Unable to get decent results using 2.5.0 - can you?

(Morgan Hardwood) #1

I’ve been trying to use 2.5.0 to do some real processing, but have failed completely.

The requirements are simple:

  • Create a HDR file using bracketed input images so it does not contain artifacts or color casts.
  • Compress the dynamic range of the scene to not lose any highlight or shadow detail.
  • The result should encompass a higher dynamic range than any of the input images alone.
  • The result should not have more noise than any of the input images.

I already ran into a wall at the first point - the HDR files have artifacts in the highlights. Tried all formats, all profiles, using source images both raw and TIFF, with both sRGB and linear encoding. Opened an issue https://github.com/LuminanceHDR/LuminanceHDR/issues/43

Ignoring the artifacts, I was eager to see the dynamic range being compressed. Not so. A 5-image bracketed set, tone mapped, had as much clipping as the 4th image alone. Opened an issue https://github.com/LuminanceHDR/LuminanceHDR/issues/44

The last point also turned out to be a problem, but I found that using Profile 6 lead to a noiseless HDR file.

Here are two sets of bracketed photos:

(alternate download): 2015-12-04_karin_library.zip (93.9 MB)

Have a go at it. I’m eager to find out that I’m doing something wrong.

One of these two images is a tone-mapped HDR made from 5 images created in LHDR and then I applied only one simple curve in RT to fix the tones (because LHDR lacks curves and the levels tool is stone age). The other is a single image from the set I uploaded above.
Which one has more dynamic range, less clipped highlights, less noise, i.e. which one is the HDR?

HDR from single raw?
(Karl) #2

First off, I really like the composition of set #2 (the library.) It has a really good feeling of depth.

I’m not in front of my home computer right now, so I won’t be able to try Luminance 2.5, but my first impulse is that you haven’t captured the entire dynamic range of the scene - the bare candle flames look overexposed in your darkest frame, which is likely the cause of the color artifacts in the highlights. (In my experience Luminance doesn’t deal well with clipped highlights.) I also see a little TCA in the candles in the left side, which is going to be amplified when you tonemap after merging, which could also contribute to color artifacts.

IIRC profile 6 uses gamma-weighted response curve. If you’re working from RAW (or TIFFs that have not had a gamma curve applied) you should use a profile with linear response.

My advice would be to process the images with Darktable first - remove the Pentax base curve and sharpening that DT automatically applies, camera white balance, apply TCA correction (there should be a lens profile for the DA 18-55) and do profiled denoise, then export to 16bit TIFF. (If you want to get adventurous, try making a duplicate of your darkest frame, lower exposure by 1 or 2 stops and see if you can pull back some detail in the candle flames with the highlight reconstruction module.)

Then try merging the DT exported TIFFs with Luminance HDR using profile 5. This should remove color casts. I’ll see if I can try tonight when I get home.

(Morgan Hardwood) #3

@Karl thank you for responding.

I’m not interested in preserving candle flames or light bulb filaments. The parts that matter are not overexposed. It’s not realistic to expect someone to cover the full dynamic range, e.g. if bracketing 3 shots 2EV apart would cover everything I’m interested in showing, why bracket 9 shots only to capture the sun or headlights?

It does. I tried both linear TIFFs and sRGB ones, and ones which used a tone reproduction operator in the input profile and ones without. No difference as far as artifacts and highlights are concerned. There were small differences in the default saturation or the default lightness of the TMO’s output, but these differences were insignificant. See the linked bug report for a list of other things I tried. I even tried linear TIFFs without white balance applied, just for the fun of it, but they too had the same problems.

I already tried all that, but by all means give it a shot.

I hope this discussion brings attention to these issues so they may get identified fixed, as they’re not new.

(Karl) #4

You’re welcome - and thank you for giving me my first workout with 2.5! :slight_smile:

This statement confuses me… your post says that the HDR files have “artifacts in the highlights”, and the screenshot in the issue you opened on Github shows that the grey artifacts are in the highlights of the lamps (the center of the flame of oil lamp on the desk, and the sconces on the wall.) These gray spots are caused by overexposure (specifically, the green and blue channel) in your darkest frame.

(this is what I mean here - I thought these gray areas in the sconces were what you meant by “artifacts in the highlights” - top is Darktable’s raw overexposure indicator, bottom is from your post on issue 43.)

If this isn’t what you mean, perhaps you could post small 100% crops of the issue?

OK; I’ve noticed that mismatches between the images and response mapping can cause colour banding and false colours. It looked like there was some of that in the shadows around the right window, so I thought that’s what you could be talking about. Interestingly this banding also shows up when I use 2.5, but not when I use version 2.4.0, so this is definitely a bug in 2.5’s HDR creation. (In fact, profile 5 - my go-to for RAWs in earlier versions - was unusable with 2.5.)

Yes, the way Luminance HDR deals with blown highlights has always bugged me. It took me awhile to understand what was causing it and how to correct it in my workflow. Unfortunately, from this first real test, it appears that there could be some regressions as well (The banding artifacts might be caused by a library on my system, I’ll have to do some experimenting to see if I can track down the culprit, then file a bug report.)

I tried making a highlight reconstruction image (underexposed by 1 stop) with the rest and was able to recover some of the grey area of the sconces, but it’s not perfect and would still require some retouching (as well as curve adjustment.). This is about the best I can get without more time, which I won’t have for another few weeks.

(Morgan Hardwood) #5

@Karl yes, it’s the black-gray in the highlights I was referring to - clipped highlights should stay clipped, not any other color. Please note that when I write HDR I mean HDR, not the tone-mapped image. If I wanted to use the HDR image as a lightmap for instance I can’t because of those artifacts. Writing from memory as I don’t have the image before me, there also also artifacts of a different sort on high contrast edges, visible in the full-sized image, not in my downscaled sampled in the first post.

(Morgan Hardwood) #6

Great news, Franco of Luminance HDR is fixing the issues! I haven’t confirmed the patches yet but am thrilled to see that they were addressed!

(Franco Comida) #7

I guess we can say the clipping issue has been fixed (mantiuk06, ferradans, fattal).

(Morgan Hardwood) #8

I confirm that the clipping and the artifacts issues have been fixed - good work @fcomida!
For anyone interested, see the two GitHub links in the first post.