This is why I use the word ‘proliferation,’ in the nuclear sense. Since the days of SD1, these illegal capabilities have become more and more prevalent in the local image model space. The advent of model merging, mixing, and retraining/finetunes, have caused a significant increase in the proportion of model releases that have been contaminated.
What you’re saying is ultimately true, but it was more true in the early days. Animated, drawn, and CGI content has always been a problem, but photorealistic capability was very limited and rare, often coming from homebrewed proprietary finetunes published on shady forums. Since then, they’ve become much more prolific. It’s estimated that roughly between a fourth and a third of photorealistic SDXL-based models released on civit.ai during 2025 have some degree of capability.
Just as LLM benchmark test answers have contaminated open source models, illegal capabilities gained from illegal datasets have also contaminated image models; to the point where there are plenty of well-intentioned authors unknowingly contributing to the problem. There are some who go out of their way to poison models (usually with false association training on specific keywords) but few bother, or even known, to do so.
ImgurRefugee114@reddthat.com 19 hours ago
This is why I use the word ‘proliferation,’ in the nuclear sense. Since the days of SD1, these illegal capabilities have become more and more prevalent in the local image model space. The advent of model merging, mixing, and retraining/finetunes, have caused a significant increase in the proportion of model releases that have been contaminated.
What you’re saying is ultimately true, but it was more true in the early days. Animated, drawn, and CGI content has always been a problem, but photorealistic capability was very limited and rare, often coming from homebrewed proprietary finetunes published on shady forums. Since then, they’ve become much more prolific. It’s estimated that roughly between a fourth and a third of photorealistic SDXL-based models released on civit.ai during 2025 have some degree of capability.
Just as LLM benchmark test answers have contaminated open source models, illegal capabilities gained from illegal datasets have also contaminated image models; to the point where there are plenty of well-intentioned authors unknowingly contributing to the problem. There are some who go out of their way to poison models (usually with false association training on specific keywords) but few bother, or even known, to do so.