Graphical neural networks

Graphical neural networks

Neural networks are developing rapidly, and those results that were just fun and psychedelic-like a few years ago have now completely changed. Graphical neural networks produce excellent results and, of course, this topic is very much discussed by all designers and not only. I have been testing and studying different neural networks. Trained on cats.

I made this image in Photoshop 25.0 Beta AI by request:

I went through different options and modded, replacing the option with a tie.

Neural networks I reviewed with cats:

BlueWillow
Leonardo.ai
StableDiffusion on Google Colab
clipdrop.co Stable diffusion
StableDiffusion on local server
stablediffusionweb.com
EasyDiffusion
FusianBrain
replicate.com Stability aistable diffusion
Photoshop 25.0 Beta AI

To demonstrate the results and to further compare them, I created one query:

I tested in all free or semi-free neural networks and Adobe Photoshop 25.0 AI.

The bow tie does not have to be blue, only the eyes are specified as blue, but many neural networks apply blue to the eyes and to the tie and to the background, add lighting in blue tones or other colors that you specify as the color of one element in images In this case, I added other requests or wrote a negative blue tie prompt. Also, if you prescribe, for example, blue eyes and a golden butterfly or burgundy, the eye color can be changed to gold, and the tie can be set to blue, or gold objects or burgundy fabrics can be added to the interior.

Of all the neural networks I tested on a cat with a tie, only Adobe Photoshop 25.0 Beta AI generated a non-blue tie and the correct blue eyes. With other neural networks, it was necessary to go through combinations of requests and change the color of the eyes and the tie differently. Each neural network shows a certain model with which it is trained and changes its appearance, perspective. Some neural networks have the ability to select, add new models, and choose different design styles.

BlueWillow

BlueWillow в Discord generates an image as in Midjorney according to the principle of sorting through the 4 proposed options. But the results in BlueWillow are not the same as in Midjourney.

По запросу:

Discord's BlueWillow neural network generates images that look more like vector graphics than realistic photos. It can be seen that there is a red tie and three variants of blue, only two cats have blue eyes and wool is not gray, but with different patterns, one cat has dark wool and is dressed in a suit. Clarifying request:

Three of the options have a golden butterfly and only one has blue-green eyes. Next, I choose one option and make a clarifying request for eye color:

I can't find the option to register a negative request. For example, if I want to make gray wool on a cat, and all the images have patterns, white face, white wool on the body, I can't change this pattern, I need to change the request. Other images on request:

They do not respond well to the request. I continue to search in options with a gold tie and blue eyes and improve the result to blue eyes.

To get better results, you need to improve the queries and test the neural network a lot.
The BlueWillow neural network generates inaccurately, all images are in the same graphic style, they do not look like realistic images, it takes a long time to select what you want to get. For example, one image option is selected and a clarifying eye color is set, but the neural network partially replaces the eye color with blue, but not as bright as needed and changes the other background from abstract to nature.
Of all the proposed options for cats, there is no cat with a pure color of gray wool without patterns and white colors.
It may be beautiful, but it doesn't answer the question.

There is an interesting possibility of mixing several images to get the result.
BlueWillow does this, but the results are not at all like Midjorney, Midjorney is very creative in this regard.
BlueWillow's results are worse.

Leonardo.ai

I generated 4 variants of cats in Leonardo.ai by request with a negative request.
Generated several bow ties, put in a negative prompt.
The results are shown quickly, and I am satisfied with their quality.

You can immediately see what a royal result, what bright blue eyes the cat has, they are definitely blue, not brown, not green. You can see the detailed pattern of the tie, how noble the cat looks in all the images, and the interior itself is not indicated, and by default it is also beautiful, fabulous.

The differences between Midjorney and Leonardo.ai can be understood when you test Midjorney.
The Midjorney neural network is paid, it mixes the image well, can use several images, text requests and generate graphics based on them, offering very realistic and creative ideas. I tested how BlueWillow mixes one image in discord and adds text requests with other images to them. The results are not as realistic as in Midjourney. They are similar to the results of Stable Diffusion, when there are no models, the mixed images turn out to be inaccurate, extra elements appear that are poorly detailed.

