site banner

Friday Fun Thread for September 8, 2023

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

3
Jump in the discussion.

No email address required.

Yeah. As someone who has been using stable diffusion/ai art generation nearly every day for quite a while, your box metaphor is pretty good. I feel like I keep hitting the walls of what's possible with it, and augmenting the capabilities of the AI take a lot of time and tweaking and technical skill that I'm lacking. It's also like it's so easy to get 90% of what I want from the AI but it can take ages and ages to get 100% what I want and even then it involves some old fashioned photoshop tweaking post generation most of the time. I ran into the same issue with chatgpt stuff too but my skills are more in visual arts than in language arts so I hit the walls faster with chatbots as I'm less able to coax out useful stuff with chat than I can with image generation.

Have you messed with controlnet much? Getting posing right before generation has made a huge difference for me.

I've messed around a bit with controlnet, but it's usually not people/poses I have issues with but the style of imagery. Like most models are trained on tons of anime and deviantart/pixiv style artwork and I am always going for a really specific style (like, say, Fujifilm documentary style photography from a specific year) so I'm always adding tons of tags in the negative to try to get away from the irrelevant styles in the training set that it wants to keep spitting out. Training my own LORAs with imagery I pick out has given me better results w this but I've just started doing that the other day so I'm still figuring it out