Esoteric Software

[Misaki]

The face expression AI is fun As you mentioned, some facial expressions seem to be difficult to recognize, like angry and fearful. However, just being able to recognize happiness, surprise and sadness seems to give me a lot more freedom!
I would definitely like to add facial expression animations to my model when it is possible to change the model's facial expression with the AI. (My understanding is that it is not possible to do that yet, but is it possible already?)

[SilverStraw]

Have you tried to lower the confident threshold for the harder facial expressions and raise the confident threshold for the easier facial expression?

This picture might help people understand what the AI is looking for when it tries to recognize facial expressions.

I would definitely like to add facial expression animations to my model when it is possible to change the model's facial expression with the AI. (My understanding is that it is not possible to do that yet, but is it possible already?)

Yes it should be possible.

[Misaki]

Have you tried to lower the confident threshold for the harder facial expressions and raise the confident threshold for the easier facial expression?

Yes, but even when set threshold to 1, "angry" and "fearful" were not recognized well. As long as I've tried, "disgusted" is the most difficult to be recognized.

Yes it should be possible.

Oh, so could you tell me what animation name is required for each expression?

[SilverStraw]

Yes, but even when set threshold to 1, "angry" and "fearful" were not recognized well. As long as I've tried, "disgusted" is the most difficult to be recognized.

I might not have explained it well. The values should be between 0 and 1 where 1 makes it harder to trigger and 0 makes it easier to trigger. If you set those expressions to 1, then it makes it harder for the expressions to be recognized.

Oh, so could you tell me what animation name is required for each expression?

I am sorry for the misunderstanding. The current prototype is not ready yet for adding facial expression animations.

[Misaki]

I might not have explained it well. The values should be between 0 and 1 where 1 makes it harder to trigger and 0 makes it easier to trigger. If you set those expressions to 1, then it makes it harder for the expressions to be recognized.

Ah, I see. Indeed, when I tried to set the threshold for "angry" to 0, my facial expression was often recognized as "angry" even if it is almost expressionless.

I am sorry for the misunderstanding. The current prototype is not ready yet for adding facial expression animations.

I got it, so I'll try to add the animations when it is ready

[skarasuko]

Is there a guide on how to export correctly? I tried out a very big Spine file, but some images are completely black and shaped like their mesh. It would also be nice if this supports multiple skins as well.

Edit 1: I found out that this does not accept multiple pngs.

[SilverStraw]

@skarasuko Sorry you are having issues.

[skarasuko]

@skarasuko Sorry you are having issues.

No worries at all. This is already feels very usable among most aspiring V-Tubers! The Spine Team should consider fast-tracking official support for this, because it's still the perfect time to be part of the market!

[SilverStraw]

Spine Vtuber Prototype

Update 1.1.3

* Separate error catching on Spine animation tracks.
* Allow "left eye open" Spine animation track to keyframe both eyes while detecting eye wink.
* Add wink threshold to determine the ease of detecting eye winks. Range from 0 ( easiest ) to 1 ( hardest ). Setting located in Model Settings > Single Value Properties > wink threshold. This property can be saved into and loaded from svp file.
* Fixed "brow raise" incorrect landmark tracking.

https://silverstraw.itch.io/spine-vtuber-prototype

Updated Spine Vtube Test Model for synchronizing both eyes on "left eye open" track and keyframe only the right eye on "right eye open" track.

https://silverstraw.itch.io/spine-vtube-test-model

Some ninja patching with the help of LustFire's bug catching.

---

A user notified me an issue with the experimental motion capture not exporting the JSON file. I patched the bug.

[Misaki]

I have tried 1.1.3, and it certainly seems to make it easier to detect eye wink! Here is the result of my test (the model was also modified slightly):

It had been a while since I tested this tool, but I found it enjoyable once again
Thank you for keeping this cool tool updated!

One thing I thought was it might be a good idea to add a threshold value for the eye pupils, such that subtle changes in position could be ignored. My model's pupils wobbled while winking, so I adjusted the left/right pupil pitch strength and left/right pupil yaw strength parameters, but if these values are set too low, the model’s eyes will not follow the eye movement at all and it is not ideal. FYI, in the model in the video above, left/right pupil pitch strength is set to 4 and left/right pupil yaw strength to 5. I would be happy if you could consider it!

Here is my latest model data:

chara-for-Spine-Vtuber-Prototype_20221223.zip

[SilverStraw]

1.1.4

• Update to Spine WebGL 4.1.24 .

• Add skeleton debug mode into scene render. The checkbox is located under "Canvas Settings" menu.

• Add debug bones checkbox under "Canvas Settings" menu. This include options for bone center color and bone line color, and bone line width ( minimum of 0 and max of 10 ).

• Add debug region attachments checkbox under "Canvas Settings" menu. This include option for region attachment line color.

• Add debug mesh triangle checkbox under "Canvas Settings" menu. This include the option for mesh triangle line color and mesh line opacity ( minimum of 0 and max 100 ).

• Add debug clipping checkbox under "Canvas Settings" menu. This include the option for clipping line color.

• Add left/right pupil pitch/yaw threshold settings to "Single Value Properties" drop-down list under "Model Settings" menu.

