There is a risk of the community getting "entrenched" in an old model/architecture. In stable diffusion land, for instance, the community is pretty much stuck

There is a risk of the community getting "entrenched" in an old model/architecture. In stable diffusion land, for instance, the community is pretty much stuck on the "old" architecture. Newer innovations, like Huggingface diffusers and various optimizations derived from that like PEFT, torch.compile support, AITemplate/TensorRT compilation and various other bits are largely unused. They are also pretty much stuck on SD 1.5, even though 2.1 is a good base for finetuning. This has happened in the past too, with ESRGAN.