Discussion
OMFG I just solved yesterday's Live2D physics mystery.
The trick was to stop trying to make sense of the numbers geometrically, and just try writing buggy code.
Wrote the dumbest bug I could think of, matched Live2D behavior 100% on the first try.
Spot the bug...
@lina this is amazing! In hindsight, was it fun at all or just frustrating?
The issue reminds me of all the weird behaviors emulator developers need to support for games that rely on old bugs. Same with Windows, actually.
@MostlyBlindGamer The physics part is more frustrating than the rest tbh. The rest was more like uncovering mysteries, but for this I'm having to dumb myself down...
@lina haha. What an interesting experience.
@lina
Oh, overwriting p.pos[0] before doing the second calculation?
@lina This is _such_ a dumb bug omg.
@lesto 0=x, 1=y. The 001/110 part itself isn't the problem.
@lina@vt.social what should the code look like?
@m Ideally it should just use a vector math library so you don't get these dumb mistakes reimplementing rotations from scratch... (hint hint)
@lina
I was so confused, looking at formulas and stuff before realizing what it's being assigned to 😭
This is now the third or fourth time I waste hours trying to "match the numbers", sketching out geometry, and trying random "sensible" formulas to match or approximate Live2D physics, and then the answer ends up being "just try doing the dumbest broken stuff you can think of".
Now, having finally understood the behavior, I know it's nonsensical, and not even invariant with the original pendulum angle.
Which means it will not be implemented in Ayagami. I hope nobody depends on this bug, but if they do, they can fix their model. Sorry.
The goal I have is to implement physics that "feel" the same, in the sense of fluidity/drag/etc., as Live2D physics. To that extent I'm willing to implement fudge factors to correct for imprecision in the original.
But not stuff like this, this is just a bug.
@lina dang. kind of annoying that if a rigger has tweaked values to make it look ok in Live2D, fixing this might make it look worse (although hopefully a very minor difference on most models?).
Assuming Live2D don't already know about this bug I wonder if they would fix it or leave it be when they do find out.
@kepstin My gamble is that the bugs are mostly things riggers work around (like the sticking to zero bug, it's notorious and everyone works around it).
This particular one would essentially affect the dynamics of fast large angle movements, though at worst by 50% or something. So it's not that bad, you'd basically need to have an instant toggle (or a wildly over scaled parameter) making a large angle physics change to notice. And it only affects angle mode, not X mode (I think most people use X mode more). Essentially, physics will react a bit more slowly than the buggy version, for extreme inputs.
At some point if people are doing "crazy" things with physics it would be impossible for them to behave identically with a non identical physics engine. I'm hoping most people don't do that.
Install bonfire.cafe
Get the full app experience
![# Compute rotation of the pendulum when incoming parameter
# is set to "angle". Besides changing the direction of
# gravity, Live2D has been observed to apply 1/5th of
# the delta angle "immediately", probably as some sort of
# fudge hack to make the pendulum start moving faster than
# it otherwise would towards the new gravity vector, making
# it more "responsive" or something. This 1/5th factor is
# accurate for small angles, but Live2D actually overshoots
# it for large angles. Why? Well...
input_angle = pi / 2 # Instantaneous +90 degrees rotation of pendulum system
da = input_angle / 5 # Apply 1/5th of the angle on first frame
# Rotate the bob angle by da
p.pos[0] = p.pos[0] * cos(da) - p.pos[1] * sin(da)
p.pos[1] = p.pos[1] * cos(da) + p.pos[0] * sin(da)
p.show()
# Live2D value: a=0.346603990 (radians)
# Above prints: a=0.346603960
# Correct answer: a=0.3141592 (pi/10 radians, 18 degrees)
# Spot the bug.](https://media.vt.social/media_attachments/files/116/956/291/577/708/341/original/96a06ac21de9983a.png)
