[AI] Graphics and image generator on your own computer - web interface for Stable Diffusion
.
How do I generate images on my own computer? How to change the content of a picture? Is AI able to suggest an interface sketch for a smart home controller? This is what I will try to find out here, all thanks to an easy-to-use web interface based on the Stable Diffusion XL architecture.
Presented here Fooocus is a completely free, open and 100% local (no internet, on your computer) environment for creating images using AI models based on Stable Diffusion. Fooocus offers, among other things:
- image generation based on prompts
- GPT2-based prompt development
- image upscaling, i.e. increasing the resolution
- outpainting, i.e. adding parts of an image outside its frame
- inpainting, i.e. editing part of an image
- variation, i.e. creating different versions of the image
- image prompt
- describe image, i.e. describing the image to the prompt of the selected model
- negative prompt, i.e. what you don't want in the image
- the possibility of integrating LORA add-ons, styles and setting their weights, which gives us greater control over what we generate
Install and run .
Download the package from the repository , extract, run.bat:
.
In newer versions there will also be separate scripts run_realistic.bat and run_anime.bat, they work in the same way - just select the style of graphics at the start.
After a while in the browser our Fooocus page will open:
.
Time to start the fun.
First generations .
I must warn at the outset that the graphics take quite a long time to generate. On mine on an Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz, 64GB RAM, GeForce GTX 1060 they take about three minutes to generate.
We simply type in the prompt and select the options. Sometimes you have to scramble to get the desired effect. In addition, the keywords used affect the style of the whole image, so it's hard to make one 'series' of graphics. But let's give it a go, first a sample prompt. All generated by my laptop.
Model used: juggernautXL_v8Rundiffusion.safetensors
A cozy cabin in a snowy forest at sunset, warm light glowing from the windows, smoke rising from the chimney
.
Pretty good, but why is the chimney on fire?
Futuristic city skyline at night, neon lights reflecting on wet streets, flying cars in the sky
.
Colour scheme great, but the sense of vehicles and backgrounds is quickly lost.
.
Astronaut floating peacefully above Earth, with stars and galaxies in the background, contemplative mood
.
Strange details again.
.
A cat café with cozy decor, wooden furniture, and playful cats lounging around while people sip coffee
.
Here you can see the artefacts much more strongly, you often end up like this trying to generate a scene with many characters.
.
A cute fox wearing a small backpack, wandering through an autumn forest filled with orange and red leaves
.
Quite good, just always a slightly different character.
.
A massive floating island with waterfalls cascading into the clouds, bioluminescent plants glowing softly
.
.
A colossal mechanical dragon flying over a steampunk city, gears and smoke in the background
.
Electronics through the eyes of AI .
Just out of curiosity - does AI know what an Arduino looks like?
arduino, electronics, table
.
Something must have been in the teaching data, but still the heresy comes out....
And the diagrams?
schematic, full bridge rectifier, graetz, electronics
.
I wasn't expecting much, but this font is cool too. Still, this
"RECCTITICIFER".
The next generation has done... a picture of the element:
.
How about AI generating us retro hardware?
unitra, radio, retro, table, receiver
.
Here's maybe another weird experiment....
polish car, fiat 126p, retro, polish city, street
Inpainting .
Inpainting allows us to modify or completely change part of the image content. For example, we can add a travelling companion to our fox. To use inpainting, turn on the "input image" option and drag the image we want to edit there. We then select the zones we are going to edit.
.
The prompt I gave was similar to the one before, only now it included the word "owl".
The first two results:
.
Trials of practical application .
Finally, I tried to use this generator to create sketches/proposals of various kinds of devices for a "smart home", for example maybe some kind of heating controller, weather station or some other type of touch panel there. I know that the text of these models does not generate readable text, but maybe some idea will give me?
smart home, ui, design, chart, heater, interface, screenshot
.
From a distance it still resembles something, the colour scheme is probably ok too, but the text and details are completely wrongly drawn.
Summary .
