AS-Bridge: A Bidirectional Generative Framework Bridging Next-Generation Astronomical Surveys¶
Conference: CVPR 2025
arXiv: 2603.11928
Code: https://github.com/ZHANG7DC/AS-Bridge
Area: Image Generation / Astronomical Image Translation
Keywords: Brownian Bridge, Astronomical Surveys, Cross-Modal Translation, Anomaly Detection, Probabilistic Reconstruction
TL;DR¶
This work proposes AS-Bridge, a bidirectional Brownian Bridge diffusion model designed to model the stochastic mapping relationship between two major astronomical surveys: the ground-based LSST and the space-based Euclid. It achieves probabilistic cross-survey translation and rare event detection (strong gravitational lensing), and demonstrates that the epsilon-prediction training objective benefits both reconstruction quality and likelihood estimation.
Background & Motivation¶
Background: Over the next decade, astronomical observations will be dominated by two flagship surveys: LSST (ground-based, multi-band optical, ~0.7 arcsec seeing) and Euclid (space-based, high-resolution near-infrared, 0.1 arcsec pixel size), which overlap over an area of approximately 7,000–9,000 \(\text{deg}^2\).
Limitations of Prior Work: The two surveys exhibit vastly different PSFs, noise statistics, bandpasses, and scanning strategies, making direct joint analysis highly challenging.
Key Challenge: Cross-survey mapping is ill-posed in both directions: LSST-to-Euclid requires recovering details blurred by atmospheric seeing, while Euclid-to-LSST requires inferring multi-band spectra from a single band. Deterministic mapping struggles to capture such inherent ambiguities.
Goal: To build a unified bidirectional generative framework that models the probabilistic conditional distribution between the two surveys, and leverage reconstruction inconsistency for detecting rare astronomical events.
Key Insight: To model cross-survey translation as a Brownian Bridge diffusion process, anchoring both ends of the bridge to the respective data distributions of the two surveys.
Core Idea: Astronomical survey translation is inherently a bidirectional stochastic process. Thus, the Brownian Bridge is naturally suited for modeling the probabilistic relationship between two imperfect observations.
Method¶
Overall Architecture¶
The two survey observations are treated as noisy projections of the same unobservable astrophysical process. AS-Bridge establishes a stochastic path between them using a Brownian Bridge: the endpoints correspond to the observations of the two surveys, and the intermediate states are defined under a Gaussian distribution. Training is performed on paired data from the overlapping sky region.
Key Designs¶
-
Bidirectional Brownian Bridge Diffusion
- Function: Trains a bidirectional generative model on the overlapping sky region.
- Mechanism: The forward process performs stochastic interpolation from source to target (instead of mapping data to pure noise), and the reverse process utilizes Bayesian posterior transitions.
- Design Motivation: Standard diffusion requires passing through high-noise states, which is inefficient. The Brownian Bridge directly interpolates between the two distributions.
-
Epsilon-prediction Maximum Likelihood Training
- Function: Replaces the traditional bridge loss with a noise prediction loss.
- Mechanism: Proves that the \(\epsilon\)-loss is equivalent to the standard loss scaled by a weight of \(\sqrt{\Delta_t}\), balancing both high-noise likelihood and endpoint stability.
- Design Motivation: Directly using \(\Delta_t\) weighting results in weights tending to zero near the endpoints.
-
Cross-Survey Rare Event Detection
- Function: Translates the detection of rare objects (e.g., strong gravitational lensing) into reconstruction inconsistency.
- Mechanism: Uses midpoint fusion combined with reverse reconstruction, multiple sampling to find the minimum error, and flux normalization.
- Design Motivation: The model learns representations of "normal" astronomical objects well, whereas reconstruction failure of rare events generates a strong signal.
Loss & Training¶
- Epsilon-prediction loss, equivalent to the standard bridge loss weighted by \(\sqrt{\Delta_t}\).
- Flux normalization of anomaly scores to prevent bias towards bright sources.
- Dataset: 115K regular galaxies + 5K strong gravitational lensing events.
Key Experimental Results¶
Main Results¶
| Method | CRPS (L to E) | CRPS (E to L) |
|---|---|---|
| SPADE | 3.39 | 16.52 |
| OASIS | 4.65 | 13.33 |
| Pix2Pix | 4.35 | 73.03 |
| Palette | 2.43 | 7.98 |
| Joint Diffusion | 3.14 | 15.15 |
| AS-Bridge (epsilon) | 2.38 | 7.90 |
Ablation Study¶
| Training Objective | CRPS (L to E) | CRPS (E to L) |
|---|---|---|
| Standard Bridge Loss | 2.55 | 7.90 |
| \(\sqrt{\delta} \times \epsilon\) | 3.59 | 11.24 |
| epsilon (Ours) | 2.38 | 7.90 |
| Anomaly Detection | FPR@1%TPR | FPR@5%TPR | AUPR |
|---|---|---|---|
| AS-Bridge | 0.00% | 0.18% | 0.80 |
| Deco-Diff | 1.1% | 5.0% | 0.61 |
| CFM | 0.24% | 1.2% | 0.75 |
Key Findings¶
- Diffusion and Bridge methods consistently outperform non-diffusion baselines.
- Epsilon-prediction achieves the best CRPS in both directions.
- Multi-modal anomaly detection significantly outperforms uni-modal approaches.
- LSST-to-Euclid translation successfully recovers the multi-object structure of blended sources.
Highlights & Insights¶
- Translates cross-astronomical survey translation into a probabilistic inference framework for the first time.
- Provides a rigorous proof of equivalence between epsilon-prediction and the weighted standard bridge loss.
- Anomaly detection does not require annotations of anomalous samples.
Limitations & Future Work¶
- Validated only on simulated data; a sim-to-real gap remains.
- Anomaly detection is evaluated only on strong gravitational lensing.
- The dataset scale is limited.
Related Work & Insights¶
- Brownian Bridge Diffusion offers a new paradigm to replace conditional diffusion.
- The framework can be generalized to other multi-modal astronomical data.
Rating¶
- Novelty: ⭐⭐⭐⭐ First to apply the Brownian Bridge to astronomical survey translation.
- Experimental Thoroughness: ⭐⭐⭐ Thorough simulations but lacks real-world data.
- Writing Quality: ⭐⭐⭐⭐ Clear motivation and rigorous theory.
- Value: ⭐⭐⭐⭐ Inspiring for both astronomy and cross-modal translation fields.