profiling sensor and photon noise

… and how to get rid of it.

[update 20/12/2012: ‘how to profile your camera’ includes instructions with the new gen-profile script]

[update 15/12/2012: no more recompile needed, updated the instructions in the benchmark section and how to run]

to summarize the current situation in dt: we have a lot of cool tools wrapped around great algorithms with almost all the knobs you need to get perfect results. while you can actually get really great results it’s this sheer number of knobs that makes finding a good parameter set quite a time consuming task. even creating per-iso presets is not straight forward, as most of the current modules depend on a lot more stuff early on in the pipe (whitebalance, exposure, basecurve, etc).

in the following, i’d like to look into automating this process and (after some calibration work, which could be optional, but less optimal) then something closer to a single-click solution. the plan is:

  • design a module where the denoising performance only depends on camera and iso setting.
  • profile (camera, iso) pairs and store presets for each one of those (can be interpolated potentially).
  • auto-apply them in switched-off state, so you can enable the right thing by just clicking the module’s ‘off button’ once.

[update 20/12/2012: this work is now in git master]


a lot of research is centered around denoising images corrupted by additive white gaussian noise. this means that your input signal x is observed as

$$y = x + n, n \sim \mathcal{N}(0, \sigma^2).$$

unfortunately, digital cameras don’t follow that pattern, so the noise is mostly dominated by some sort of signal-independent sensor noise and signal-dependent photon noise (or ‘shot noise’). this non uniform noise follows a poissonian-gaussian distribution [2]

$$y = \alpha p_i + n_i, p_i \sim \mathcal{P}(x), n_i \sim \mathcal{N}(0, \sigma^2).$$

a poissonian is a discrete distribution, emerging from quantized photon behaviour. alpha converts this to a number range we use in image processing ([0, 1] with extrapolated hdr headroom in the case of dt). to be able to apply general denoising procedures, we need to stabilize this noise via a transform [0].

denoising is then done in three steps:

  1. transform to map variance to unity uniformly all over the image,
  2. denoise with a generic method against additive white gaussian noise,
  3. backtransform

academia seems to fight additive white gaussian noise with current king-of-the-hill bm3d [6] and extensions. dt uses bilateral filters or non-local means [5]. the latter combined with a relatively simple extension [4] has been shown to perform quite well (fast, next to no implementation overhead for us).

there are very elegant wavelet based methods, too [3], so i also tried to drop in our wavelet code using bayes-shrink with an overcomplete wavelet transform, i.e. do a soft threshold shrinkage with threshold:

$$T = \sigma^2/\sigma_x = \sigma^2/\sqrt{\max\left\{0, \sigma_y^2-\sigma^2\right\}}.$$

the results had certain wavelety artifacts, so the following is using our nlmeans code to remove additive gaussian noise.


to illustrate the above noise model, here’s a plot of noise standard deviation sigma vs. brightness from a canon eos 5dm2 at iso 3200:


note how noise increases at higher brightness (caused by the poissonian distribution) and has a global offset (caused by the additive gaussian). only the green channel is saturated, the other two channels absorb too much of the light to go all the way to the right.

noise is estimated via mad (median of absolute deviations) on the finest hh coefficients of a wavelet transform, assuming zero mean. this is why the calibration shot needs to be out of focus, we want to minimize the spill of edge information into the hh coefficients. the median estimator is robust to a few outliers, but we still want to violate the assumption that the finest wavelet coefficients only contain noise as little as possible.

to fit an analytical model to these curves, we approximate

$$\mathcal{P}(x) \approx \mathcal{N}(x, x)$$

and fit this noise model to it [2]:

$$\sigma(y) = \sqrt{\max\left\{0, a\cdot y + b \right\}}.$$

note that our fit in the above graph (smooth curve, green) goes below zero, which can be caused by the black level (which is subtracted in these input numbers).

these fits can be used to construct conversion curves [1], which look as follows:


applying those to an image, and then computing the noise levels vs. brightness again, we get this nice flat graph that mostly shows sigma=1 (which indicates white additive gaussian noise, yay!):


this is easy to process with your favourite denoising filter (wiener filter and dct, bm3d, wavelet-based methods, nlmeans etc).

all that remains to be done is to map the values back to their original range. note that the algebraic inverse of the mapping is the wrong thing to do; we want an unbiased inverse [1].


this table shows some numerical results in terms of peak signal to noise ratio before and after denoising. the reference image is the mean of two iso 100 shots.

