LocalOptima I've never heard of that logic before, though your results seem to show it works.

LocalOptima - "It's because repeats prevent buckets from getting dropped when the bucket size is less than the batch size. When batch size = 1, all buckets are large enough for a batch."

This makes no sense to me. Can you explain it or point to a resource that does?

LocalOptima - I'm sorry, but I was able to read and comprehend that entire thread, and it says nothing at all remotely like what you have said here. What are you trying to say? What does it mean for a bucket to be "dropped"? How does the number of repeats or batch size have anything to do with bucketing?

LocalOptima - Thanks for the info and the link. I was absolutely remiss... my bad. I read "everything" and saw nothing relevant to your comment... and it was there in the last sentence which I skimmed over in my hasty frustration.

Sorry.

I have learned something valuable that I did not know, even after successfully training hundreds of LoRAs.

-------------------

Why would data get “dropped”?

It’s not that your images are literally deleted — they just don’t get used during some iterations/epochs because of how aspect ratio bucketing and batching interact.

Most bucketing systems (including diffusion-pipe and Musubi) work like this:

They group your images into buckets based on their aspect ratios (e.g., 1:1, 3:4, 9:16).

Each bucket is used to build batches so all images in a batch have the same target resolution (reducing distortions).

If a bucket doesn't have enough images to fill a full batch (e.g., you have a batch size of 2, but only 1 image fits that bucket), some images might sit unused in that pass.

With very small datasets (like 50–100 images), this can mean certain aspect ratios almost never get trained because they don’t neatly fill batches.

Where do repeats come in?

num_repeats multiplies the dataset artificially.
For example:

75 images, num_repeats=10 → treated as 750 images per epoch.

Each image is just duplicated logically (not on disk) to make the buckets "fatter."

The idea is:

By having more "copies" of each image, the bucketing code has enough items to fill batches consistently.

Without repeats, small buckets (like rare aspect ratios) might get skipped entirely during a given epoch.

Why does global batch size matter?

If your global batch size (per-GPU batch size × number of GPUs) is big compared to your dataset:

Buckets with only a handful of images may not have enough to fill one batch.

Those images sit out that epoch because they can’t form a complete batch.

Repeats "inflate" the dataset so those odd buckets can participate.

--------------------

This blows me away. Love learning!!!

LocalOptima This is most useful with a diverse dataset yeah? Can I ask how many poses/angles you have in the 23 videos?

yes please it was deleted

I would like that one too. ^_^

@LocalOptima can I get it too please??

@LocalOptima please ❤

How are you getting it to work with 2.2? It throws a "lora not loaded"

I don't really get desirable behavior with wan 2.2. The lora clearly does something, but it doesn't swing naturally like this version

Let him train.