[Bug] train_dreambooth_lora_flux2_klein.py: batch size mismatch with --with_prior_preservation

[vishk23]

When using --with_prior_preservation with train_dreambooth_lora_flux2_klein.py,
the prompt embedding repeat logic doubles the batch incorrectly.

The line:
num_repeat_elements = len(prompts)

should be:
num_repeat_elements = len(prompts) // 2 if args.with_prior_preservation else len(prompts)

Because prompts already contains both instance and class samples from collate_fn,
but the repeat creates 4 embeddings for a batch of 2 latents.

Reproducible with: train_batch_size=1, with_prior_preservation=True

Diffusers version: 0.38.0.dev0 (main branch)

[chenyangzhu1]

I found that the same problem also appears in Qwen-Image's dreambooth lora script.

diffusers/examples/dreambooth/train_dreambooth_lora_qwen_image.py

Line 1468 in 8070f6e

num_repeat_elements = len(prompts)

For convenience, I didn't start a new issue, but instead submitted a pull request in #13396.

@sayakpaul

[sayakpaul]

All reviews done. Thanks a lot!