There are public static diffusion models trained on/like midjourney, but IMO they're still lacking.

What's behind MidJourney, and how has it not been matched/defeated? There's a lot of interest in the space and money to be made, and MidJourney proves that quality results are possible. Is it because of the sheer costs or a secret sauce?

They combine multiple custom algorithms that are designed to run on high-RAM (>24GB) GPUs. And, they incorporate a lot of user-provided feedback into the training of new models. If you have more time on your hands than cash, you can get some free generation credit by rating images.

That's mostly because they make a finished end-user product, which nobody else does (except OpenAI and DALL-E, but they are mostly focused on language processing right now).

- It runs in the cloud with more VRAM available than on consumer machines, so it's able to use more parameters and custom layers. This makes it more coherent.

- It's finetuned on a large RLHF dataset, which they have from the users voting for bot-generated images in their massively popular Discord server.

- It includes a decent share of custom processing and manual tuning (system prompts, custom layers trained on god knows what etc)

Their closest contender, Stability AI, is currently busy training SDXL, their new model, on a RLHF dataset they are collecting on their much smaller Discord channel. So Midjourney is SOTA for text-to-image.

There's DeepFloyd IF which is large enough and includes a 11B LLM for natural language understanding, so it has much less issues understanding complex prompts such as "koi fish doing a handstand on a longboard rolling down the slope of a moon crater". But it's not a finished product (no complex pipelines, no RLHF), and the license is restrictive.

That said, Midjourney's strengths are also its weaknesses. Midjourney is good for the text-to-image, but not much else, and image generation is much more than just text-to-image. MJ gives great results off the bat, but struggles with unusual requests, complex compositions with multiple objects, and can't use various elaborate trickery available in the much smaller Stable Diffusion (which is mediocre by default but has powerful higher-order guidance and training, and is also much better at complex things). Midjourney is also censored - you can't have Joe Biden eating a slice of porcupine.

Well you can gain a free GPU hour by rating images. I am pretty sure midjourney is doing some kind of reinforcement learning from human feedback.

The 5.1 raw model is simply unbelievable for photo realistic images. There is 4a,4b,4c, 5, 5.1, 5.1 raw, the anime model I never use. All are interesting but how often do you read anyone mention what midjourney model they are talking about? Basically never.

There are certain circumstances that I have seen stable diffusion match the output but overall I don't even think it is close.

Personally, it wouldn't even occur to me to look for stock photos now as opposed to just making variations in midjourney. There is a bit of a learning curve with so many options in midjourney so I take anything I read about it with a grain of salt. Most people are using these tools to make absolute shit.

Stable Diffusion is still good but you need to fine tune it (not that expensive with Lora) and use controlnet to get good results. The model is just too small on its own.

Civitai is a good place to get these models if you don't want to train your own. NSFW partially

Kandinsky 2.1 is worth a shot if you haven’t tried it.

what about stable diffusion or the tools by adobe ?

The new Bing runs on a better version of dall-e and is much better