Stable Diffusion is the only one you can host yourself or even run locally on desktop & mobile if your hardware allows. You can train your own styles and concep

Stable Diffusion is the only one you can host yourself or even run locally on desktop & mobile if your hardware allows. You can train your own styles and concepts using Dreambooth method (aka fine-tuning on a very small dataset) for 1.x models, or use textual inversion / LoRA / embeddings for 2.x models. SD2 is much more difficult to fine-tune so folks have been using these alternative methods. Embeddings can represent a specific style and are easier to compose. They are storage-efficient as well (<= 1 MB if I remember correctly). DALL-E is good for surreal memes or some kinds of art. It works well for outpainting. For other use-cases it falls behind on realism and flexibility. Midjourney has come a long way and generates very aesthetically pleasing images. But it's very opinionated to allow for ease of use for new users. And it can get expensive if you need to generate many images.