autoshade-raw-denoise-v2.pth
AutoShade's RAW-mosaic denoiser: KAIR/DPIR's DRUNet-colour architecture (MIT) carrying weights fine-tuned by the
AutoShade project for its own pipeline (generalised Anscombe transform,
per-plane noise model and noise field) on real RawNIND pairs (Brummer & De Vleeschouwer, CC BY-SA 4.0) and
synthetic sensor noise; v2 continues the v1 fine-tune with point sources injected on the clean side so that faint
stars are no longer removed as noise (scripts/train_raw.py --stars 0.5 --loss l2x, 60000 steps, the step-50000
checkpoint chosen by the pre-committed acceptance lines).
This repo is the project's own SECOND host for the file it publishes on GitHub:
https://github.com/skymanbp/autoshade/releases/download/v1.5.2/autoshade-raw-denoise-v2.pth. The sidecar (python/denoise_raw.py) tries this copy first and the release after it, and
accepts either only when the bytes match the pin it carries:
- sha256
ffafa40a53f52092149db2fcf03636117ad6855e1068142d4f6b03b634e9f9c4 - 130,590,559 bytes
Licence: MIT for these weights and the architecture; see LICENSE beside this file. The RawNIND pairs the weights
were fine-tuned on are CC BY-SA 4.0 and are credited in the project's docs/TECH_STACK.md.