planet-deconv: a small experiment in single-frame planetary deconvolution

A short note about a small thing I’m poking at: planet-deconv is a U-Net trained to map individual short-exposure planetary frames (or short bursts) onto the corresponding lucky-imaging stack, using existing PIPP / AutoStakkert!3 outputs as supervision.

The question I’m trying to answer is whether you can get something close to stack-quality output without keeping every frame from a session — useful either as a real-time preview during capture, or as a deconvolution stage in a pipeline.

Status, as of right now: research code, not a polished tool. The single-capture sanity loop works. Multi-capture training works on a frame-level holdout. Cross-capture generalization is the open problem and it’s where most of my time is going.

If you’re doing planetary imaging seriously and want to look at the training setup, the repo is public. Feedback welcome — especially failure modes, since that’s what I need most right now.