Tied Spreadeagle

When two consenting adults love and trust each other, sometimes they engage in acts that require, what we call, a "safe word". LORA's require something called a "trigger word", or at least they used to. Those don't seem to work for HunyuanVideo.... something to do with llamas? But I digress...

Tied Spreadeagle pretty consistently gives you a tied spreadeagle position. It doesn't have to be on a bed necessarily and it seems pretty flexible. I only trained front angles so try side ones at your own peril.

The showcase videos include full metadata with a healthy wildcard prompt that almost never contradicts itself ;)

Example {Wilcard} Prompt

tiedspreadeagled, {low|medium high|high} angle {full body|close up} view of a {curvy|slim|fit|slender} {18|25|30|50} year old {European|Latin|Russian|Swiss|Irish|American|African|Swedish|Asian|Austrian|German|Siberian|French|Dutch|Japanese|Ukrainian} woman with {very long|medium length} {dirty blonde|blonde|brown|black} {|straight|ponytailed|styled up-do} hair and {dark|pale|tan} skin is lying back tied spreadeagled on a {metal frame hospital bed in a bright operating room|dark wood table in a bdsm room with a wooden headboard|small white sheet bed with a wooden headboard with lots of candles around her|narrow metal table in an alien abduction and test laboratory}. 

She is {completely naked|wearing a bikini|wearing a bondage harness} with {huge breasts|tiny tits|small breasts|medium breasts} and a {hairy|hairless} ***** except for {pink|black|white} high heels.

Her arms extended towards the top corners of the bed and her legs are spread very wide. Her wrists and ankles are tied with a beige rope to the {wooden|wrought iron} top and bottom bed posts. The rope is taught keeping the girl from being able to move much. She struggles lightly against the restraints but fails.

She looks from side to side with {a red ball gag in her mouth held with white and brown straps around the back of her neck|{white|black|red|pink} tape over her mouth|a blindfold over her eyes}. She looks very {scared|happy|excited|surprised}.

{There is a Hitachi magic wand vibrator on the bed beside her along with a giant black dildo.|There are beer bottles and used condoms beside her} She has shiny oiled skin making it glisten with wetness.

The camera is {handheld|held steady}.
smooth natural motion, oiled skin

Versions:

One

This was a pleasant surprise since it was my first attempt at this LORA. It understands the position and [lack of] movement well. I think I trained it on the wrong aspect ratio though because it REALLY likes close-ups and doesn't seem to want to show feet.

Two

This one hits the mark, most of the time. No more "All Close Ups All The Time!". I threw this one back in the oven for a couple hours to see if it would stabilize, but more steps/epochs didn't improve it... Alas, Two really isn't that bad, with the right prompting I find it worthy of it's place here.

Three

Trained on a revised set of higher resolution (576x324 and 384x384) videos for 50 frames. LR set at 6e-5, run for 65 epochs, 3900 steps. This one does a better job with rope/chains/straps. Is it perfect? No, and I think in general BDSM concepts are hard for AI because of the fine details in the restraints. Alas, this one seems flexible and all around very capable. It can do low/high angles, wide angles, close ups, aerial views and a variety of bindings.

I think that does it for me for this concept as I'm not sure where to go from here. If you notice any serious issues or have suggestions, feel free to post a comment below!

Four?

I had a long conversation with ChatGPT about HunyuanVideo and my understandinf of captioning needed work, deep thought and some experimentation. Given that this dataset was the most mature, I've re-captioned things again and run a training... run. The loss curve was compelling! I just haven't had time to test it, I'm expecting either a breakthrough that will allow us to greatly reduce prompt size, or more head scratching. The suspense!!!

I also have a new concept trained, also awaiting testing. I'm optimistic after so many captioning revelations, feeling like I finally understand what, and what not to, caption.