Hi there, here's another of my nieche lora. This one is for the generation of nana, from "Nana to Kaoru" serie.

This Lora was made for testing the possibility of training one character and different styles together, on civitAi trainer and i must say that the result is far better than expected.

Disclaimer: even toh this lora isn't aimed to produce nsfw it can be used to generate those kind of images, so be carefull :)

Base version

This version has different keywords:

n4n4            <-- It's the keyword to obtain the character

bow             <-- (optional) it's usefull when using low Lora Strenght



----------------------------------------------------------------------------



from_anime      <-- this tag was trained successfully, but the images used to train it were low quality images, so it now works as a negative tag (watch the details section to have a rough idea)



----------------------------------------------------------------------------



from_manga      <-- This tag should be used if you want to portrait nana with the style of the manga, it works for monochrome images or normal ones (if you want to produce a monochrome image, it should be used with "monochrome" and "greyscale" tags)



----------------------------------------------------------------------------



from_illustration <-- it makes the image style more similar to the illustrations of the serie, it has a soft bias for backgrounds (more specifically color gradients background), so if you want a specific background describe it in the prompt, and don't use only generic expressions as "simple background"

The Lora Strenght should be set between 0.5 and 1 depending on the desired result (for more details check the "details" sections), i suggest a strenght around 0.7 - 0.8.

I personally tested most of the time using euler a sampler, but it shouldn't have any major issue with other samplers (if those works good on the model you are using).

As concerning the models, it works decently on most models, but depending on the model the Lora Strenght might vary (for example using hassaku you need to lower the lora streght or you'll get oversaturated images)

Details:

This version was trained from 33 images, around 10 images per source (manga, anime and illustrations) so that the model could have a solid idea of the character and a rough idea of the style.

The model has been tested many times to check if there are any bias and those are the results.

(The tests in this section were made using anylora ckpt pruned), at 512*768, 40 steps, without the use of inpaint,highres.fix or adetailer)

Test 1

The first test consisted in verifying if the lora had acquired the n4n4 concept, and as you can see starting from 0.5 the impact on the generation becomes clear, and works as intended (ps. there wasn't an intended outfit for the characters, so you might need to specify things like "school_uniform" or "white_shirt").

The tests from 3 to 5 have the same lora weights, in the same order (+ they test the results in both positive prompt (on the topside) and negative prompt (on the bottom))

Test 2

The second test was to check if the lora had any bias for camera angles, so using lora strenght of 1, we've tested multiple times per angle, and it looks like there's not any problem when the camera angle is specified. When it's not specified it tends to generate "upper body" images, similar to "cowboy shot" + "portrati" camera angle.

Test 3

The 3rd test was aimed to check the results using "from_anime" tag in the positive (top images) and in the negative prompt (bottom images).

As you can see between lora strenght 0.5 and 0.7 the tag start impacting heavely on the style. When used in the positive prompt it will produce a low quality images, but when used in the negative, it can make the generation more stable.

Test 4

The 4th test was aimed to check the results using "from_manga","greyscale" and "monochrome" tags in the positive (top images) and in the negative prompt (bottom images).

As you can see the style match the style of the manga when the tags are used in the positive prompt, and when used in the negative, it will bring the model towards the generation of images more "illustration" /"manga_cover" oriented <-- without the "issue" of having the face of nana with thick lips (as it usually is in the illustrations)

Test 5

The 5th test was aimed to check the results using "from_illustration" tag in the positive (top images) and in the negative prompt (bottom images).

Both cases are pleasing, but i have to state the using it in the negative prompt will produce images kinda nsfw.

Note: on most of test i've continued to use this tag (and "from_manga" tag, some times even combining them obtaining cool results), but you might need to tweek the strenght of this tag to achieve some specific results (for example, i had a cool image, but the mouth of n4n4 was generating weirdly, that's because the angle was too narrow and the "from_illustration" tag was trying to generate a thick kind of lips, in an image it couldn't handle it; so the solution was to lower the weight of the tag to around 0.7 - 0.8)

Test 6

This test was made to understand what would've been the outcome obtained by mixing the style tags ("from_manga" and "from_illustration in the positive prompt; "from_anime" in the negative).

The results are stable, even with lora at full strenght. But when you use this strategy at lora weight 1, you might wanna tweak the style tags strenght. A good rule of thumbs would be that the sum of the 2 strenghts doesn't go above 1.5

Conclusions

This isn't my best lora, and it won't probably be usefull to most of you; but i wanted to still share it for few reasons.

The first reason is that the output can be decent, and i haven't found another lora of this character around.

The second reason is to remark the importance of a good dataset -->the images from the anime used to train were too bad, and probably not using those would've made the lora even better.

The 3rd reason is that i genuinely think that testing a lora and trying to resume all the toughts and effort you put in it will increase the amount of things you'll learn from it.

As always if you like the model i encourage you to share your generations and to tip some buzzs to support me :).