We show ablation results for four key components of our method (part of which are shown in Fig. 6). We evaluate the following variants:
- Without collision handling (w/o CH) - removing mesh-based collision handling techniques results in Gaussians floating unnaturally above surfaces instead of accumulating realistically.
- Without motion simulation (w/o Motion) - using only data-driven optimization with Video-SDS yields incoherent and physically implausible motion.
- Without appearance optimization (w/o App) - using only physics simulation with LLM-initialized appearance lacks photorealism.
- Without physics guidance (w/o PG) - fixing motion trajectories from physics simulation and optimizing only appearance prevents effective joint optimization, as the model cannot refine motion to achieve photorealistic results.
| w/o CH | w/o App | w/o Motion | w/o PG | Ours |
|---|---|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|