iso psnr noisy psnr denoised psnr handtuned
100 41.3101 41.9915
100 40.6879 41.3296
400 34.5583 38.4054
800 31.0149 33.6109
1600 29.5853 33.9332
3200 25.8187 29.4477
6400 23.5462 28.7782 28.5843
12800 20.4019 25.5775 25.6052/25.629 (*)
25600 17.1598 24.7214
(*) handtuned using old nlmeans/equalizer and handtuned using the new module.

interestingly, i get better results out of iso 1600 than from 800. this can have a lot of causes and it would have been a good idea to measure the camera’s native iso levels instead of powers of two.

here are a couple of 1:1 screen grabs to illustrate the numbers. from top to bottom: iso 3200, 6400, 12800 and 25600 with one-click denoise results:





iso 25600 seems to have trouble with the black level, judging by the color cast (lifting the black level manually helps a little, fixing whitebalance instead of using auto made things worse). apart from that i’m still not going to recommend it for everyday use, but maybe downsized resolutions will be useful for some purposes at times.

how to profile your camera

the parameters (a, b) have to be found for each (camera, iso) pair. these are then stored in a preset for a new denoising module. there is a generic (poissonian only) preset, but you can’t dial in these values manually, you have to create a preset for your camera and iso (if it doesn’t already exist).

to generate a profile, dt ships a couple of scripts and tools in tools/noise/. it will create presets as well as a couple of detailed data plots used to verify the usefulness of the result, such as the curves seen above.

the steps in detail:

taking noise profile images

take one shot per iso. it has to be out-of-focus and have a widespread histogram, i.e. be over- and underexposed at the same time (see example image below). if you want to be thorough, take 2–3 to verify the numbers. this image looks very green, because the ‘raw linear’ style has been applied to it (find it in darktable/tools/noise). this makes sure we get linear camera raw rgb with no basecurves, no whitebalance, etc in the way. it is essential that you apply this style.

you can either take those shots manually or have


take them for you via gphoto2. you need gphoto2, exiv2 and gnuplot installed to run the script (well, plus various coreutils).

your images should look something like this:


generate plots and fits

open a terminal, go to the directory you just exported your images to. from there, type

~/darktable/tools/noise/gen-profile -d .

this will create a ton of informative pdf files, which should look like the above examples. if the fit doesn’t match the curves, or the curves look wildly different than the general shape in the plots above, something went wrong (most likely your image wasn’t overexposed enough, so check the histogram). this script will output a file presets.txt with copy/pastable presets for your camera, to be shipped with dt. one line will look something like:

{n_("canon eos 5dm2 iso 3200"), "Canon", "EOS 5D Mark 2", 3200, {1.0f, 1.0f, {4.494e-05, 4.494e-05, 4.494e-05}, {-1.063e-06, -1.063e-06, -1.063e-06}}},

the exif data (model, maker, and iso) is used to do automatic matching of the correct preset, so you only need to enable the module.

additionally, this script will copy your darktable database, inject the newly created presets for you to test, and instruct you how to run dt with that library for a test.

run a benchmark (optional)

this requires additional shots. since we want to verify what we did was reasonable, we want (and need) independent input.

to test the presets you just created, you have to run darktable as instructed at the end of the script, something like:

darktable --library /tmp/library.db

take one image for each iso setting you’re interested in (same iso sequence as above). for iso 100, take 2–3 of them. it is essential that these images only differ in iso and exposure time (take them with a sturdy tripod and aperture priority mode).

create a reference image (use the ‘create hdr’ button in lighttable mode with a couple of iso 100 shots).

you don’t need to apply any styles this time, leave the images as natural as possible (but all the same).

export your images as float pfm to the same directory (filename iso$(EXIF_ISO) is a good idea), also put the reference as reference.pfm in there.

apply the newly created presets to your images, export them in that directory too (float pfm, use iso$(EXIF_ISO)_denoised as a filename for example).

then run


to verify that the psnr values are reasonable. the denoised version should always have a higher psnr than the input (even for iso 100), the values should be in the range of the above table.

ship it!

send your new presets to one of our mailing lists, or upload them to some pastebin or similar and let us know in irc. we might also be interested in the pdf files with the plots, to verify your fits worked out and everything looks fine.

using the new ‘denoise (profiled)’ module

the goal of this work was to give you a simple, single-click denoiser that always just works (tm). there are still a few options exposed for your tweaking pleasure, though these affect how the nlmeans core works. the conversion curves are not exposed to the user, but there is a generic preset for non-profiled cameras which will try to roughly do something reasonable.

