Here a zip with the node setup i tried (not really working, discard that or get only the useful things) and a sample video: deflicker test.zip (580 KB)
Not sure why Blender wont let pack the video inside the file…
In my case a deflicker node setup would be useful for timelapse photos; i used for some “experiments” virtualdub+msu deflicker but i was looking, if possible, to do something better/more flexible or comparable directly in Blender.
I tried to use a normalise function in the compositor but I couldn’t get a satisfactory result. The deflicker algorithms in vdub etc can do frame averaging, that is predictive sampling. Allowing them to build an exposure profile over time. Blender is not temporally aware at the moment so you don’t have access to frames in either direction.
Well, i tried with something like doing the average luminance/value of the next and previous three/four frames (similar to the nodes in the test example) but is not enough…i think also frame “weighting” is needed; i guess that theoretically can be done but would be very slow and relatively complicated compared to vdub+filter…
Btw the video sample was very “extreme”, in fact also vdub+filter was not able to do a good work…
Good on you for trying. I found that often the issue was with image profile, that is the baked in gamma and slope of exposure. I tried to standardise the brightness around this mid point then stretch from there. Alas the levels never really match up. Perhaps some sort of frame blending would help at the most extreme deviations?
