#zemai — Public Fediverse posts on home.social

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-08-08 · 04:24 UTC

Techniques to improve prompt entry. #zemAI

With the release of director tools in NAI I wanted to write and cover how to use them. However they're all pretty self explanatory and don't need a guide, just mess around with them and you'll figure it out. Instead this will be a guide on how to better improve your prompt entry and use director tools to replicate images with it so you don't have to look up tags whenever you make AI art.

As mentioned previously NovelAI and anime based models use Danbooru tags. This should be the primary source for the tags you use.

First and probably best way to practice is to take an image off of Danbooru and try to replicate that image using tags. This will help you to better recognize what you need to be doing when writing out a prompt. This is important because to actually replicate an image using only tag you actually have to use tags and phrases not in the Danbooru image you're trying to replicate.

Step one: use the "random" button on Danbooru to get a random image.

In this case I wanted to use official art as an example for this guide. So you can search "official art" and press the random button to get a random official art image.

Step two: use your eyes and take apart the different elements in the picture. Then try to figure out what those tags are.

The random button game me an image of jougasaki mika from idolmaster.

As you can see we can take apart certain parts of the image, she has pink hair, yellow eyes, a hand fan, a flower hair ornament, a purple kimono, a pink obi, she is also sitting down and looking at the viewer. As for the background it seems to be taking place on a sidewalk at sunset. The image also has an orange tint over everything.

There's obviously more identifiable information visible, but this is a good start for this example and we can look that stuff up and enter it later.

Step three: enter in those tags and see what it outputs.

After this step you may find some stuff missing. One major part is the body positioning may not fit the correct angle. This is why you need to practice prompts because body angles are not usually in tags in certain Danbooru images.

The camera position tags (from above, from below, etc.) are usually the primary way to position the body, even if that doesn't seem to make sense. Keep in mind though there are limitations to positioning the body in AI art. But don't worry about that for right now this is just practice.

At this point we should have an image that replicates the elements of the original image.

Now why did we do this? So we can demonstrate how to use director tools. Some of these tools have some practical applications for artists so it's important to recognize how to take an idea to a prompt.

---
The director tools and what they do.

Remove BG:
Isolates a character from an image. Very useful to pull a "sprite" from an image.

Line art:
Converts the current image into a monochrome line art image. VERY useful for an artist as this allows you to study how to draw curtain angles and things.

Sketch:
Less coherent version of line art.

Colorize:
Can change the current image. Very broad in use. Can change outfits, looks, and background with prompts.

Emotion:
Changes the face for a particular emotion.

Declutter:
Apparently removes stuff like text boxes or something. Haven't had a chance to really test it.

For these examples I used remove BG to extract the sprite I initially generated. Then used colorize to insert a background and color theme that fits the image were trying to replicate.

With the introduction of director tools we now have some greater freedom to make the images we want.
__jougasaki_mika_idolmaster_and_1_more__sample-fec314476d1d2f284f4875e02a8b03ef.jpg
1girl, pink hair, yellow eyes, orange background, from side, head tilt, pink han s-2729567701.png
1girl, pink hair, yellow eyes, orange background, from side, head tilt, pink han s-2729567701_bg-removal.png
1girl, pink hair, yellow eyes, orange background, from side, head tilt, pink han s-2729567701_lineart (3).png
1girl, pink hair, yellow eyes, orange background, from side, head tilt, pink han s-2729567701_sketch.png
1girl, pink hair, yellow eyes, orange background, from side, head tilt, pink han s-2729567701_colorize (1).png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-08-04 · 17:32 UTC

More director tools demonstration in NAI #zemAI

There's pretty much no limitations on how you can edit existing images. Even going so far as to edit 3d images into other 3d models.

In the other demonstration you can even edit Ellen Joe's artwork into something that looks like her other in game outfit. :back_from_gab:
so-whats-your-reason-for-pulling-zhu-yuan-v0-iaqrq9fv18cd1.jpeg
1girl, 3d, bikini, s-4079824210 (2).png
be5571cfd72d5dbbdb27a19ba630fda8.png
1girl, 3d, bikini, s-1281454591.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-08-02 · 21:37 UTC

@mactonite @Pawlicker @hispanicweeb @mischievoustomato
Forbidden AI knowledge: secret arts edition #zemAI

You can get nude edits of official art using the colorize feature.