  • patch size: 0 will degenerate to a bilateral filter. increasing the number will increase sharpness in your output, but you might have to increase ‘strength’ to counter balance the effect.
  • strength: make denoising stronger (> 1) or weaker (< 1). mostly a matter of taste. the default setting was chosen to maximize psnr in my tests (with a patch size of 1) .

future work

it seems that the fit parameters (a, b) behave quite nicely with iso, so it might be possible to interpolate them for your exact iso setting. see for example this plot of the parameter a (poissonian component) for a canon eos 5dm2 indicating that simple linear interpolation will be enough:


it should also be easy to auto-create this graph as a quick sanity check for new presets.

the influence of exposure time on the gaussian part is another thing to take into account, especially the case when the black frame is subtracted in camera.

[update 23/12/2012: this interpolation is now happening automatically in git master when you enable or reset the module]

the bottom line

if you’re not into signal processing and don’t like to profile your gear, just hope someone already measured your camera, enable the new denoising module and enjoy parameter free optimal denoising. it’s magic!


  • [0] paul prucnal, bahaa saleh: transformation of image-signal-dependent noise into image-signal-independent noise, optics letters, vol. 6, no. 7, 1981.
  • [1] markku makitalo, alessandro foi: optimal inversion of the generalized anscombe transformation for poisson-gaussian noise, ieee trans. image process.
  • [2] alessandro foi et al.: practical poissonian-gaussian noise modeling and fitting for single-image raw-data, ieee trans. image process. vol. 17, no. 10, 2008.
  • [3] frolian luisier, thierry blu: a new sure approach to image denoising: interscale orthonormal wavelet thresholding, ieee trans. image process. vol.16, no. 3, 2007.
  • [4] charles kervrann and jerome boulanger: optimal spatial adaptation for patch-based image denoising. ieee trans. image process. vol. 15, no. 10, 2006.
  • [5] antoni buades, bartomeu coll, jean-michel morel: a non-local algorithm for image denoising. proc. cvpr 2005.
  • [6] k. dabov, a. foi, v. katkovnik, and k. egiazarian: image denoising with block-matching and 3d filtering. proc. spie electronic imaging, 2006.


in this short time, you guys contributed presets for:

  • canon eos-1ds mark ii
  • canon eos 20d
  • canon eos 30d
  • canon eos 350d
  • canon eos 400d
  • canon eos 40d
  • canon eos 450d
  • canon eos 50d
  • canon eos 550d
  • canon eos 5d mark ii
  • canon eos 5d mark iii
  • canon eos 600d
  • canon eos 60d
  • canon eos 7d
  • canon powershot g10
  • canon powershot s90
  • konica minolta dynax 5d
  • nikon d300
  • nikon d3100
  • nikon d5000
  • nikon d5100
  • nikon d600
  • nikon d7000
  • nikon d80
  • nikon d800
  • nikon d90
  • olympus e-30
  • olympus e-400
  • olympus e-420
  • olympus e-m5
  • olympus e-pl1
  • panasonic dmc-fz18
  • panasonic dmc-g3
  • panasonic dmc-g5
  • panasonic dmc-gf1
  • panasonic dmc-gx1
  • pentax k100d
  • pentax k10d
  • pentax k200d
  • pentax k-5
  • pentax k-7
  • pentax k-x
  • sony dslr-a200
  • sony dslr-a230
  • sony dslr-a550
  • sony dslr-a700
  • sony nex-3
  • sony nex-6
  • sony nex-7
  • sony nex-c3
  • sony slt-a55v
  • sony slt-a65v

which seem to work great, thanks for that!

this also enables us to do nice cross-camera noise comparisons, like so:


lower would be better for both axes, but note that lower gaussian noise level might just mean higher black level. so i wouldn’t base my camera purchase on these plots just yet.

note how the four versions of the eos 400d gather around almost the same spot. three of those have been taken at almost the same time, and one of them comes from a different day with different lighting. this shows that the method can produce consistent results (and that you might gain some accuracy by averaging a few runs). it would now be interesting to see how far away from that cluster a different device of the same type is, to verify using someone else’s presets is a good idea.

someone noted that, at times, noise still stays in your image, visible as large color blotches (not fine grained anymore). these can be alleviated by the equalizer, using something similar to this style.

