excellent work
hello conflictx, i think you have done a great job with your TI/embeddings, have you worked also on anything based on v1.5 ? 2.0 is not so popular and few people use the models, but I am going to try to use yours for some! Would be lovely to see what you are up to nowadays as many months have passed and a lot of progress has been done with AI gfx and safetensors models
Hey, sorry for the late reply and thanks. I moved back to 1.5 completely to be fair, as Stability.AI dropped support and stopped releasing for SD 2.0 I didnt see the point in sticking around for it. Extensions only supporting 1.5 didn't help either, and let's be honest ControlNet is a major one. I did continue working on 1.5 models, but mostly for private/professional use and I have a 1024x resolution model that surpasses anything base 2.0+ could do.
I also started using local LLM's and Voice AI so my GPU time has been split over multiple things now. For the moment I have mostly been having fun with a Discord Bot that uses my local LLM that can be used as an assistant and also has access to Stable Diffusion for both creating the image and stories for that image.
o_O Oh wow! that's awesome, so that's a private engine on your computer? looks amazing! It's like you have your own "midjourney" kind of thing going on :D!
I am very curious and interested in models trained at a higher resolution than 512px because that is a very huge limit. I have heard some people were talking about making a custom 1.6 model of SD but that's just rumors for now, with higher resolution images used as source and a lot of data about perfect hands so to have a superior model to the 1.5, but we'll see ;)
It looks like you have done something similar though and would love to know more!
Exactly, everything runs locally and goes trough an API to connect to Discord. It needs to a bit more Discord features to compare to Midjourney I guess, but output wise it comes close. The images on discord run trough a High-res as well before being sent so the image output is at 1535x2048.
As for 1.5 models trained on higher resolutions, I think I might have been one of the first to try back in the day with https://huggingface.co/Conflictx/Complex-Lineart, which used 768x768 but on only 100 images. For a more generic model you need a bigger dataset, and preferably all good images with decent captions.
do you have any plans of releasing your model out there? or perhaps is it a work-in-project?
it surely sounds like an amazing idea yours !!