Waifu wearing too many clothes? Just strip them off with this technique.
1girl, s-1406403242.png
1girl, s-1406403242 (1).png
Kuro_2wei (1).png
Kuro_2wei (1) (2).png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-08-02 · 12:49 UTC

@smugumin this will increase my workflow.

I'm now truly evil. AI can now progress to the point where I could make full hentai CGs sets if I really wanted to. #zemAI :back_from_gab:
Noise s-3988101317 (28).jpg
perspective, close-up, from above, 1girl, cow ears, pov hands, holding hands, bl s-1458888463.png
perspective, close-up, from above, 1girl, cow ears, pov hands, holding hands, bl s-1866327226.png
perspective, close-up, from above, 1girl, cow ears, pov hands, holding hands, bl s-1866327226 (1).png
perspective, close-up, from above, 1girl, cow ears, pov hands, holding hands, bl s-1866327226 (2).png
perspective, close-up, from above, 1girl, cow ears, pov hands, holding hands, bl s-1866327226 (3).png
perspective, close-up, from above, 1girl, cow ears, pov hands, holding hands, bl s-1866327226 (4).png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-08-02 · 11:29 UTC

@mischievoustomato @Pawlicker
Also, yes you can do this with Colorize tools. #zemAI

(Don't look, lewd)
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (5) (1).png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (5) (2).png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-08-02 · 11:27 UTC

@mischievoustomato @Pawlicker
New NAI tech found, this is more advanced. #zemAI

With colorize tools, prompts can be used to fully replace parts of images. Making it possible to "pivot" images into different outfits, looks, and backgrounds.

Very useful tool. Game CG like variations just with AI only.
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (5).png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (5) (2) (1).png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (5) (2) (2).png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (5) (2) (3).png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-08-01 · 23:52 UTC

NovelAI released director tools today. #zemAI

They're self explanatory, so they don't need a guide. But basically they're simplified AI tools without you having to mess with controlnet or something. Noob friendly.

They're all pretty useful tools, especially remove background which is highly important to get clean human figures, sprites, and whatever without messing with an image editing tool to clean it up.

Sketch and lineart are good for training your own art.

Emotion change is good for alternate image expressions. Keep this in mind for the future with Aetherroom as they will be using sprite expressions.

I don't have a good example for declutter because I don't mess with comic image sources.
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525.png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (1).png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (2).png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (3).png
1girl, character_chloe von einzbern, classroom, teacher, dark-skinned female, pi s-2375732525 (4).png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-29 · 04:16 UTC

@grips @FruitpilledPeachcel @bronze
I was gonna respond to this earlier but I got massively sidetracked.

>also you got really good with AI :illyahuh:

We're almost at the point where we can make anything with AI. I can do and recreate most things but there's still stuff it can't get right. Still, it took a lot of effort and research to get to this point so I'm happy to have taken the time to learn it.

Have some Terry and Mai Kuro cosplays. #zemAI
1girl, white background, character_chloe von einzbern, terry bogard, dark skin, s-2966824884.png
1girl, white background, character_chloe von einzbern, terry bogard, dark skin, s-1252801282.png
1girl, white background, character_chloe von einzbern, character_shiranui mai, p s-609151735.png
1girl, white background, character_chloe von einzbern, character_shiranui mai, p s-1480692316.png
1girl, white background, character_chloe von einzbern, character_shiranui mai, p s-3842527964.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-24 · 17:41 UTC

AI method: vibe transfer outlining

Don't know if I wrote about this earlier, but I did discover this yesterday.

I've said before that cosplays with AI tend to be a mixed bag and sometimes don't work. However with vibe transfer you can push the AI to use a particular outfit.

Basically, generate the target outfit with the proper skin/hair/eye coloring then use that as the vibe transfer reference (RS 0.1). This will push the AI to make outfits close to the reference, which is good for various things such as the phase Connect brown lady post on Maid001. Normally the AI likes to stick to whatever it is referencing so with vibe transfer you can change that.