Filed under: blog development upcoming feature
These are comments from the old website, archived as static HTML
  1. Wow, that looks awesome! Can't wait to profile my K5.

    Few questions:
    * should the light causing the overexposure be white one (will LED one be too "bluish")?
    * can the underexposed part be made by covering half of the lenses?
  2. Wow can't wait too with my pentax k-5 also !
    Is it possible to make it in Darktable 1.1 released or do I have to use dev version ?
  3. Oh man, this is amazing! Excellent work. I was hoping after my recent message on the mailing list that you smart guys were working on this!
  4. Another question: most cameras allow "partial" ISO settings... so in addition to the 100/200/400/800/1600/3200, you can set it to 1000, or 1250, or whatever. It seems that this is either "auto" or in 1/3 stops.

    In the case of auto, it's not possible to do all measurements, of course... but for 1/3 should we try? You mentioned interpolation, which is why I ask.
  5. Any camera doing Auto ISO will still stick either to the full stop ISOs (100/200/400/etc) or the half or third stop ISOs (200/250/320/400/etc). As far as I'm aware no camera does truely stepless ISO..
  6. Bruce Guenter on Tue Dec 11 20:48:17 2012:
    I've taken several shots at each ISO setting. Can I average the a[3] and b[3] values together to make the result more general?
  7. AMAZING!!!

    Have to try it this weekend.

    Thanks for the hard work.
  8. José Carlos on Tue Dec 11 20:56:02 2012:
    This is just great!
    Can't wait to try it.

    Excellent work.
  9. small remark, to generate the plot "gnuplot" is needed
  10. I've already profiled the K-5, the results are amazing.
  11. Thank you.

    This is really great news. Looking forward to test it in master.
  12. Just trying it now ;-)

    There is an error in the which is looking for the "noiseprofile" program in ~/vcs/darkatble/tools/noise and not ~/darkatble/tools/noise
    And you must run ./ to compile noiseprofile too !

    Great job !!
  13. José Carlos on Wed Dec 12 23:09:52 2012:
    I also had to install gnuplot.

    As the source code is too much for me, I sent the files to the mailing list. (Olympus E-PL1)
  14. see `future work' section, i think you can interpolate a and b for every iso setting (take the values for 100 and 200 and use the average for 150). didn't try it though, but the data seems to indicate it makes sense (modulo weird stuff that happens at certain iso values for certain cameras).

    currently i'm only using a[1] and b[1] (green channel) and derive the others runtime via whitebalance settings (which makes some assumptions about white light and reasonable whitebalance i'm not quite sure about).

    [edit: i think you meant same iso and average the values (a, b). yes, that makes sense too. even though the MAD statistic is robust, it comes with variance that you can take down using several shots]
  15. the light should be reasonably white, so that some information ends up in the other two channels at all. the rest will be done by whitebalance, which i conspire with to derive a joint what-is-white-and-how-does-my-cfa-look factor which is used to create these curves for the denoising module, currently. i still match the red and blue curves, just to verify at some point that what i do makes sense ;)

    underexposed could be a half covered lens, but an edge to a dark foreground object is better, as you're likely to get some gradient black->white out of it (and thus data in between, too).
  16. Hi,
    really great work.
  17. I read this post and it is just a little to technical for me to understand. Would somebody care to summarize what exactly this is about in layman's terms? Is it a bit like lens profiling for distortion only with cameras and ISO noise? Will this module essentially automagically get rid of noise in photos once my camera has been profiled?
  18. My darktable doesn't have $(EXIF_ISO) as a thing for file export, so it's crashing Dt when I try and follow these instructions.
  19. are you sure you're on the denoise branch?
  20. yes.
  21. i'm going to try to setup some interpolation as a default for the module, so you'll get precise numbers for 1/3 iso as well (and only need to switch on the module, the promised single-click mode).
  22. This is offtopic, however, I would absolutely love if you could implement the paper:
    "Illumination Decomposition for Material Recoloring with Consistent Interreflections"
  23. Gary Thompson on Wed Dec 26 07:15:53 2012:
    Hi, I just downloaded the master GIT and compiled on Ubuntu 12.04 (Mythbuntu flavour).

    I'm very happy with the results and have performed a subjective comparison on two specific photos of mine that I've been unhappy with. I was considering buying Lightroom and had just tested it (which also meant I needed to set up a temporary Windows computer).

