<h1 id="v1-attention-nsfw-supported-2m44rgx1u">V1 - Attention NSFW SUPPORTED</h1>Hi there, this is my Marin Kitagawa model. I've spent about 3 days testing things here and there to have a decent output, and i think i finally achieved my goal.<h2 id="description-bhsvfu7re">Description</h2>The model was trained on 200 images of marin with a really low learning rate, for 13 epochs.The model was trained on Blue-Tear v2, and it is intended to work on models based on it.The activation keyword is:<pre><code>M4R1N</code></pre>It is suggested the use of additional tags as:<pre><code>choker, piercing</code></pre>The model cannot reproduce the earclip of marin, it is intended to reproce just the earrings.There's no need to use the "1girl" tag, but there are no big complication in using it (compared on the nagatoro lora this is a huge improvement).The hair, the color of the hair, and the eyes are already included in the "m4r1n" tag so there's no reason to add additional tags.ps. the eyes are the feature that was understood less by the models, so they can be affected easily by other tags (such as empty eyes), and lose their standard color. To fix this you can tweak this tag strenght, use adetailer, or trying to use a "color + eyes" tag depending on the desired result.The suggested weight is 1, even toh the model function at lower values.<h3 id="tests-and-info-2dsa9e42g">Tests and Info</h3>The model was trained on different image sizes, so to have a better understanding of which dimension worked best with it, i've used a generic prompt and swapping only the tags in the y axis<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/cdf53df8-7d0a-440f-b9bd-99c24c717be3/width=525/cdf53df8-7d0a-440f-b9bd-99c24c717be3.jpeg" />As you can see different sizes can produce pretty different results.For example using the 896x1152, you will achieve a better "close-up" than using 832x1216.<hr />I then wanted to test the expressions and the flexibility of the lora, and i have to say that the results are kinda good.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d01cc609-9146-4b00-8f1b-7aa6a1be4d88/width=525/d01cc609-9146-4b00-8f1b-7aa6a1be4d88.jpeg" /><hr />Another test i wanted to do was checking the effect of changing the weight:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/33f1fa5f-89a7-47a5-9b15-f4583d6bfd13/width=525/33f1fa5f-89a7-47a5-9b15-f4583d6bfd13.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/fc2a491c-fb16-4ca8-9fee-2dad96dd374d/width=525/fc2a491c-fb16-4ca8-9fee-2dad96dd374d.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8d3212c8-5a09-4244-ade8-afd6e4b0e8f8/width=525/8d3212c8-5a09-4244-ade8-afd6e4b0e8f8.jpeg" />I checked it with 3 different times (the 1st testing "mochrome, lineart, greyscal" ; the second testing "upper body"; the 3rd testing "close up")And as you can see the weight does have a different impact on case basis. But overall 0.7~1 is the value which gives the most stable output, and which doesn't corrupt other tags.The overall outputs are visibly pleasing, but if you want to really step up with this lora i suggest the use of adetailer or of the upscaler (i personally use ultrasharp cause it doesn't ever betray me, but there are a lot of which gives different results.)That said i hope you will like the model, even toh it is a little heavy it really does its job.As always if you like it you can support me by leaving a recension, liking the model, uploading your creation or even donating buzzs.

Files with SHA256: af7ba6067e511bddda88d997df0136b84bd2815ab017bc73852babf8e6a0184a