GPT5 minimalist work flow in PixInsight.

9 replies306 views
Himanshu Pandey avatar

I asked GPT 5 for a minimalist workflow to process a dual narrowband OSC image in PixInsight. I asked it to consider all the freely available 3rd party plug ins as well native PI processes and to recommend the best process for each step. I constrained it to avoid separating/recombining stars as well use of masks, just a very simple direct process to a final image. I understand how much power PI has under the hood and this is just my way of diving in.

Below is what it gave me, what do you think?

———————————————————-

  1. DynamicCrop

    • Remove stacking artifacts and edge gradients.

  2. Gradient Removal

    • GraXpert (Astro model, Medium strength)
      or ABE (Degree 2–3, Subtraction) if you want native only.

  3. Linear Noise Reduction

    • MURE Denoise (if gain/RN known)
      or MLT (5 layers, conservative thresholds) if not.

  4. Channel Extraction & OIII Mix

    • Extract R = Ha, G/B = OIII.

    • PixelMath: OIII = 0.7*G + 0.3*B (adjust weights by SNR).

  5. LinearFit

    • Reference = Ha, apply to OIII.

  6. HOO Combine (Normalized)

    • PixelMath:

      Hn = Ha/(Ha+OIII+1e-6); On = OIII/(Ha+OIII+1e-6); R: Hn G: 0.2*Hn + 0.8*On B: On

  7. BackgroundNeutralization

    • Small preview on true background.

  8. Stretch

    • GHS (2 passes) → HT for fine black/midpoint placement.

  9. Local Contrast

    • LHE (Kernel ~64, Amount 0.3) global.

  10. Color Tuning

    • Curves (global saturation/hue).

    • SCNR (Green, 0.25) only if needed.

    • CorrectMagentaStars (global).

Helpful
Quinn Groessl avatar

“minimalist” I’d say it failed like it normally does with such tasks.

Himanshu Pandey avatar

Quinn Groessl · Sep 10, 2025, 09:38 PM

“minimalist” I’d say it failed like it normally does with such tasks.

You’d get a kick out of its 1st response…literally twice as long as what I posted….

Oscar avatar
I disagree with the linear denoise (I dont think anyone really uses Mure, and MLT/MMT is hard to deal with without stripping away detail; just use NXT, or DeepSNR - which is free), and also the local contrast part (simply applying LHE without masks is horrible advice), also what I assume CorrectMagentaStars does, can be easily done with SCNR Green on the image, inverted.

EDIT: Also, I cant understand the pixelmath, but it honestly doesnt look good
Himanshu Pandey avatar

I asked it again with a slightly different prompt and got the following which is inline with how I started this work in Siril.

Here’s the same minimalist, mask‑free, star‑intact OSC dual‑narrowband workflow for Bortle 8.6 skies, restated as plain text so you can copy and paste without table formatting.

1. Linear Preparation
1.1 DynamicCrop — Remove stacking edges. Define crop once on a reference frame and save as a process icon for reuse.
1.2 AutomaticBackgroundExtractor (ABE) — Remove light pollution gradients. Use function degree 3–4, correction mode set to Subtraction, and replace the target image.
1.3 NBColourExtraction (Bill Blanshan script) — Extract Ha and OIII from the OSC stack. Set the correct CFA pattern and filter wavelengths (Ha = 656.3 nm, OIII = 500.7 nm).
1.4 MultiscaleLinearTransform (MLT) — Linear noise reduction. Use 3–4 layers with low thresholds (L1 = 2.5σ, L2 = 1.5σ, L3 = 1.0σ) and an amount of 0.6–0.8.

2. HOO Combination
2.1 PixelMath — Combine Ha and OIII into an HOO composite:
R: Ha
G: 0.85*OIII + 0.15*Ha
B: OIII
Create a new image in 32‑bit float, linear.
2.2 SPCC (Spectrophotometric Color Calibration) — Use narrowband mode with correct filter wavelengths. Enable background neutralization.

3. Stretching
3.1 GeneralizedHyperbolicStretch (GHS) — Global stretch with stretch factor ~3–4, symmetry point ~0.25–0.3, and highlight protection enabled.
3.2 CurvesTransformation — Apply a mild S‑curve in the L channel to boost contrast.

4. Non‑Linear Refinement
4.1 CurvesTransformation — Adjust hue for the desired HOO palette and apply a mild global saturation boost.
4.2 BlurXTerminator (if available) or Deconvolution — Apply global sharpening with low iterations to avoid ringing.
4.3 MultiscaleMedianTransform (MMT) — Non‑linear noise reduction. Use layers 1–3, amount 0.4–0.5, with low thresholds.
4.4 MorphologicalTransformation (optional) — Reduce star sizes without masks. Use erosion with amount 0.15–0.20, 1–2 iterations, and a circular structuring element size 3.

5. Finalization
5.1 CurvesTransformation — Apply a small global saturation boost in the S channel for final color pop.
5.2 ProcessContainer — Save all steps with fixed parameters for one‑click reproducibility.

Well Written Helpful Insightful
Shinpah avatar
It spit out mostly a bunch of gibberish (which is typical).

As far as I can tell "NBColourExtraction (Bill Blanshan script)" doesn't exist. Perhaps it means NB Normalization?

There are numerous redundancies, particularly the color calibration for mono being duplicated across NB normalization, and the Pixelmath, and SPCC, and the Nonlinear Hue adjustment. The deconvolution step explicitly is misplaced.

I wouldn't trust any of the specific numerical values it assigned as a setting for any pixinsight process. That's really the meat of processing and naturally has some trial and error.
Helpful Insightful
Gilmour Dickson avatar

What do I think? I think there is a rather odd irony that you come to ask a forum of real people their opinion of the results from you asking AI for a processing workflow. Why don’t you skip all the AI stuff and just ask real people right from the beginning? I see more and more posts like this here and on cloudy nights and each and every time it leaves me scratching my head.

Well Written
AstroStew avatar

Gilmour Dickson · Sep 11, 2025 at 06:02 AM

What do I think? I think there is a rather odd irony that you come to ask a forum of real people their opinion of the results from you asking AI for a processing workflow. Why don’t you skip all the AI stuff and just ask real people right from the beginning? I see more and more posts like this here and on cloudy nights and each and every time it leaves me scratching my head.

Well what do you expect from a forum that openly promotes AI, being moderated and practically run by it, it’s the future unfortunately, people just can’t think for themselves anymore, and can’t be bothered to ask and talk to other people, social media is bad enough for killing human interaction, but now there won’t even be this in the future, instead we will just ask AI the questions. It’s a sad works we now live in, and after last nights news, it’s just getting worse..☹️

PS, I doubt this post will pass AI moderation 😂

Stjepan Prugovečki avatar

Everyone has his/her own processing flows and tips and tricks. One thing in your processing flow is by my opinion very doubtful : Deconvolution (either “classical” or AI like BXT) being done AFTER noise reduction. This does not make much sense…

Himanshu Pandey avatar

Stjepan Prugovečki · Sep 11, 2025, 08:48 AM

Everyone has his/her own processing flows and tips and tricks. One thing in your processing flow is by my opinion very doubtful : Deconvolution (either “classical” or AI like BXT) being done AFTER noise reduction. This does not make much sense…

Thank you for using your time to comment constructively.

Well Written Respectful Supportive