    Default Denoise (profiled) settings on my particular photos created noticeable colour banding and removed fine textures (people where wearing colourful costumes with similar colours, feathers etc...). Saying that, I compared to the photos produced by Lightroom and although they were slighyly better, also produced colour banding. I then went back and very quickly tweaked the patch size and strength settings and the results were excellent.

    Great work on the noise profiling!!!

    Before I e-mail the profile, the camera has built in noise handling for high ISO and long exposures under custom settings, should I keep these at default? Do we need profiles to take into account these settings or do they not impact the RAW images?

    Just a note on how I compiled:
    * I installed as per the instructions on this website. I might have needed to add the Darktable Release Plus PPA to get additional libraries but can't remember.
  24. On this post you say that once the profiling is finished, a file named presets.txt is generated and the presets are injected into the database.

    Ok, I wanted to profile my camera (Oly E30) so I did some shots from ISO 200 to 3200, copied the ORFs into the tools/noise folder and ran ./gen-profile -d .

    The script did his job, the file and the pdfs are generated but I cannot see any presets when I run /opt/darktable/bin/darktable --library ./library.db from tools/noise folder. I assume that the presets should be on the denoise (profiled) module, right?

    So, what am I doing wrong?

    Thanks !
  25. the presets are filtered based on exif data. if you're comfortable with the sqlite3 command line tool, you can query the presets table (where operation='denoiseprofile') to see if your data is actually there. a quick sanity check would be looking at presets.txt (if there are any lines at all, if the exif data in there is correct etc).

    did you read through the output of the script? maybe it was complaining about your shots being not over/underexposed enough and aborted.
  26. Yes I read the output and everything was fine. Even I opened the presets.txt and the lines are there. This is the last line of the file:

    {"E-30 iso 200", "OLYMPUS IMAGING CORP.", "E-30", 200, {1.0f, 1.0f, {1.25000544229574e-05, 7.02465111426353e-06, 1.02274762732865e-05}, {-9.87788730741878e-08, -3.9579366931526e-07, -1.55476386202853e-07}}},

    I notice that it ends with a comma; is it ok?

    I don't know anything about sqlite3...
  27. very interesting. our presets for olympus cameras so far have "Olympus" as the maker, not all capitals and not with the lengthy appendix you have there. can you try to replace that and see if it helps?

    sometimes there's some exif confusion between exiv2 and raw reading libraries.
  28. I profiled again and at the end of the process the scripts says:

    ===> Record preset in library for testing purpose
    make: Entering directory `/home/faktor/darktable/tools/noise'
    make: Nothing to be done for `all'.
    make: Leaving directory `/home/faktor/darktable/tools/noise'

    I also ran the query select * from presets where operation='denoiseprofile' with no results..
  29. wow....! just updated and now my presets are there !!
  30. supersubexy on Fri Dec 28 08:51:16 2012:
    Need help. I'm using the latest mint with cinnamon. I installed dt from the ppa. I can't find the tools directory anywhere, let alone the noise dir or the gen-profile. Should i reinstall using git?

    Thanks for the help. I really want to try this to extend my pentax k-x life. I was seriously thinking of spending money on newer, less noisy models before reading this.
  31. supersubsexy on Fri Dec 28 15:08:20 2012:
    Got it. After re-installing dt from git, not the PPA, everything works.

    Thank you so much, guys. It is simply amazing. Compared to the results I got from Rawtherapee and in-camera jpeg engine, images from dt after profiling has less luminance noise and somewhat sharper. Against the unprofiled denoise, the advantage is less obvious, but it's there. However, chroma noise is more noticable on dt's images. I hope this can be addressed in the future.

    Now, to whom do I send this noise profile file?
  32. send the presets to darktable-devel@lists.. or upload them somewhere where i can find them.

    re: luma/chroma. if you want to keep your luma noise, just switch blend mode to `color'. if you want to blend luma/chroma, you can create multiple instances of the module and blend separately with different blend modes.

    if you want a quick workaround against larger color blotches, try this style for the equalizer
  33. When I profiled my camera, I took one shot per ISO (same picture).

    If I take multiple shots per iso (different pictures), will the profiling be improved?
  34. Looks very cool. I just wanted to contribute by profiling my nex5 and nex7, but the raw images (*.ARW) neither for the nex5 nor for the nex 7 seem to be recognized. Directory with particular raw files is properly indicated with -d.
  35. hm, both cameras should be supported by darktable, since RawSpeed supports them:


    darktable is not able to import them or what do you mean by "not recognized"? Do you try to import from camera or from a folder? I suggest trying the latter one.
  36. I have test images for the NEX5R, NEX6 and NEX7 imported into my library for testing, which all seem fine.

