With very large open models like SD3 medium and Flux.1 gaining popularity It's becoming comon to provide the diffusion model (unet/diffusion transformer) part of the model and the text encoders separately, since the text encoders can often be reused across different models, to save internet bandwidth and storage space.
I think it would be cool to support these split models here. It could also be a way to use different quantization for different parts of the model.
VAEs can already be provided separately, and this is the same kind of thing.
SkutteOleg
commented
1 month ago
With very large open models like SD3 medium and Flux.1 gaining popularity It's becoming comon to provide the diffusion model (unet/diffusion transformer) part of the model and the text encoders separately, since the text encoders can often be reused across different models, to save internet bandwidth and storage space.
I think it would be cool to support these split models here. It could also be a way to use different quantization for different parts of the model. VAEs can already be provided separately, and this is the same kind of thing.