffff455/width=525/878f6c70-3e50-4aa6-83c5-bcd8dffff455.jpeg" />I made a <a target="_blank" rel="ugc" href="https://drive.google.com/file/d/1lR7dKvFoKK9mCwKU4hVcR8jwYnYXoZgE/view?usp=share_link">video</a> on how to achieve the same quality or just repeat my artHands finally reached the style of the 90s. There were many difficulties, dozens of versions, but I found a golden mean, from which I will continue to build on future updates. just use hires fix as much as your video card is capable of, so as not to give an error, with hires fix 2, upscaler Latent (nearest-exact), the picture will be much more detailed. If you want a style close to the original but with more detail, don't use Denoising strength above 0.55, but to avoid artifacts, do not use too low a value, optimally 0.5 For SD 1.5: Use <a target="_blank" rel="ugc" href="https://civarchive.com/models/9942?modelVersionId=11811">Abyssorangemix3aom3_aom3a3.safetensors</a> model with these negative prompts: (worst quality, low quality, extra digits:1.4) Combined with my other <a target="_blank" rel="ugc" href="https://civarchive.com/user/OneRing">LORA</a> How to use LoRA's in auto1111:<ul><li>Update webui (use <code>git pull</code> <a target="_blank" rel="ugc" href="https://www.youtube.com/embed/mn8fMF10XN4?start=31&amp;end=60">like here</a> or redownload it)</li><li>Copy the file to <code>stable-diffusion-webui/models/lora</code></li><li>Select your LoRA like in <a target="_blank" rel="ugc" href="https://www.youtube.com/watch?v=-bMeyXOZwN0">this video</a></li><li>Make sure to change the weight (by default it's <code>:1</code> which is usually too high) *Information taken from <a target="_blank" rel="ugc" href="https://civarchive.com/user/Lykon">Lykon</a> Please post your work with or without comments, it will help me improve. Thanks! If you like my work, click on the heart above, I will be pleased :3</li></ul>

NEW ERA / LORA / PONY DIFFUSION

A completely new milestone in model training, my initial desire was to make not just a retro model, but a full-fledged retro tool that does not imitate old anime, but fully recreates them, and with this release I am a few more steps closer to this.

Files with SHA256: 0cdd1a282c167074ac08e3b241a336e5f5dc93b13430b232e63de345c29b6675