    Are you sure you're using a recent version of Darktable (1.1.1 or git master).
  37. darktable itself works fine with the raw files of both cameras, but the gen-profile script doesn't accept them and tells me, that I have to "check existing profiling shots" although I correctly referenced to the folder containing the raws.

    i have done the following (ie for nex5):
    1) manually taken the 7 shots with different iso-settings
    2) copied shots from sd-card to ~/Pictures/tosafe/20121231.nex5.noiseprofile
    3) imported the folder into darktable (git 1.1+289~g32d9d47)
    4) applied the raw linear stile from the /darktable/tools/noise directory
    5) ~/darktable/tools/noise/gen-profile -d ~/Pictures/tosafe/20121231.nex5.noiseprofile

    I suppose there is a step missing, because 3) and 4) don't have any impact on the raws itself.

    The shots can be downloaded here:
  38. Update: Profiling script now works. I expected the packages to be necessary only for taking the shots with gen-profile. As mentioned in the manual after installation of gphoto2, exiv2 and gnuplot, the script is running fine.
  39. Walter Sonius on Thu Jan 3 13:24:03 2013:
    Ubuntu 12.04.1: darktable PPA release plus + compiled version in /opt

    Error: Unable to find camera in database: NIKON CORPORATION NIKON D2Xs

    It shows this error when starting the compiled version of darktable and selecting a nikon d2xs nef and enabling "noise profile" module.

    The Profile made for the Nikon D2Xs cannot be located in its Database since darktable thinks its name is "maker:NIKON CORPORATION" and "model:NIKON D2Xs" but the metadata in the nef files shows only "maker:NIKON" model:"D2Xs".

    It does not give this error for the Nikon D3 and its profile since the "maker" and "model" name "NIKON CORPORATION NIKON D3" is both the same in the database of darktable as the metadata .nef files from the camera.

    Both camera profiles has been made with the gen_profile script and gphoto2.

    Any fix for this?

    When tethering Unmount/Remove/Eject the Camera first from Finder/Nautilus otherwise it will show as locked!
  40. The generated pdf-plots look different to the ones illustrated in the tutorial above. When I apply the presets for profiled denoising in darktable, I get complete black results.

    Is there something wrong with my testshots? They can be found here:
  41. Walter Sonius on Thu Jan 3 15:24:44 2013:
    Cameras Profiling problems/succes:

    Canon 1Ds mark 3: Succesfull
    Profile generates ok, havent really tested the result yet.

    Nikon D2Xs: Almost succesfull
    Problem with database camera name see previous post.

    Nikon D3: Succesfull
    Profile generates ok, havent really tested the result yet.

    Hasselblad H4d-60: Problem
    Running the script ends as soon as it detects that all jpegs images are NOT over exposed, but in fact they are in Hasselblad Phocus / Gnome Preview. All generated jpegs from the .3fr file look the same no matter what iso and all Whites look clipped so it doesn't recognize overexposure but the 3fr files itself are really different and over exposed...

    Hasselblad CFV-39: Problem
    Almost the same problem as the H4D 3fr files. Running the script to make a profile works, but the profile itself is totally wrong. All develeped jpegs and Graph PDF files look the same no difference between iso 50-100-200-400-800. Just as the H4D files the 3fr raw files are really overexposed and look different for each iso level.

    PS: testing the "denoise profile" module
    Don't forget to start darktable from its "/opt/darktable/bin/darktable" location! Because the local darktable binary doesn't have the "denoise profile module".
  42. did you check whether those hasselblad images work in dt at all? maybe the whitepoints are off, such as for iso 50 for some cameras (where you just get a brighter image instead of lower noise).
  43. Oh. Wow. Thank you - it was already available (I was about to profile my D90). The results are visible and fantastic.
    Thank you :)
  44. In the process of doing this for my Sony RX100. When I try to create the reference HDR from iso125 shots (lowest real ISO), the resulting dng has some seriously freaky colors... is that expected? I looked and the result also had some modules enabled highlights, basecurve.. I disabled those, but still looked bad. Is this expected?
  45. if you're sure your images align perfectly (or, tripod-perfectly), could you upload these for me somewhere so i can have a look? if you can reproduce by creating a hdr from just one raw image even better.
  46. ben robbins on Sun Jan 13 02:31:04 2013:
    I generated presets for the Sony A850 and added them to library.db. They show up in the presets menu for the denoise module, but images taken with that camera are not matched automatically. If I select an A850 preset manually, the image just turns black. Any ideas?
  47. Walter Sonius on Sun Jan 13 14:04:44 2013:
    Hi Jo,

    Thank you for your time and patience and sorry for the late response. At this moment I'm uploading some Hasselblad h4d and cfv39 raw 3fr to 'wetransferbeta' and mail the download link to the developers list of darktable on as soon as it finishes.

