After that one guy asked for it, the process was unstoppable.
I present to you
One Bar Prison
Versions
One
~15 videos for 50 frames at 25FPS. LR 4e-5, repeats 12. This one is fortified with close ups as I've been doing lately to reinforce learning of the right details.
As always, you need to prompt properly with my LORA because I train them to be flexible. The spotlight videos have a ComfyUI workflow and wildcard prompt embedded in them. Simply download and drag them over the ComfyUI window.
---
Training Notes
When training a concept that requires both a pose and more subtle detail like restraints, crop to the restraints and caption those accordingly:
trigger
close-up of a midsection,
part of a restrained poseMake these represent up to ~20% of your dataset, making sure the wider views are more prominent and numerous, and you'll be able to train on lower resolution full body shots, naturally de-emphasize faces (which presumably will not appear in close-ups) and get the subtle restraint details respected.
For the wide body shots, use rich natural language in full sentences, minimalism doesn't work as well imo. Capture everything you want to be flexible at least once across your dataset. When captioning things like hair color or restraint color, only mention the color once or twice in your set to avoid it from becoming a token pattern.
If you do not caption something that appears in most/all images, it can become default and thus be hard to prompt away from.
Final though:
Caption and prompting are two critical pieces and must be considered as both equally important to the final results. A LORA that forces it's concept will not be flexible. A flexible LORA will not "force" a concept, so you'll have to prompt for what you want, and that's OK!