3 ms·
The model that's used by Stable Diffusion is only around ~4 GB in size.
by zirgs 4y ago
The model that's used by Stable Diffusion is only around ~4 GB in size.
- KeplerBoy 4y agowhich would make traditional distributed training prohibitively slow. You would have to sync the weights with residential internet speeds (10s of MB/s) instead of PCI-e Speeds (10s of GB/s).
- zirgs 4y agoIf you want to add new stuff to Stable Diffusion - you don't have to retrain the main model from scratch. You can train it only on a few hundred images at a time and then add the resulting model as an extension or merge it into the main model. People train their models on a single celebrity or a single artist that way. Other types of AI could be trained in a similar way.
- zamadatix 4y agoIs this amenable to merging from thousands in one go or do you have to train -> merge -> train -> merge to not overwrite each other's trainings?