    You or other developers might have a look how those files are opened and handeled by darktable to fix the profiling issue.

    The hasseblad files themself seem to be opened and edited by darktable but with standard kind of underexposing.

    The .tar.gz archive contains tripod shot raw under/overexposed images from both of the camera's in all isos. Its big 1.1GB.

    As soon as its online you receive a link.
  48. ben robbins on Sun Jan 13 15:35:01 2013:
    Also, here is a line from presets.txt:

    {"DSLR-A850 iso 100","SONY","DSLR-A850",100,{8.03274706694361e-06,3.05589202633716e-06,4.98677350288896e-06},{-5.92030857781154e-08,-6.70027773398613e-08,-7.26577465835965e-08}},
  49. As the sensor of Sony Nex-3 and Nex-5 are complete equal, would it be possible to copy the already included presets for the Nex-3, so that they are applied to images from the Nex-5? Currently I manual select the preset using shortcuts.

    Would it be possible to display the the selected preset? I think that would be advantageous to determine, whether the presets for a specific model is already available or the standard settings are used, when the module is turned on.
  50. I get the same behavior out of most of the profiles I have generated for my RX100.

    A few of them do not turn the picture black, but also do not seem to have an effect on the noise. The generic function does reduce noise, however.
  51. i can't tell anything by the numbers, but usually that means the fit failed and your profiles are broken. check the pdfs and see if they match the example in the blog post above.
  52. I know next to nothing about denoising but this looks really impressive! The goal of the approach seems to be to remove all noise from the image. Would it also be possible to apply this noise model only to color information? My reason for asking is that it seems in the examples that the algorithm may be a bit overzealous in removing information. In the shadowy parts of the picture the fabric looks very smooth (e.g. at ISO 3200) whereas in the lighter parts there's still a sense of the texture of the material. I guess that if you could remove only color noise the result may look more pleasant while textures would still be intact.
  53. This is indeed possible (and Jo for example promotes this as nice solution) to only apply it to color and leave the rest untouched (or just remove part of it). You can use the blend mode "color" with an opacity of 100% for this. If you additionally want to remove some luma noise just add another instance of the module and pull down the overall opacity.
  54. I think jo meant this issue for discussion and workaround:
  55. I'm working on a color profile for my camera. The text says that whitebalance should be off but when I load the raw-linear style from tools/noise (git version from yesterday) darktable says that whitebalance is on.
  56. I'd like to contribute a profile for my camera (Fujifilm X100) but before I start I have two clarification questions about the procedure. You say "go to the directory you just exported your images to". So do I understand correctly: I take the pictures, one for each ISO, apply the raw-linear style in darktable, then export and run gen-profile on the exported images (not the RAWs). Usually people export JPGs but I wonder if the artifacts from the compression mess with the model fitting. So what is the recommended export format?

    Another question: the text says that whitebalance correction should be switched off ("this makes sure we get linear camera raw rgb with no basecurves, no whitebalance, etc in the way"). However, when I inspect the raw-linear stlye, darktable says that whitebalance is on under this style. I would switch it off manually but that seems impossible as there is no switch for that in the GUI.

  57. ben robbins on Sun Jan 20 22:25:42 2013:
    Hmm. I keep getting a conversion curves chart with blue stuck at 0. Would that be causing "broken profiles"? How can I diagnose this issue better?
  58. Hello,

    Love DT.

    I’m trying to generate presets for the Canon 1DMk.IV
    I have DT1.1.2 on OSX 10.6.8, took pictures and put them in the 1D4 folder.

    Then from the termianl:

    Li-Loos-Mac-Pro:1D4 liloo$ ~/darktable/tools/noise/gen-profile -d .
    -bash: /Users/liloo/darktable/tools/noise/gen-profile: No such file or directory

    BTW. No success with tethering with E-M5 and 5D3.