The results are not as realistic as Midjourney. They are similar to the results of Stable Diffusion, when there are no models, the mixed images are inaccurate, extra elements appear, poorly detailed. I haven't tested Leonardo.ai and BlueWillow neural network much, maybe you need to get used to them.
After three or four generations of 4 images, I got a message about the need to top up tokens. Free tokens for 8 hours are enough for a small number of generations for one-time tasks, not for a large number of raids.

Midjorney and BlueWillow work in discord, a bot is created there and requests are written in a separate private channel, then one of the 4 results is selected and the requests are clarified. You can write new requests based on the proposed results, see the history of requests. Leonardo has a discord channel, but I could not generate requests on the private channel, they are not entered for some reason. Midjorney (paid) and BlueWillow (free) have the /imagine prompt command and a request on discord in Leonardo I did not find such an opportunity.

Leonardo.ai generates by text

all cats with blue ties and at the same time adds extra ties, so for testing the neural network I improved the query with the burgundy butterfly:

If you do not specify a negative prompt, options with several ties are obtained.
The negative prompt is not set initially by default and is hidden in prompt magic.
There is also a mismatch in eye color. I generated this from another attempt, so I chose a different model.
150 tokens are given, the generation of one 1536x1536 image costs 26 tokens, there are different styles: Cimematic, Creative, Vibrant.
For some reason it generated with two butterflies, one blue, the other maroon.
The cat is realistic, but the eyes are not blue, the quality of the image, the realism suits me, the accuracy of the request does not suit me.
If you generate on the wrong model, the results are terrible, I set the default SDXL 0.9 and the cat's eyes and nose were deformed, extra paws appeared, defects appeared in various image elements.
You need to test which model is suitable for the image, there is a choice.

I changed the generation model for the test to Absolute Reality

After several unsuccessful options, the result turned out to be quite decent, the cat is gray, realistic, with blue eyes, the color of the butterfly is suitable, only the type of tie is not specified.

Tokens run out fast because it takes a lot of testing to get the right results.
Tokens are updated after 8 hours for 150 tokens.
Made a new prompt:

If you change the request and add a gold tie, Leonardo does not understand it, the neural network changes the tie to yellow and at the same time changes the eyes, and sometimes adds a yellow interior.
In half of the cases, eye defects and tie defects appear, the neural network tries to put the blue and yellow tie together. There are cases of graphics defects, individual parts of the body are stretched, objects are divided into several parts.
To get a clear result, you need to test the queries.

I chose the last version of the cat, finished it in Photoshop 25.0 Beta AI, changed the eye color to blue as requested, rearranged and finished the tie, added ribbon around the tie, added fur to the cat.

StableDiffusion with the model AbsoluteReality

По запросу:

The tie is always generated in light blue and its color changes only after a clarifying request.

I've added variations of the gold bow tie, burgundy royal:

It is possible to complicate the queries and improve the background, styling, but the results are clear, the same cat model.
When trying to change the color of the butterfly to gold:

the color of the tie still remained blue,
changed the request to:

added some ties, added a negative prompt:

The angle has changed a little, but the tie itself does not change in color.

The angle has changed a little, but the tie itself does not change in color.
Adds elements of gold to the interior, but not the gold color of the tie.

When specifying the color of the tie to yellow, it adds yellow color to the interior, can show two cats, can greatly change the shape of the tie, a yellow tie is added by highlighting (((yellow))) with three brackets, after that a yellow tie and blue eyes appear, but this tie shape is rarely generated as a butterfly.
Sometimes several cats or body parts may appear, there may be mixing of one image and parts of another image from the next generation.

If you make the following request: "Cute Gray Cat with blue eyes, wearing a (((yellow bowtie)))", then mixes the parts of the butterfly with the cat.
If you remove a few butterflies in the negative direction, you will add a lot of them to the cat.

More specific request:

To determine the desired option, it is necessary to test a lot, there are various disadvantages, for example, whiskers can smoothly transition into yellow tissue. Two more cats can be seen on the shelves in the interior.
The last two images are interesting in my opinion, I refined them in Photoshop.

clipdrop.co/stable-diffusion

clipdrop.co/stable-diffusion

I tested with the standard request:

It can be seen that all the cribs are generated by blacks, adding a prompt:

The neural network created 4 images, 1 with a gold tie, 3 with a blue one, even created a gold ornament on one.
Sometimes the images are detailed, sometimes they are blurred, the image with the gold tie is blurred.
You can see that the graphics are very detailed, it is similar to leonardo.ai and at a larger size these details should be visible.

