To be fair, this has been my initial impression with basically all image generation models. The current generation is finally at the point where throwing untuned prompts at untuned models gives good results, but those never match the results of finely tuned (positive and negative) prompts with parameters adjusted from experience; ideally with model fine-tunes added.

If you want the best results there is still skill and work involved. Consequently a showcase by people experienced with the model far surpasses what you get when you shout prompts over the internet for somebody else to try

The real value of Stable Diffusion models are the finetuned models when the base model is released.

Non-open licensing may adversely impact that.

I didn't think those were bad at all

Not bad, but also not the big jump over fine-tuned SD1.5/SDXL checkpoints that some expected.

Fine tuned 1.5 checkpoints are amazing. If only the models could comprehend instructions like SDXL or better.

That woman at the top (at least I assume it's supposed to be a woman) should be wearing a beauty pageant sash reading "Ms Uncanny Valley of 2024". Creepy AF, IMO.

Agreed that some of the others aren't so bad, however the bar for a convincing swamp creature is a lot lower than for a convincing human being.

If breasts come out looking like that, I'd be taking a hard look at my training data.

The whole picture reeks of bad taste. The hands are particularly terrible. Total fail! But then again, maybe it's due to a poor prompt.

Those are shockingly bad.

I am sure someone will tell me there is a reason why I am wrong and these aren't that bad.

Midjourney has never needed an explanation though with words. The proof is in the output. Everything else is nonsense.