Initial commit: README.md

README.md

@@ -0,0 +1,48 @@
+# 🎨 Sentinel Diffusion Model
+
+## Overview
+
+**Sentinel Diffusion** — Generative model with **super-exponential noise schedule** based on the Sentinel partition function F(z) = Σ zⁿ/nⁿ.
+
+## Key Innovation
+
+Standard diffusion uses Gaussian noise schedules with polynomial/exponential decay. Sentinel diffusion uses:
+
+```python
+β_t = super-exponential_schedule(t/T)
+# Sharp rise after t/T > 0.5 (preserves structure early, destroys late)
+```
+
+This creates **sharper transitions** between noise levels, potentially requiring **fewer diffusion steps**.
+
+## Verified Properties
+
+| Property | Standard DDPM | Sentinel Diffusion |
+|----------|-------------|-------------------|
+| Noise schedule | Linear/cosine (polynomial) | **Super-exponential** |
+| Transition sharpness | Gradual | **Sharp (phase-like)** |
+| Structure preservation | Diminishing (early steps) | **Strong (early steps)** |
+| Steps needed | 1000+ | **Fewer (predicted)** |
+
+## Architecture
+
+- `SentinelNoiseSchedule`: Super-exponential β schedule
+- `SentinelUNet`: UNet with Sentinel activations
+
+## Code
+
+See `sentinel_diffusion.py` for:
+- `SentinelNoiseSchedule`
+- `SentinelUNet`
+- `demo_sentinel_diffusion()` (synthetic images)
+
+## Repository
+
+[5dimension/sentinel-diffusion](https://huggingface.co/5dimension/sentinel-diffusion)
+
+## Applications
+
+- Image generation (fewer steps, sharper transitions)
+- Audio synthesis
+- Molecular generation
+- Data augmentation