New request:

Gives a clear burgundy tie, blue eyes, the color of the cat's fur is light gray or beige with spots of white. Some images come out detailed and you can see that there will be good quality when the image is enlarged, but some images come out blurry, where the background is the same color.

The neural network allows you to generate 12 1024x1024 images for free during the day. In incognito mode, you can generate additional generations and in other browsers as well. There are different styles: Anime, Photographic, Digital Art, Comis Book, Fantasy Art, Analog Film, Neon Punk, Isometric, Low Poly, Origami, Line Art, Cinematic, 3D Model, Pixel Art.

I added a request:

Anime

Photographic

Digital Art

Comis Book

Fantasy Art

Analog Film

Neon Punk

Isometric

Low Poly

Origami

Line Art

Cinematic

3D Model

Origami

The neural network generates 4 images per minute, the queries are not accurate, you need to set negative prompts and sort through the results a lot, it adds a lot of gold elements depending on the chosen style.

StableDiffusion

StableDiffusion is installed on a local server without a model.

I tested with a standard request:

The result is inaccurate: green eyes, the cat's head is too dark from above, turns into a light color from below, instead of the cat's body is a human shirt.

stablediffusionweb.com

https://stablediffusionweb.com/#demo

Generates each 512x512 image for free in about 20-30 seconds.
There are 94 different cat styles on the site that I checked.
Since there are many styles, I did not add a negative prompt, all cats came out with blue eyes and a blue tie, although I wrote the following prompt, where the color of the tie is not specified:

cinematic-default

cinematic-default

ads-advertising

ads-automotive

ads-corporate

ads-fashion-editorial

ads-food-photography

ads-luxury

ads-real-estate

ads-retail

artstyle-abstract

artstyle-abstract-expressionism

artstyle-art-deco

artstyle-art-nouveau

artstyle-constructivist

artstyle-cubist

artstyle-expressionist

artstyle-graffiti

artstyle-hyperrealism

artstyle-pointillism

artstyle-pop-art

artstyle-psychedelic

artstyle-renaissance

artstyle-steampunk

artstyle-surrealist

artstyle-typography

futuristic-biomechanical

futuristic-biomechanic-cyberpunk

futuristic-cybernetic

futuristic-cybernetic-robot

futuristic-cyberpunk-cityscape

futuristic-futuristic

futuristic-retro-cyberpunk

futuristic-retro-futurism

futuristic-sci-fi

futuristic-vaporwave

game-bubble-bobble

game-cyberpunk-game

game-fighting-game

game-gta

game-mario

game-minecraft

game-pokemon

game-retro-arcade

game-retro-game

game-rpg-fantasy-game

game-strategy-game

game-streetfighter

game-zelda

misc-architectural

misc-lovecraftian

misc-macabre

misc-manga

misc-metropolis

misc-minimalist

misc-monochrome

misc-nautical

misc-space

misc-stained-glass

misc-techwear-fashion

misc-tribal

misc-zentangle

None

papercraft-collage

papercraft-flat-papercut

papercraft-kirigami

papercraft-paper-mache

papercraft-paper-quilling

sai-analog-film

sai-cinematic

sai-comic-book

sai-craft-clay

sai-digital-art

sai-enhance

sai-fantasy-art

sai-isometric

sai-line-art

sai-lowpoly

sai-neonpunk

sai-origami

sai-photographic

sai-texture

Easy Diffusion

https://easydiffusion.github.io

I installed the program on the computer. I tested on standard model sd-v1-4. Requested by:

By specifying many ties in the negative prompt, half of the tie appears and the second tie is located much lower. The color of the tie is blue instead of gold. The eyes are bright blue instead of blue, and the cat has a patch of white fur in the front instead of gray. Added a prompt with a golden tie, out of four options only one option generated correctly.

The poses of the cats are bad, the cats are toy, sometimes the tie is missing.
I have added a model absolutereality_v181.safetensors з site в Easy Diffusion *. I wrote a prompt:

The tie is still blue, you need to test the new prompts.

