HDR module in 1.8.515 Public Alpha is extremely slow

[BrendanC]

Oh good grief, I'm making a bit of a dick of myself aren't I?

So sorry about the FITS file. I've put the right one in now. Oddly enough, when I was supposed to be capturing the Bubble Nebula a while back I accidentally typed in the coords for the Cocoon! What is it about those two objects that I keep getting confused in my head?

Anyway, love your analysis Mike, and I hope this all helps. Ivo, I'm really pleased that the HDR module has seen some tweaking as a result of all this - that's what alphas are for!

What I'm wondering is whether the Low quality could go any lower and therefore be faster? It's so close to the Medium currently, and whereas Mike thinks it could be jettisoned, for people like me, a faster Low setting would be more useable and more distinct from Medium.

[Mike in Rancho]

No worries, Brendan. It still gave us data points, and another comparison for percentages.

I guess not so much that Low Q needed to be jettisoned, just that there's some room there to do something. Perhaps Med Q can have higher quality and take a little longer, if that even makes sense or would help with the image? I think Ivo said Low Q is already bottomed out as far as it's going to go.

As we are already starting out well past the default point of old-HDR, I do wonder just what image results difference we will achieve from increasing the context level higher and higher. I'm not sure an assumption that it is "better" just because it takes longer actually holds up.

I didn't continue testing deeper due to the time (both of the module and on the clock lol), but it may be that Low Q doesn't really show its value until pushing quite high context pixel levels, and on large resolution sets.

Likely to be dataset dependent, of course. But also as Guy noted, the scale being pixel-based will thus be relative to your capture resolution and what kind of binning was done.

[BrendanC]

Yes, the scale being dependent on the capture might be a good way to do this. Bigger images could, on the whole, indicate a more generally sophisticated setup.

[LuckyEddie]

Very unscientific but, the '20 minute initial calc' dataset is now down to just over 3 minutes for the default settings on module entry, with 1.8.516 on the same machine. That's fine for me, and similar to what I get with the Super Structure module.
These new defaults are, to my mind, a much better starting point, and of course you can still crank it up if you have the horsepower to support that.
Thanks Ivo

[xonefs]

Are you guys getting good results with the new module?

The presets all make the few images I've tried look comically bad. This is with reveal on defaults (the other presets look even worse)

HDR:

before:

I just want to reveal some structure in the middle I know is there but not like that

I'm trying to change parameters to mess with it, but it takes so long to render each time anything is changed I'm still waiting and not sure what parameters need adjusting since each adjustment can be 20 minutes. I had similar issues when I was working on an andromeda image the other week it made the core look very bad/fake and wrinkly so I just gave up on it.

[admin]

Are you guys getting good results with the new module?

The presets all make the few images I've tried look comically bad. This is with reveal on defaults (the other presets look even worse)

HDR:

before:

I just want to reveal some structure in the middle I know is there but not like that

I'm trying to change parameters to mess with it, but it takes so long to render each time anything is changed I'm still waiting and not sure what parameters need adjusting since each adjustment can be 20 minutes. I had similar issues when I was working on an andromeda image the other week it made the core look very bad/fake and wrinkly so I just gave up on it.

That would be (likely) the result of using the module on oversampled data, while not starting off with a reasonably balanced larger scale dynamic range to begin; the input image appears over-stretched, and does not appear to have the Contrast module applied first. The dynamic range problem in that image is very much still one of large scale, not small scale;

Your workflow should look like this and deal with progressively finer detail;

1. If applicable, bin to make sure you're not massively oversampled.
2. Use AutoDev with a value set for 'Ignore Fine Detail <'
3. Use the Contrast module
4. only then use the HDR module to resolve finer detail.
5. Use Decon to resolve the finest detail

Running your input image at your published (small) resolution through Contrast, and only then the HDR module yields expected results;

JW549gg.jpg (280.88 KiB) Viewed 6587 times

(all defaults)

[xonefs]

That would be (likely) the result of using the module on oversampled data, while not starting off with a reasonably balanced larger scale dynamic range to begin; the input image appears over-stretched, and does not appear to have the Contrast module applied first. The dynamic range problem in that image is very much still one of large scale, not small scale;

Your workflow should look like this and deal with progressively finer detail;

1. If applicable, bin to make sure you're not massively oversampled.
2. Use AutoDev with a value set for 'Ignore Fine Detail <'
3. Use the Contrast module
4. only then use the HDR module to resolve finer detail.
5. Use Decon to resolve the finest detail

Running your input image at your published (small) resolution through Contrast, and only then the HDR module yields expected results;

JW549gg.jpg

(all defaults)

thank you! that's probably likely this is data from my old 294mm in bin1 mode I'm just editing so it is likely a bit oversampled, but not massively at .9"/pixel. I thought my seeing should be able to support that. I can try binning it and I'll try all that.

[Mike in Rancho]

Also don't forget that it's only the first calculating that really can go long, or if you change the pixel scale up. Or I guess, change to high quality now that those options are available. But as noted in the discussion, even the default 50 pixels is really more than was available under old HDR.

The other changes should be quite fast, and you can dial things back for a minimal, soft HDR-ing. To quickly do that from the default settings, back off the gamma highlight (way back if you want) as well as the highlight and shadow boosts.

Of course every image is different in target detail, as well as possible scale that you are at when entering HDR.

[fmeireso]

Just recently turned to St 1.8.525 and i think the new HDR works really well. I could get alot of detail out in my M42. I found also that indeed context size is somehow good between 30 and 50 pixels.

Still have to test this for globs, i did one run on my M92, it did not quite came out as i would have prefered, but i did only one try sofar.

As for CPU and GPU time, HDR needs a bit of time but on my system not bothersome...of course i bin alot but evenso works pretty smooth. I use a Ryzen 5 3600 and a modest GTX 1650 GPU.

Set aside the HDR module , it looks, for some reason my stars come out better in the newer version. In the former ones they came out as having a strong core and small halo like. But in the new version this seems strongly reduced, depends of course of stretching and other things but overall it has become less then an issue...

Another thing that i noticed, allthough this has nothing to do with Startools, that using another stacking program then DSS like PI in my case really does a better job. This just come out better and then afterwards processing in Startools sure goes alot better. It is just a better start.

To conclude, i really like ST v 1.8, as the versions go up, all seems to get better