Practical use, you can use this to make interesting cosplays. #zemAI :back_from_gab:
1girl, character_ tamamo no mae (fate_extra), dark-skinned female, pink hair, ye s-2649802128.png
1girl, character_ chloe von einzbern, pink hair, yellow eyes, dark-skinned femal s-2291382244.png
1girl, saber alter (ver. shinjuku 1999) (fate), dark-skinned female, pink hair, s-3499289945.png
1girl, character_ chloe von einzbern, pink hair, yellow eyes, saber alter (ver. s-3281581615.png
1girl, character_ chloe von einzbern, pink hair, yellow eyes, saber alter (ver. s-1614853187.png
1girl, character_ chloe von einzbern, pink hair, yellow eyes, saber alter (ver. s-4008168215.png
1girl, white background, character_ miyamoto musashi (swimsuit berserker) (secon s-3678678878.png
1girl, character_ chloe von einzbern, pink hair, yellow eyes, dark-skinned femal s-118378384.png
1girl, character_ chloe von einzbern, pink hair, yellow eyes, dark-skinned femal s-72864877.png
1girl, character_ chloe von einzbern, pink hair, yellow eyes, dark-skinned femal s-1422655577.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-19 · 20:07 UTC

@skylar @hispanicweeb
Also have these lewd realistic AI image tests from a few days ago. They're lewd don't look. #zemAI
realistic, {cosplay photo, photo (medium), real life,} bedroom, 1girl, blonde ha s-3839305809.png
realistic, {cosplay photo, photo (medium), real life,} bedroom, 1girl, black hai s-1464465000.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-19 · 16:56 UTC

@skylar @hispanicweeb
NovelAI is based on Stable Diffusion.

NAIV3 really broke out of the "AI slop" that AI used to have.

Anyway the intention of this realism research is to make unrealistic Amazon queens that look realistic. (Not really) #zemAI
realistic, photorealistic, {cosplay photo, photo (medium), real life,} 1girl, ou s-3355022494.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-19 · 16:27 UTC

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-19 · 01:07 UTC

Using color for stronger AI image realism

Note this is theory and subjective.

Realism, uncanny valley, and cartoon/anime style exist on a spectrum. So when achieving "realism" we first have to get to the uncanny valley.

There's an obvious problem though, "realism" as in a real life photograph of a human model isn't actually "real" and more of a product of the camera used and whatever touching up before the final published product. If this is confusing here is a simple spectrum as an example:

Photoreal. Airbrush
|. |
0--------x-------0--------x------0
|. |. |
Fake. Uncanny real

(No idea if that formats proper)

We're concerned with photorealism and airbrushing. Photorealism doesn't cross over to uncanny because of the coloring used seems too "fake". Airbrushing doesn't cross into uncanny because there's a real base used and therefore has "weight" in the image.

The reason why AI has a problem generating real images is because AI doesn't operate on logic. So it doesn't have the concept of natural weight and shading.

However we can probably solve this problem. The theory being that the techniques used to achieve airbrush realism can be used with photorealism to cross the uncanny valley.

The tags (realism, cosplay photo, photo (medium), real life) can give us a good base but it's hard to develop realism just from that. So we must find tags that can offer the proper shading (the term is contouring apparently) to improve realism. Some examples (not a complete list)

Shiny skin, sweat, oiled - they do what they say. In photography these are used to provide contouring and a good base for airbrushing. On the AI side it can help with the illusion of realism as real photography already does this.

3d, blender (medium), etc - they're not good tags to use because they make things plasticky. They do offer depth to images due to the fake 3d look.

Artist: (enter artist name here), (style tags) - for artists you want to avoid anything heavily stylized with a focus more towards color. For art styles I haven't tested this too much but nothing harsh like game cg, watercolor was one I tried. You can even try attaching the artist prefix onto something else and see what you get.

(Background tags) - outdoors is the primary tag to focus on as the AI handles realistic trees well. Indoor stuff can be messy and ruin the illusion. Basically you have to be mindful of the backgrounds because they help with lighting but can easily ruin the illusion.