Fusian brain

fusionbrain.ai

I made a request:

Generates images quickly, I tested 4 options without style, one angle of approximately the same format came out. The images all come out similar, the angle of view, the pose, the pattern of the texture of the cat's wool are the same. The cat's eyes are brown, although in the request they are blue, the color of the tie is blue everywhere, although the color is not specified anywhere. The color of the wool in some images is more light brown than gray. But considering the free quick result, you can edit something, the image size is 1024x1024 with no usage restrictions.
You can pay attention to the fact that there is only one tie and there is no problem with their number, location, deformation of the body, several cats, etc.

Additional request:

generates golden eyes and a blue tie. The result is high-quality, fast, but does not meet the request and needs to be changed somehow so that the neural network understands that exactly blue eyes and a golden tie are needed.
Also, the cat's wool is not completely gray, but with patterns and white shades.

Additional request "Cute Gray Cat with blue eyes, wearing a royal gold bowtie" generates golden eyes and a blue tie.
The result is high-quality, fast, but does not meet the request and needs to be changed somehow so that the neural network understands that exactly blue eyes and a golden tie are needed.
Also, the wool of the cat is not completely gray, but with patterns and white shades.
It should be noted that there is only one tie and there is no problem with their number, location, deformation of the body, several cats, etc.

After several attempts, the neural network creates successful blue eyes and a golden butterfly, but adds blue and gold shades to the interior.
The neural network does not understand which color refers to which object.
After many different attempts (more than 20), blue eyes appear

Fusionbrain has an editor, where you can delete the area with the eraser tool and add a new request, I deleted the golden eyes and wrote the request in blue, but the eyes came out blurry, unrealistic, I finished the last image in photoshop.

When trying to change the color of the butterfly to burgundy: Cute Gray Cat, blue vivid eyes, Royal burgundy bowtie
the neural network creates a red interior, while the eyes become a different color and the butterfly turns blue again.
The neural network does not understand that the tie should be red and paints the interior in red, you have to go through the results or somehow change the query to find the results:
Cute Gray Cat, blue vivid eyes, bowtie of burgundy fabric texture
Cute Gray Cat, blue vivid eyes, burgundy fabric bowtie

The neural network generates a bright red tie, but adds a lot of red to the interior and the eyes are not blue. We need to improve the request and test the results.
Fusionbrain.ai has different styles, I generated them to evaluate the results.

CyberPunk

Anime

Christmas

Kandinsky

Aivazovsky

Malevich

Picasso

Goncharova

Classicism

Renaissance

Oil painting

Pencil drawing

Digital painting

Modieval painting

Soviet catoon

3d render

Cartoon

Studio photo

Portrait photo

Khokhloma

Stability aistable diffusion

replicate.com/stability-aistable-diffusion

I made prompt:

Size up to 1024x1024, free.
Generates quickly, but makes mistakes: cat without a butterfly, camera view of the cat is very close, two heads of the cat, made the gradient of the eyes yellow-blue. It is necessary to add additional requests and test them.

The results are fast, interesting, but do not match the request.
Other generators работают долго. Пишет "This can sometimes такое 3 to 5 minutes при modelе boots up."
Some of the results I tested were also not clear.

Photoshop 25.0 Beta AI

What's new in Photoshop Beta 25?

I made a request:

I tested different results, chose this one

I enlarged the area, refined the ears, the bottom of the body.

Generates fast and clear results in a realistic style of three options with the ability to generate other options.
Produces quality results, angles are different, images are not duplicated, different ideas are generated. For example, it does not show a blue crib in all images, but offers various realistic ideas that can be modified. Even if the butterfly bed of a different type is generated, you can make a new request specifically for the style of this bed, for example, "bow tie gold color" or "bow tie burgundy royal look".
Disruption with body proportions, lack of tie is not noticed, does not add blue or other color to the background.
Different neural networks generate different models of cats. I tested options with different colors of ties, it offers everything correctly.

I complicate the query:

Images are generated quickly, there are occasional failures, the tie may be on the head or elsewhere, the defects of compliance with the request are minimal.

Photoshop 25.0 Beta AI distinguishes between cat breeds, complicating the query:

In the last image, he finished the eyes, bow tie and ribbon.
If you delete part of the image of the tie and add the request "yellow bow tie", Photoshop 25.0 Beta AI adds part of the tie in the selected area, understanding that you need to restore part of the image, does not add a new image of the second tie to this area.