To was a short presentation showing what images could easily be generated on typical consumer hardware. It didn't turn out to be that difficult at all and the results are pretty good too, although I'm sure opinions will be divided. Out of curiosity I also tried to generate slightly more practical things, but from this angle the model used limps along so I wouldn't expect much from it.
I'll leave the final verdict to you, only here I'll emphasise that all generation is 100% local, as well as highly customisable and modifiable, so it might be worth taking an interest in, if only out of privacy concerns.
Do you use image generators and if so, what for?
p.kaczmarek2 wrote 12174 posts with
rating 10128 , helped 580 times.
Been with us since 2014 year.
Comments
Cool stuff. I'm going to test it on my "stationary" tomorrow. Especially interesting are these additional options in addition to the usual image generation. Maybe you will get used to it and make your... [Read more]
You certainly need to be patient, as generating good graphics can take a little more trying than with ready-made, closed generators that are not available for download. It is also a good idea to disable... [Read more]
I like the owl-shaped backpack the best (something went wrong here ;) ) and that cat: https://obrazki.elektroda.pl/7909671700_1742286449_thumb.jpg https://obrazki.elektroda.pl/5886031800_1742286461_thumb.jpg... [Read more]
. A while ago I installed myself fooocus with the default models (it actually installed itself). I was most curious to see what the difference in generation time would be between the laptop GTX1060 card... [Read more]
. The character's hands (fingers) are a typical example, and as I tested, the wires cause problems too. The more complex the concept the worse, but it also depends what was in the teaching data. I... [Read more]
Lol, good one! Neither Bing nor ChatGPT actually can (both are the DALL E model). Even after pointing out the error and asking to raise the wine level by 100%,. The glass is always half full. [Read more]
Apparently ChatGPT 4o has the ability to generate simple templates for 3D printing, has anyone checked how far this stands out or is suitable? Let's give it a simple phone stand.... [Read more]
There are some such generators, but I haven't tested them, I don't have a printer :) E.g. meshy.ai [Read more]
This illustrates how influential the learning data is. By manipulating the learning data, you can strongly influence the performance of the model. It's not so much 'real' intelligence as some people think,... [Read more]
Unfortunately you are right, and this is the biggest threat from any AI. If people start to believe uncritically what AI generates, whether in terms of information or images, it is a very simple way to... [Read more]
In my opinion, a good test of these generators is requests for technical and application graphics, e.g. "spectrum of a high-pressure sodium lamp", "phase diagram of water", "inside of a quartz resonator",... [Read more]
There is a lack of a model for technical drawings, let alone diagrams. There's quite a niche here, it will be interesting to see if they ever get around to it. Although it won't be easy, because how do... [Read more]
If the AI were fed with the diagrams described, it might understand something. At the moment, based on a photo, it is able to match a part to a motorbike and even determine the brand, although it does... [Read more]
Playing AI on a graphics card with VRAM=6GB had to end up like this :) On this hardware you run SD 1.4 or 1.5 - i.e. relatively "poorly educated" models, from 2-3 years ago. SDXL 1.0 already needs 8GB... [Read more]
The demand for AI has definitely had a negative impact on GPU prices, which of course is classically lost on gamers. [Read more]
This is influenced by many factors, AI being just (or as much as) one of them. I personally have been unlucky, due to a crash I had to buy graphics in 2022. (peak of cryptocurrency digging + covid) and... [Read more]
I for one regret not selling my RTX 3070 in the peak as used ones were going for 6-7k, I bought it literally before the price spike as it was December 2019 :D . If I sold it then, I'd probably have a... [Read more]
I would rather not expect a good AI model for technical drawings/schematics any time soon. Text is easier to generate, because it is enough to predict successive tokens, but a technical drawing? A schematic,... [Read more]
. Not to look far - model o1 for my prompt: "Generate a schematic of an electronic circuit that, when connected to 230V AC power, will blink a red LED in a system of 2 seconds on, 1 second off. To this,... [Read more]