Vibe transfer - these can only be used for color. Using a real photo doesn't give you realism unfortunately. So VT should be used to nudge the AI to maintain something about the image. (Color, outfits, environment)

Below are some random tests I've done to nudge the AI past uncanny valley #zemAI :back_from_gab:
realistic, {artist_cosplay photo, photo (medium), real life,} 1girl, small breas s-4146541044.png
realistic, {artist_cosplay photo, photo (medium), real life,} 1girl, small breas s-959091649.png
realistic, {artist_cosplay photo, photo (medium), real life,} 1girl, small breas s-3178322086.png
realistic, {artist_cosplay photo, photo (medium), real life,} 1girl, small breas s-157886924.png
realistic, {artist_cosplay photo, photo (medium), real life,} 1girl, small breas s-2899405667.png
realistic, {artist_cosplay photo, photo (medium), real life,} 1girl, small breas s-2545563171.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-17 · 17:18 UTC

@raccoon
Walking out of the uncanny valley :back_from_gab: #zemAI
realistic, {cosplay photo, photo (medium), real life,} indoors, 1girl, character s-3256547070.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-17 · 17:04 UTC

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-09 · 17:18 UTC

The further away you get from human tags the more the AI branches off from looking realistic. When you focus on realistic body types and looks the AI can pull off a near realistic image. #zemAI
{realistic, cosplay photo, real life}, bedroom, 1girl, close-up, from above, sma s-1070043547.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-09 · 17:15 UTC

Various AI realism tests #zemAI

NovelAI isn't designed for this as the data used is mostly anime art. But because Danbooru has some real people photos you can push it for realism.

Obviously I covered this before, but I figure to try testing negative vibe transfer to push more consistent realism.

The process is:
-generate standard image (no artist no realism just anime style image)
-use that image for negative vibe transfer (low values, around -0.05)
-then generate realistic image. Tags to use...
"realistic" (for color)
"cosplay photo" (focuses on human models)
"real life" (not exclusively for realistic, must be used with other tags)
"photo (medium)" (used for realism, again it's not exclusively for realistic image and must be used with other tags)
"asian" (weak tag, use if you want)
-then use that realistic image as another reference for the vibe transfer. (Around RS 0.2)

You can generate realism without it, but the negative vibe transfer with the positive vibe transfer can be used to strengthen the realism. This isn't an official guide post because I've only tested this minimally.
realistic, cosplay photo, real life, outdoors, 1girl, medium hair, upper body, g s-3678636781.png
realistic, cosplay photo, real life, outdoors, 1girl, red hair, short hair, red s-113473005.png
realistic, cosplay photo, real life, outdoors, 1girl, red hair, short hair, red s-2969597544.png
realistic, cosplay photo, real life, outdoors, 1girl, red hair, short hair, red s-3715765691.png
realistic, cosplay photo, real life, outdoors, 1girl, {{{{large sombrero}}}}, ha s-1067972034.png
realistic, cosplay photo, real life, outdoors, 1girl, {{{{large sombrero}}}}, ha s-1422113887.png
realistic, cosplay photo, real life, outdoors, 1girl, long hair, blond hair, blu s-3211512326.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-05 · 12:54 UTC

BTW before I forget, this masking can be used to create contrasting parts of your image. (RS -0.3, -0.6)

From all the tests I've done about this, this seems to be the most helpful as AI doesn't like to do contrasting stuff like this and helpful for creating more interesting environments. #zemAI
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-2102164635.png
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-156508241.png
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-802223411.png
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-3079178727.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-05 · 12:35 UTC

Also, a blooper.

This technique generated a lewd image accidentally. Feel free to enjoy it. #zemAI
_apple, white background, s-449414980.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-05 · 12:32 UTC

Prompt changes along with this masking can be done but the art style seems hard to change. However it comes up with some interesting stuff. #zemAI
_1girl, city, night, cafe, indoors, dark-skinned female, huge breasts, glasses, s-3328285097.png
_1girl, city, night, cafe, indoors, dark-skinned female, huge breasts, glasses, s-3385200394.png
realistic, photorealistic, 1girl, city, night, cafe, indoors, dark-skinned femal s-3774861750.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-05 · 12:29 UTC

Shape killing with masking #zemAI

From my understanding the way image 2 image and masking works is that the AI considers the totality of the image when making decisions on what to change in the image. Meaning that stuff like limbs can avoid being changed by the AI if the AI thinks it should be there. So you don't have to be clean with masking.

At full negative vibe transfer values and using the original image as the negative reference (RS -1) we can see what happens when only parts of the image are changed. Theoretically the masked part should be completely different.