• Fix bug that did not load pre-multiplied alpha when it was assigned false.

https://silverstraw.itch.io/spine-vtuber-prototype

One thing I thought was it might be a good idea to add a threshold value for the eye pupils, such that subtle changes in position could be ignored. My model's pupils wobbled while winking, so I adjusted the left/right pupil pitch strength and left/right pupil yaw strength parameters, but if these values are set too low, the model’s eyes will not follow the eye movement at all and it is not ideal. FYI, in the model in the video above, left/right pupil pitch strength is set to 4 and left/right pupil yaw strength to 5. I would be happy if you could consider it!

I had to reinstate the threshold mechanism after I forgo it for moving average. You cannot export those settings to .svp yet. I want you to test it out before I make more commitments to the pupil threshold.

[Misaki]

Thank you so much for adding left/right pupil pitch/yaw threshold settings!! I have tried them and changing the threshold values actually helped to reduce the eye wobbling. Here is the result:


The settings for my model related the eyes are as follows:

• left/right eye strength: 50

• left/right pupil pitch strength: 10

• left/right pupil pitch threshold: 0.15

• left/right pupil yaw strength: 4

• left/right pupil yaw threshold: 0.07

While testing these settings, I found I should modify my animations, so the model in the video above is updated. Here is the updated project files:

chara-for-Spine-Vtuber-Prototype_20230117.zip

As shown in the following image, I made the pose will not be changed immediately after detecting changes in eyelids and mouth movements captured in some animations:

eye-open-animation.png

This adjustment was made so that minute movement changes would not cause the eyelids or mouth subtle opening. Also, since I set the thresholds for pupil pitch higher values than for pupil yaw, I adjusted pupils in the pitch down/up animations do not move from frame 0 to frame 10, so that these do not appear as if the position suddenly jumps when the threshold is exceeded.

By the way, you said:

You cannot export those settings to .svp yet.

but somehow I can export the threshold settings to .svp. (Maybe you have updated this tool after replying to this thread?)

Anyway, I am happy with the results this time! There are some things I would like to fix in my rig (e.g., The half-eye pose is not very good, although I have adjusted it many times), but I think the current specification of this tool is already great for vtubing. I am looking forward to the day when facial expression animations can be added. Great work!!

[SilverStraw]

but somehow I can export the threshold settings to .svp. (Maybe you have updated this tool after replying to this thread?)

It has been a while since I worked on the source code. I forgot that I have a function that exports settings from a list of default setting values. I updated that list for the left/right pupil pitch/yaw threshold default values, therefore, they got exported.

---

Update 1.1.5

• Add "Backface culling" checkbox to "Canvas Settings" menu. The backface of attachments are invisible if checked. The back side is kept consistent with Spine editor.

• Add "Flip Horizontal" checkbox to "Canvas Settings" menu . When checked, the world X-axis is flipped.

• Add empty animation tracks: custom tracks 1 to 4.

• Add customizable expression buttons 0 to 20 underneath the rendering canvas. Button 0 resets all custom tracks thus removing all customizable expressions active. The buttons are usable once the expression slot has been setup.

• Add customizable expression setup interface under "Model Settings" menu.

1. Add customizable expression slot drop-down menu ( 1 to 20 ) in the setup. Each slot allow you to setup multiple custom track index ( 1 to 4 ), transitional animation, and animation loop that follow after the transitional animation.
   Include a button to add a setup interface row for adding more custom tracks ( 1 to 4 ), transitional animations, and animation loops.
2. Include a button to add a setup interface row for adding more custom tracks ( 1 to 4 ), transitional animations, and animation loops.
3. Add a button to remove the last setup interface row. You do not want any rows with incomplete information as you will not be able to finish setting up the customizable expression slot.
4. Add a button to assign all the custom tracks ( 1 to 4 ), transitional animations, and animation loops to the customizable expression numbered slot ( 1 to 20 ).
5. Each setup interface row has three parts: custom track index ( 1 to 4 ), transitional animation, animation loop. Custom track index ( 1 to 4 ) and transition animation are required for setup while animation loop is optional. The transition animation and animation loop input fields are drop-down menus that lists all the animation found in file ( .json | .skel ). Transition animation does not loop and the animation loop plays after the transitional animation. There is not any mix duration between transition animation and animation loop.
6. The customizable expression slots are saved into and can be loaded from SVP file.

• Remove animation selection from Single Value Properties drop-down list with "Model Settings"

https://silverstraw.itch.io/spine-vtuber-prototype

Updated model for testing custom expressions.

https://silverstraw.itch.io/spine-vtube-test-model

[Misaki]

The expressions feature is really fun and wonderful!! I haven't been able to test it much because preparing the animations still takes more time, but here are the results of a quick test I did:

I know that the transition animations back to the default pause also need to be registered in the expression buttons, but I have not yet been able to do that at this time.

My current model is in a very half-assed state, but I'll leave the data here for anyone who wants to test it:

chara-for-Spine-Vtuber-Prototype_20230130.zip

When I make more improvements, I will share the data here again.

[SilverStraw]

I should add another setting for 'returning to default pose' animation to the customizable expression feature. That way you save another expression slot.

After this feature, facial expression recognition AI should be as easy as setting which expression slot you want to use for the AI result.