Thanks 
I did an analysis, I hope it’s okay .. I took our images:
Analyse-dialog with ChatGPT
Me
”Take a look at these two pictures and tell me which one looks more physically realistic in terms of the broken glass and the distribution of the wine. Divide into: Picture 1 - green background Picture 2 - black background”
GPT
“
Image 1 – green background
Physical evaluation:
Wine distribution:
Very realistic. The liquid forms thin, connected filaments (ligaments) and drops of varying sizes, as would be expected from a glass breaking quickly + the inertia of the liquid.
Splash direction:
The movement plausibly follows the direction of momentum of the tilting glass: wine spills outwards, not symmetrically, but slightly preferentially in one direction.
Glass breakage:
The glass shards are irregular, vary in size and follow a ballistic dispersion. No exaggerated ‘explosion effect’.
Interaction between glass and wine:
Particularly good: the wine ‘flows around’ the broken edges, with a few drops sticking to the shards → surface tension correctly represented.
Conclusion for image 1:
Very high physical plausibility for high-speed photography (similar to schlieren/stroboscopic photography).
Image 2 – black background
Physical evaluation:
Wine distribution:
Highly aestheticised. The liquid forms a symmetrical, almost sculptural crown that looks more like staged splash art than a real break.
Splash pattern:
Too uniform, too ‘clean’. In reality, more chaotic fragmentation would be visible.
Glass breakage:
Some of the shards appear too large and too evenly distributed, almost as if deliberately choreographed.
Interaction between glass and wine:
Fewer microdroplets, less viscous distortion → appears like ‘frozen aesthetics’ rather than genuine dynamics.
Conclusion Image 2:
Physically possible, but highly stylised – more visual drama than real fracture physics.
Overall assessment
More physically realistic: Image 1 – green background
More aesthetically dramatic: Image 2 – black background
”
Me
“Is Image 1 wild?”
GPT
”Why image 1 is not wild
Chaos yes, but causal: The image shows controlled chaos: liquid follows inertia, broken glass creates irregular shards, nothing looks ‘staged’.
Asymmetry is correct: Real breaks are never clean or symmetrical.
The scale is right: Droplet sizes, filaments and shards are in a plausible relationship to each other.
The timing is realistic: It looks like a real high-speed freeze, not a staged splash.
In your language:
Image 1 is not wild, but honest.
It’s untamed physics, but not exaggerated.
Another test, thanks @LarisaHaster for your input here 
Your attempt was again more aesthetically oriented, I tried to incorporate physics.
But as you can see, the fine splinters don’t really look like glass. Similar to the ‘wet sand’ effect in a snowstorm.
I guess this is a ‘bias-variance trade-off’ fail.
Analyse
- High bias
The model has a strong aesthetic priority:
- ‘Dramatic breakthrough’
- ‘Cinematic debris’
- ‘High-speed impact look’
→ These assumptions dominate the generation.
- High variance (untamed)
At the same time, too many concepts are allowed simultaneously:
- Glass
- Dust
- Water
- Smoke
- Splinters
- Particle light
→ No clear decision for a material class.
Sorry here for German 
| In ML-Sprache | Im Bild |
|---|---|
| Underfitting | Physik nicht gelernt |
| Overfitting | Filmästhetik reproduziert |
| Bias dominiert | Symbolik schlägt Kausalität |
| Varianz unreguliert | Material-Matsch |
Then I readjusted and got this:
With a more realistic template, the priority in the model shifts:
- less cinematic bias,
- less symbolic exaggeration,
- clearer material separation (glass ≠ dust ≠ liquid),
- better fragment geometry.
The result immediately appears ‘more correct’, not because it is suddenly physically correct, but because - the aesthetic bias is closer to the real reference.
I guess, it’s not about the “physics engine” or that physics has been better learned - regularisation has been improved 
This is precisely why good references work so well:
They reduce variance before generation instead of smoothing it out afterwards.