Of course without any prompt changes a bias is generated with some purple hair smirking girl getting generated. Why? Still don't know. However this art style seems pretty familiar. I'll leave you to think what the art style looks like.
_apple, white background, s-531103701.png
_apple, white background, s-661039911.png
_apple, white background, s-1945961734.png
_apple, white background, s-2246960721.png
_apple, white background, s-908983753.png
_apple, white background, s-1718554056.png
_apple, white background, s-4237454959.png
_apple, white background, s-3986307885.png
_apple, white background, s-2391771003.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-04 · 18:15 UTC

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-04 · 18:08 UTC

On another note... #zemAI

Multiple images can be used for negative vibe transfer. I used the red, blue, and green apple images from above for this test.

HOWEVER

they require extremely low values, as in -0.06 and below RS. Probably meaning that the effects of using multiple negative vibe transfers are additive (yeah adding a subtraction how does that make sense) but this means you can funnel the AI to generate certain things.

From the examples below the AI gets abstract with generation. The apples are small, tattoo like. At lower values (-0.05) the apples are red and small but still retain the shape of an apple.
_apple, white background, s-2721132925.png
_apple, white background, s-1180744623.png
_apple, white background, s-54748971.png
_apple, white background, s-4268622912.png
_apple, white background, s-1517061705.png
_apple, white background, s-2274524709.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-04 · 17:57 UTC

At extreme negative values the AI refuses to generate apples. Instead it opts for monochrome images of a girl usually with a smirking expression.

First 5 images are RS-1
4 after that are RS-0.6

There's only 1 apple I can see in these images, try to find it!

Whatever the case is there's not much to learn from this as we can't explain why the AI draws things like this at these settings. There does seem to be a bias in what it is drawing. #zemAI
_apple, white background, s-2752221349.png
_apple, white background, s-1281305926.png
_apple, white background, s-1281305929.png
_apple, white background, s-1281305927.png
_apple, white background, s-1281305928.png
_apple, white background, s-4020115044.png
_apple, white background, s-4020115045.png
_apple, white background, s-4020115046.png
_apple, white background, s-4020115047.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-04 · 17:49 UTC

Shape killing, a new AI idea. #zemAI

Since we know that negative vibe transfer will subtract colors can we subtract shapes and structures in an image?

If you don't understand think of it like this:
Imagine an apple 🍎 🍏

The apple will generally look a certain way and have certain features, such as round points, stem, red color. Taken further if you saw a black shadow of an apple you would be able to tell it was an apple from the silhouette.

Imagine if I told you to draw an apple that didn't look like an apple, what would you draw? Imagine drawing an apple using only squares, what would it look like? Basically what we're trying to do with the AI, what will the AI draw when it can't draw an apple when told to draw an apple?

First image is the control, and the image used in the vibe transfer.

Next four apples are at IE 1 RS -0.1
As you can see the apples start become more untraditional. With one being green with a golf club, one being blue, one being a different apple angle, and one being an orange that looks like an apple.

Four apples after that are at IE 1 RS -0.15
This is the maximum threshold for negative values the AI can handle (ill explain next post) and at this point the AI avoids traditional red apples and starts to go abstract in objects generated.

The lesson to be learned here is that we can use images to discourage the AI from generating things is a certain way. Doing so can encourage creativity with AI generation.
_apple, white background, s-1968115615.png
_apple, white background, s-3744394324.png
_apple, white background, s-3744394325.png
_apple, white background, s-3744394323.png
_apple, white background, s-3744394322.png
_apple, white background, s-1616520682.png
_apple, white background, s-1616520681.png
_apple, white background, s-1616520683.png
_apple, white background, s-1616520684.png

#zemai

:back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War @[email protected] · 2024-07-03 · 16:12 UTC

More of the minus vibe transfer world. #zemAI

At full negative RS the image trends towards a certain direction. These arts seem to be more "gacha character card/wallpaper" official art in how they look. The AI starts to go eccentric and colorful in the environment.

Still, there are questions:
Why the purple and gold?
Why generate a text logo?

But some conclusions:
Negative values don't hurt the quality of the generated images.
Negative values also can be used to control color in generated image, and maybe preferred to be used in some cases.
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-3169617676.png
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-3169617677.png
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-3169617678.png
artist_ realistic, indoors, cafe, night, city, 1girl, glasses, huge breasts, bla s-3169617679.png

#zemai