How to create AI pictures by yourself
First, I have to admit, that there are millions of pages, which do offer the service to generate artificial intelligence pictures for you! Maybe that is the good and maybe that is the bad news. The AI business is turning very fast and there is much going on; even the big boys like Google are coming up with their own AI picture software. The most pages you must pay for these services, and you will get standard pictures of standard avatar`s (in case you are looking for an avatar picture for a MMO game for you). A little bit later I will give you numbers; to understand how expensive this can get. The quality of the pictures is not satisfying at all. There is one page, which has one of the best pictures of all – Midjourney!
If you did already generate more than 5000 pictures at their discord server, you will have access to their alpha service. How I understand you will have to pay 25 bucks per month if you don’t fulfill this requirement. In my mind Midjourney is leading the market of AI picture regarding the quality. The pictures come with an amazing level of quality and details! Midjourney is not the best with “Controlling the AI” – means to influence the AI to do what you want, or to create a picture how you do need it.
An open-source alternative is Stable Diffusion. It is a deep-learning text to image software which is free to use, and Stable Diffusion does not claim rights of generated pictures. You really must think about what you want to do with AI. Just to give you numbers. In the past 10 months I did generate about 2 million pictures! And I am not talking about the pictures, when I did abort the generation process – that is something you will have to practice. However – about 4000 pictures of the millions I did generate are in my mind preview able. Maybe I am not good with prompting xd 😊 ! That is not very much, and this is very time consuming, and you will need – yes – a good hardware!
You can download the Stable Diffusion software for free! Cause Stable Diffusion is very slowly to work with, and it takes some time to load. I did decide to work with Comfy-UI! Comfy-UI is about 50% faster and I do like the menu of how to control the generation process. ComfyUI is a node-based GUI for Stable Diffusion.
There are so many tutorials out there. For the English speaking people, I can recommend:
https://www.youtube.com/watch?v=LNOlk8oz1nY&t=79s
For the Germans:
https://www.youtube.com/watch?v=HOIeLNNhdaU
Thats how Comfy-UI does look like when you do work with it:
With prompting “an astronaut is leaving a spaceship” I will get this generated picture:
That looks outstanding! Doesn`t it? Well, but it is generated with SDXL. SDXL 1.0, the flagship image model developed by Stability AI, stands as the pinnacle of open models for image generation. The hardware requirements are extreme high, and you should have a 4080 computer with lots of VRAM for this. AI generation is all about VRAM!
There are the first SDXl models out, which you can use with a 2080 or 3080 graphic card. In general it is not a bad idea to start just with the hardware you do have. The standard model is still SDL 1.5 and it is pretty amazing what you can do with it. Thats how I did start to work with AI for about months.
The big issue about AI generated pictures are humans. The AI is still not perfect with it at all. The preview of hands and feet is mostly broken, and you do need a lot of stuff and tools, which does help a little bit. Dont get me wrong. If you do want to generate a portrait of a human, the AI is perfect, but when you want to do this with the full body you will be in troubles!
Lets try a new prompt: “portrait of a beautiful African woman with perfect eyes in the desert at the afternoon, perfect eyes, perfect eyelashes, perfect eye-lid, dark red lipstick, masterpiece, intricately detailed, expressive, heavy shading, cinematic angle, cinematic lighting, masterpiece, best quality, 8k, UHD, shadow play, intricate details, natural skin” give me this generated picture:
The skin is not perfect and now the “work” would start with trying maybe another model or trying different prompts. The prompts are the Holy Gral of AI generated pictures. If you do have a good one – don’t share it! Each model works different with the prompts and that makes all so time consuming. One of the best pages to find free download model is civitai.com. Try to get the trained models, not the merged ones. You will even find there Lora`s (LoRA models, known as Small Stable Diffusion models, incorporate minor adjustments into conventional checkpoint models) and other helpful downloads like tutorials, embeddings and just creative ideas.
Right now there is a rush going on regarding the generation of animated gifs/videos with AI. You can find free download example I did create at my page: https://www.icarusdesignes.com/free-gifs
All the pictures and collection you can see at my overview page are generated with Comfy-UI. Art - Anime - Erotic - Humans - Landscapes… I guess there is nothing you cant do with AI!
Last week was a new model version of Stability AI was announced. SD 3.0! In the meantime there is even a new version for SDXL - Lightning! The problem about SDXL is that you do need a high power hardware! The new Lighting version of SDXL even works with a 4 to 8 step generation version! I did test it and the pictures do look awesome! It even might work with a 2080 or 3080 graphic-card. This picture I did generate with the new lightining model of juggernaut. It is a free trained model to download and good fun to use it. I did use the ip-adapter to get more depth in the picture. I hope you do like it :
Happy AI generation!