ppisp-ablation / README.md
ForenlyAI's picture
PPISP ablation renders + card
d54c984 verified
|
Raw
History Blame Contribute Delete
1.82 kB
---
license: cc-by-4.0
pretty_name: PPISP Scene→Sim Ablation (OFF vs ON)
task_categories:
- image-to-image
tags:
- 3d-reconstruction
- gaussian-splatting
- radiance-fields
- novel-view-synthesis
- ppisp
- nvidia
- scene-to-sim
- physical-ai
size_categories:
- n<1K
---
# PPISP Scene→Sim Ablation — OFF vs ON
Evaluation renders from a single controlled ablation of **NVIDIA's PPISP** inside
Forenly AI's **Scene→Sim** (real-scene → simulation) pipeline.
## What this is
The same real **3-camera capture** reconstructed with the **same 30,000 training
iterations**, changing exactly one variable — **PPISP off vs on**.
| Run | Result |
|-----|--------|
| PPISP **off** | hazy, washed-out blur; geometry and colour collapse |
| PPISP **on** | sharp, colour-correct, stable — **+9.5 dB PSNR** |
## Files
- `ppisp-off.jpg` — reconstruction render, PPISP off
- `ppisp-on.jpg` — reconstruction render, PPISP on
- `ppisp-sidebyside.jpg` — labelled side-by-side (off | on)
## Why it matters
In Scene→Sim, reconstruction fidelity **is** the training ground: a blurry scene is a
broken one — geometry the robot can't trust, textures the policy can't read. PPISP
targets the photometric artifacts (lighting/colour drift) that break radiance-field and
Gaussian-splatting reconstructions; it works across 3DGS, 3DGUT and NeRFs.
Methods benchmarked in this line of work: **PPISP · 3DGRUT · 2DGS · GOF**.
## Method / credit
PPISP is NVIDIA's method for physically-plausible compensation of photometric variations
in radiance fields. This card documents an applied ablation, not the original method.
Interactive viewer: [Space — ppisp-scene2sim](https://huggingface.co/spaces/ForenlyAI/ppisp-scene2sim)
· Study: https://forenly.ai/lab-study/ppisp
**Forenly AI — The Skill Layer for Humanoids.**