2x slower/faster training finetuning sd15 because of bug in meta data

Tested 10x and it persists, easy workaround but slows training process overall. Many users will likely be experiencing 2x slowdown.

Fine tune sd15 checkpoint. Set generate caption metadata, and generate image buckets metadata.

This completes and training starts

Training is 2x slower measured on identical training sets.

Stop training. Uncheck these two options.

Check add resume from state or start training a new,

These aren't computed this time. Not counting this time, Now it/s is 2x faster. time between epochs is 2x faster. number of steps per second 2x faster.

To workaround, stop training immediately after they have been created or precompute, then resume training/start training.

No file, test, or entry point is named. Reproduce SD15 fine-tuning with both “generate caption metadata” and “generate image buckets metadata” enabled, then compare it with resume training after those metadata files already exist. Done means metadata generation no longer causes the reported 2x slowdown, with the existing workaround behavior verified.

Written by the indexing model from the issue text.