Official code for new flow-based model

2 slothofgalaxy 1 8/14/2025, 8:39:00 PM github.com ↗

Comments (1)

slothofgalaxy · 3h ago
TL;DR

- A flow's prior is typically fixed (e.g. N(0, I)). We learn it and use a lightweight flow to model pixel dependencies;

- This makes sampling (ODE solving) more efficient, without sacrificing performance in our setting;

- We introduce bespoke training objectives for both autoregressive and continuous-time flow variants;

- Flow-SSN achieves SOTA performance on standard stochastic segmentation benchmarks!