AI Image Generation Beginner Guide

Hi class, today we're learning about AI image generation !

So, you've just stumbled into the fascinating world of AI image generation and might be feeling a bit overwhelmed with all the models, prompts, and other tricky concepts that seem to be blocking your way to creating amazing AI pictures.
Don't worry, I'm here to help you get past that initial confusion. By the time you finish reading this article, you'll have all the know-how you need to get started on generating some truly incredible images!

Summary : 

• What are base models ?
• What are models / Finetunes ?
• How should I prompt ?

 What are base models ?

Base models are the first layer of artificial intelligence. 

Pre-trained on a very large dataset, they are the building block of your generations. 

Here's the three most popular base models currently available on Dezgo

1. Stable Diffusion 1.5 (SD1.5) : Released in august 2022, SD1.5 was the first groundbreaking popular base model trained by Stability AI. 
   While it lacks in quality compared to most recent base models, it makes up with the massive community driven resources based on it.
   
   
2. Stable Diffusion XL (SDXL) : Released in july 2023, SDXL is a significant step forward in quality compared to SD1.5. 
   SDXL is natively trained with 1024x1024 pictures compared to the 512x512 of SD1.5 meaning you're able to generate pictures twice the resolution.
   
   
3. Stable Diffusion XL Lightning (SDXL Lightning) : Released in january 2024, SDXL Lightning is an architecture based on SDXL. 
   In artificial intelligence, larger resolution means exponentially higher cost.
   With SDXL being almost 10x more expensive than SD1.5, Lightning is an ingenious tweak to create pictures based on SDXL way faster for a minimal quality loss.
   
   

🤔 Alright so which base model should I use ?

• If you want the highest quality overall, go with SDXL.
• If you want the best quality over price ratio, go with SDXL Lightning.
• If you need advanced tool and large community support, use SD1.5.

 What are models / finetunes ?

A finetune refers to a base model that received a small additional training by the community.

Retraining a base model is very helpful in order to specialise a base model. It's also way less expensive to do than creating another base model ! Hence why there is so many of them.
For example, while SDXL has a large understanding of our world concept, it won't necessarily be that good at creating things we humans consider beautiful.
So if I retrain SDXL with an emphasis on a realistic dataset, I'll be able to have a finetune with way better results, like so :  

Result with SDXL base Model

Result with an SDXL Finetune

Here's the prompt if you're wondering :
"Art by florist Mandy Disher, digital illustration, portrait of a beautiful lady, perfect anatomy, centered, red lemon theme, multiple shades of yellow, burst of lemon zest and juice, close to perfection, dynamic, very detailed, watercolor, ArtStation, concept art, blur, sharp focus, illustration, Carne Griffiths, Wadim Kashin."

The finetune I used was Juggernaut 10. It's specialised on realism and you can use it for free and unlimited use on Dezgo.

We currently host 100+ finetunes so choosing one will be more tricky compared to choosing a base model. I recommend to start with the category of the finetune you are interested in (anime, realistic, artistic...) then switch around until you find the one you like the most.
Almost all of these 100+ models are used so your own tastes will be the deciding factor there.

 How should I prompt ?

Writing a good prompt is the most important deciding factor when it comes to AI image generation. 
Good thing it's also the factor you have the most control on so let's break it down !

Now I could probably write a whole article only dedicated to good prompting technique (let me know if you'd like to read it) so let's focus on the few things that matter the most with prompting.
You're in luck, you only have to remember two rules.

1. Be as precise and specific as possible.

AI generators can't read minds (yet).
If your prompt is too vague, your result will be underwhelming. 
To solve that, give the AI as many details as you can think of and see how much of a quality jump it'll make : 

A parrot flying

anime artwork of a parrot with vibrant color flying at night on a snowy mountain, full moon in background, purple snow

Model is Envy Starlight by the way, one of my favorites ! 

If you feel like your AI generations are too bland, definitely add as many as these details as you can think off.
Talk to the AI like it needs to know everything. Mention a style, a context, a background, colors etc...

2. Find the most powerful keywords.

During the training of models and finetunes, it's important to understand that pictures present in a dataset are tagged with a series of words.
These words are what the AI will remember in order to that a parrot has wings or that a full-moon is a round object.
The tricky part is to understand that some words you use yourself to describe an object can be absent from the dataset of that specific model. 
Maybe it's better to write "white moon" instead of "full moon" ?
Maybe there's more images tagged with the word "Macaw" instead of "Parrot" ?

The whole game consist in noticing what doesn't look good in your final picture and tweaking the corresponding word in your prompt with synonym until you find that one word that just perfectly works.

What about the rest ?
Guidance, Steps, LoRAs, Sampler, Seed ?

Well, if you already want to tweak those, then you are not a beginner anymore.
For every beginner reading this guide dedicated to them, my suggestion is to not play with those settings. Dezgo will help you in trying to always set the best value for those so don't worry about them too much.

And for anyone else, if one of these concepts still is unclear to you, don't hesitate to leave a comment and I'll write an article on them.