<h3 id="heading-22058">TL;DR</h3>An ahegao LoRA model intended to generate those "ahegao profile pictures". Trained with 100% monochromatic images of ahegao anime girls, zoomed in on their faces.<h1 id="heading-22059"><code>1girl, monochrome, gao, open mouth, tongue out, close up</code></h1><h1 id="heading-22059"> 🌸 aheGAO - a LoRA 🌸</h1><h2 id="heading-19834">Description</h2>Gao is a LoRA concept model trained using cropped pictures from hentai mangas, focusing solely on the female's face.The creation of this concept was intended for generating "ahegao profile pictures" out of curiosity, not specifically for use as a profile picture.However, of course, nobody is stopping you from using this for your profile picture... 🤔The training images were 100% monochromatic and were zoomed-in ahegao faces intended to be profile pictures, just for clarification.The LoRA was trained with the <a target="_blank" rel="ugc" href="https://civarchive.com/models/9409?modelVersionId=90854">Anything</a> checkpoint.You don't have to use that checkpoint, just for your information.<h2 id="heading-22060">Usage</h2>This LoRA is NOT SUITED for resolutions that aren't 1:1You CAN still generate images with a size like 512 x 768, but around ⅓ of the time, it's going to look horrendous.The following includes trigger words added to the training captions. Use them in your positive prompt to generate more specific details, or exclude them in your negative prompt if you don't want such features.<code>gao</code> = the main trigger word<code>ahegao</code> = an optional choice; could further enhance the 'ahegao' face<code>monochrome</code> = monochromatic coloring - even without specifying it, the results are occasionally monochrome, so include it in the negative prompt to decrease the probability.<code>open mouth</code> = give the subject an open mouth<code>tongue out</code> = give the subject a tongue that's sticking out<code>ecstasy</code> = add a stronger euphoric moaning effect or a feeling of ecstasy to the subject, along with a slight grin or an open-mouthed smile - essentially giving the subject a "happy" expression rather than a "neutral" or "frowning" one with their mouth being open.<code>crying</code> = give the subject watery eyes, or tears - can look somewhat odd; this aspect will be improved upon in future versions.<code>rolling eyes</code> = make the subject's eyes appear to be rolling upwards, conveying an ecstasy-like expression<code>heart-shaped pupils</code> = turn the subject's pupils into a heart shape<code>blush</code> = add a blush to the subject<code>close up</code> = gives the image a stronger profile picture effect, zooms in closer to the subject's face - also useful for when the image isn't even zoomed in at all<h3 id="heading-19">LoRA's MODEL STRENGTH</h3><code>0.7 - 0.9</code> works decently for me.And, don't forget to add this into your negative prompt:<code>(worst quality, low quality:1.4)</code>For that FULL effect 😉You can go mess around with the prompt, CFG, whatever sampler you prefer, etc.Go ahead and find what works good enough for you!<h2 id="heading-8644">Cons</h2>These issues are not guaranteed to be evident 100% of the time; these cons only occur occasionally.<ul><li>Unwanted speech bubbles / captions - the concept was trained on manga, so of course, there will be captions everywhere...</li><li>Uncanny empty white eyes - likely due to the extreme rolling back of the eyes in the training images... interesting...</li><li>Strange objects - those text captions ( from the manga / training images ) cause Stable Diffusion to interpret them as random floating objects, especially with those speech bubbles, leading to even stranger results...</li></ul>Other cons include malformed head shape, strangely-shaped tongue, etc. Adding such cons to the negative prompt can fix the problem, other times, not the case.

ahe-GAO [ アヘ顔 ]

completely reworked the LoRA's training<ul><li>Used ~30% less images for trainingv1.0 has 110+ images for trainingv1.5 has 70+ images for training ( NEXT VERSION WILL HAVE MORE, IF I EVEN RELEASE A NEW VERSION )<ul><li>Removed images with WAY too many speech bubbles / text on it - this significantly reduced the amount strange objects appearing in generation.</li><li>Removed images with weird facial features, weird expressions - reduces the chance of uncanny faces, morphed faces, morbid faces, you get the idea.</li></ul></li><li>v1.0 -&gt; single text file as a caption for ALL imagesv1.5 ( current ) -&gt; one caption file in each of the multiple groups of images - meaning that the training images were divided into multiple groups based on similar features, with each group containing images that share the same caption file; essentially, one caption file for one group ( of images that are alike ).This allowed for more control with the type of image you want from this LoRA.</li><li>LoRA was 'reworked' literally. v1.5 is a fresh start without ANY fine-tuning from v1.0.</li><li>Slightly improved compatibility for resolutions above 512, like 512x768 - v1.0 had occasional issues with faces being stretched or completely morphed. v1.5 has a somewhat lower occurrence of this happening.</li><li>Increased the Epoch amount in training from 150 to 200!v1.0 had ~31 steps epoch, and 150 epochs. ( 4650 total steps )v1.5 had ~14 steps each epoch, and 200 epochs. ( 2800 total steps )Once again, nearly 30% of the images used to train the model were removed because of how much they could cause Stable Diffusion to generate weird things...</li></ul>

gao-v1.5.safetensors

v1.5

v1.75

v1.0

Mod from <a target="_blank" rel="ugc noopener noreferrer nofollow" href="https://civitai.com/models/381618?modelVersionId=426038">https://civitai.com/models/381618?modelVersionId=426038</a>Trigger WordsGAOAHEGAOMONOCHROME1GIRLSOLOOPEN MOUTHTONGUE OUTTL;DRAn ahegao LoRA model intended to generate those "ahegao profile pictures". Trained with 100% monochromatic images of ahegao anime girls, zoomed in on their faces.1girl, monochrome, gao, open mouth, tongue out, close up

Files with SHA256: 6392ec60e3e2fea65765fe1530d4d8cf265afaa940b75c9c8947ea629a4cc472