Loading:

  • Deep Sky Colors
  • astrophotography
  • stars
  • tools
  • books

Forum

DSPX 1.0.0 – full M...
 
Notifications
Clear all

DSPX 1.0.0 – full M31 workflow test: feedback, bugs and comparison with PI/AW

2 Posts
2 Users
0 Reactions
224 Views
(@sergiusz)
Eminent Member
Joined: 2 months ago
Posts: 12
Topic starter   [#3362]

Hi Rogelio,

I spent a full afternoon running a complete M31 workflow in DSPX 1.0.0 (Windows, RTX 4050 via DirectML), from 104 raw OSC frames (Player One Ares-C, Optolong L-Pro, FRA400 + 0.7x) to a signed final, and compared it with my native PixInsight and AdaptiveWorkflow versions of the same data. Overall: very impressive for a 1.0. Here is what I found, hoping it helps.

WHAT WORKS GREAT
- Pre-processing: 2 min 03 s vs 8 min 26 s in WBPP with matched settings (LN, PSF Signal Weight, ESD, CFA drizzle 1x). Master noise measured on the drizzle masters is on par (R ~4% lower in DSPX, G/B equal). 104/104 frames registered, satellite trail detected automatically. The plain-language run report is excellent.
- Background Extraction automatically rejected a sample sitting on the galaxy halo ("nebulae, galaxies"). Very useful.
- Color Calibration with "Gaia stars, average spiral galaxy white": 1620 stars measured, clean result.
- Masked Stretch turning the screen stretch off by itself is a great touch.
- Speed overall: Deep Sharpen 40 s for the whole image on GPU, Noise Reduction 1.6 s, Deep Star Removal ~5–6 s.
- Local Contrast (clarity/dehaze) and Deep Star Mask with hand edits: exactly what was needed.

POSSIBLE BUGS / UX ISSUES
1. Switching from the Background Extraction tool to the Selection tool silently APPLIED the pending DBE using the full automatic 64-sample grid (log: "The Background Extraction tool had unapplied work… applied"). On M31 that grid sat on the halo and removed much of the galaxy. Undo worked, but I'd expect either discarding the pending work or a confirmation prompt.
2. Run report shows "engine.mode = fast" although "Best quality" was selected in the pre-processing window. Maybe two different settings, but it is confusing.
3. Ctrl+Z goes to whichever window has focus (e.g. a process window), so it looks like nothing happened. Undo/redo also leaves no line in the log, which would help confirm what was reverted.
4. After processes that lower the background (DBE, Background Neutralization) the existing screen stretch is stale and the image looks almost black/clipped. A "re-stretch after apply" option (like Masked Stretch does) would avoid false alarms.
5. Catalog Star Color preview says "the color calibration is fitted to the stars", which reads like a whole-image calibration, but only star pixels change (3.08%). Wording could be clearer.
6. Ctrl+K search for "stretch" offered a Siril script (AutoStretch, wanting to install PyQt6) where I expected the native process. Showing native processes first, or marking scripts more clearly, would help newcomers.
7. Text process: color fields are 0–1. A color picker or hex/0–255 input would make it easier to reproduce a signature color.

ALGORITHM OBSERVATIONS
8. Deep Star Removal on LINEAR M31 data (Linear checked, Additive) put the whole galaxy core, M32 and part of M110 into the star image. On non-linear data the core mostly stayed, but M32 still went to the stars. A protection for compact galaxies/cores (or a warning when large extended objects end up in the star image) would make linear star removal safe for galaxies.
9. HDR Compression on a Masked-Stretched M31 flattened the core even at 9 layers. Probably expected behavior, just noting it.

Thanks for the great work. I'm happy to share the logs, run report and comparison crops if useful.

Best regards,
Sergiusz Gobiecki



   
Quote
RBA
 RBA
(@rba)
Member Admin
Joined: 4 years ago
Posts: 22
 

Hi Sergius,
Thank you so much for such detailed report!! Answers to the bug list:

1) Look at Preferences>General. If "Auto-apply tools" is checked, that's the expected behavior. If you'd rather have the behavior where the user has to explicitly apply the tool for it to take effect, uncheck that.

2) “Best quality” and engine.mode = fast are two different settings: “Best quality” is Easy mode’s goal. engine.mode = fast means the PixInsight-parity mode is off, that's all. I'll try to make this clearer.

3) This is real. I'll make Ctrl-Z to only act on controls when one is actually typing, and all undo/redo's will get an entry in the log. Expect this on Beta 11.

4) Following your suggestion via an option in the same panel as #1

5) I'll check the wording. Catalog Star Color is definitely not a color calibration tool.

6) Results are fit to "best match" but I'll give priority to native processes and scripts over Siril's.

7) This is already a color picker.

8) Changes in what the model recognizes as stars requires retraining, not code changes, and a training session is a lengthy process, from preparation to completion. I am not training the star removal model at the moment. I'll likely get back to that when DSXP has been out in the jungle for a couple of months or so.

Thanks so much again!
RBA



   
ReplyQuote
Share:
This website uses cookies to improve your experience. Cookie Policy