dobkeratops
commented
8 months ago https://github.com/haotian-liu/LLaVA
opensource multimodal LLMs .. text+vision .. imagine being able to finetune these on community curated datasets another potential usecase for imagemonkey
observations about the current state of AI..
so over the past year there's been these diffusion models & large language models become available to be public
There's the controversy over training data (scrapes)
e.g. Steam doesn't allow games with AI generated content, or even use of ChatGPT ; they're covering themselves
That seems to me to validate the need for a efforts such as imagemonkey to produce more CC0 open training data?
Individuals such as ourselves can't currently produce foundation models.
but models like stable-diffusion can be fine-tuned (stability.ai's argument about the scrapes is that a fine-tuned model can remove its ability to recreate copyrighted material in the original scrapes, which sometimes come through)
I see that you had VGC finetuning integrated in imagemonkey.. have you considered extending your service to finetune diffusion models?
Also consider that in time multi-modal models may appear.. ones trained simultaneously on images & text