Attention: please restart VAM after updating BodyNet. Otherwise, VAM will continue to use the previously loaded dll.

changelog v1.32
  • "Thigh Apart" morph removed - no more bow legs
  • re-trained and added a bunch of networks
Pro version:
  • Added UserDefaults to the person plugin. The load priority is: settings stored with the scene > creator defaults > user defaults > factory defaults
  • Added the option to run BodyNet automatically after loading an appearance or morph preset. It's off by default and will cause issues if you load an appearance preset, which has a different scale than the previous one (just like when you execute BodyNet manually).
Attention: please restart VAM after updating BodyNet. Otherwise, VAM will continue to use the previously loaded dll.

changelog v1.31
Improvements
  • Major rework of how the plugin determines the body to head ratio. It's using collider positions with pose compensation instead of skin vertices now, making the algorithm work in any pose.
Bugfixes
  • Fixed a typo preventing loading the bundled default network.

Pro version only
Features
  • Added a preset system to save creator specific default settings for the person plugin. All settings are included but the most useful will probably be the used network and the number of refinement iterations. When you load a scene from a var, BodyNet will extract the creator's name from the var's file name and look for a matching preset.
some bugfixes but mostly internal reworks
I plan on releasing a paid version that comes with a person plugin as well as a session plugin that automatically adds the person plugin like you know it from Cheesy's PluginSuit, for example.

You can hook the person plugin up in UIAssist or a button in your scene.
There will be a creator version that you can package and ship with your vars, too. With that version, you could even ship your own custom trained networks along with the plugin. Your patreons could load and adjust their appearances with as little as two clicks.

Here's a quick demo of it in action:

changelog v27
Features
  • Added a VR mode to the UI. You can enable it in the scene plugin's settings.
Fixes and improvements
  • Fixed minor bugs when training on multiple skeletons at once.
  • Further memory optimizations
  • Increased the drawn skeleton's refresh rate from 1Hz back up to 23Hz.
vr.jpg
changelog v26
Features
  • Updated/added the pre-trained networks for the following source appearances, more to come.
    Each was trained on a set of 42 popular and diverse appearance presets from various creators.
    • VaM Default
    • CuddleMocap.Alana
    • Androinz.Ellie
    • AshenRyder.Chloe
    • ZenMocap
  • Added the ability to train on sets of target skeletons.
    By increasing the training set size we can train networks that are more versatile and less biased.
    This allows me to package pre-trained networks that are far more capable of generating good results without refining.
Improvements
  • Optimized memory usage and multi threading performance.
  • Implemented bone orientation morphs.
sneak peak v26

We'll be able to train on multiple target persons simultaneously. All computations are made in parallel, which means you can utilize more of your CPU when you train with lots of models.
The training goal here is to reduce the average loss. This also means I can create and package networks that are way better at matching any appearance to a specific source look.

Here's a shot of 21 skeletons processed at the same time:
before.png

after.png
changelog v25
Improvements
  • Major multi threading improvement. There will no longer be a huge hit on fps.
Bugfixes
  • Fixed shoulder width morph not taking effect.
  • Fixed an error in body-to-head ratio measurement.
Hey,
just a quick info for those who'd like to support my work.
There's no Patreon but you could

:coffee: buymeacoffee.com/tbdvam :coffee:

if you'd like.
You can't imagine how much of that stuff I drink while coding. I'm somewhat of an espresso snob.

Cheers
v24 changelog
General improvements
  • Implemented bone and morph simulation to speed things up.
    There's no longer a need to set a specific physics update rate and training is multiple magnitudes quicker.
  • Added the first two specifically trained models. CuddleMocap's Alana and Ashen Ryder's Chloe.
    Training is quick enough that you probably don't even need those but if you're trying to match one of them you can load the data.

Yep, that's real time.

Back
Top Bottom