acerola wrote:It is interesting, that the second picture has more contrast. Is it tufuse creating better pictures with shorter stack or just post processing?
TuFuse creates better pictures with shorter stacks -- where "better" means that contrast/brightness are more like the in-focus pieces of the original frames.
Consider the limiting case: if you run a single frame through TuFuse, what comes out is just what went in. If you run two adjacent frames through TuFuse, what comes out has overall contrast and brightness very similar to what went in, but the details are merged. The more frames you run through at once, the more each pixel in the result is affected by overall differences in contrast & brightness due to the influence of varying OOF regions across all the frames. If you run through 50 or 100 frames at once, overall contrast will likely degrade while local contrast gets increased.
The depth-map and weighted-average algorithms currently used by CombineZM and Helicon Focus do not have this difficulty, since grossly OOF contributions are simply ignored.
This suggests that some sort of hybrid algorithm might work well -- use the multiresolution pyramid approach to merge details, but only from those images that plausibly contain details at each pixel position.
Developers may be working on this, but probably it will take a while to get the details right. After all, TuFuse traces its ancestry back to work done over 20 years ago.
But if you want to play with this idea in the meantime, you might consider something like this: Break your image into multiple overlapping slabs, say images 1-15, 8-22, 15-29, and so on. Run TuFuse on each slab independently. This will merge details within the slab without changing the overall brightness/contrast very much. Take the resulting images and feed them into CZ or HF to do the final merging, which will again be done without changing overall brightness & contrast very much.
This approach avoids some of the disadvantages of TuFuse, while at the same time retaining some of its important advantages. In particular many of the standard artifacts such as halo, lost bristles, and depth-map "swirly" effects can be significantly reduced or eliminated.
Shown below is an example of what can be made to come out. I choose these words carefully because I had to do quite a bit of playing around to get good parameter values. This is bleeding edge stuff and I don't recommend it yet, but I thought you'd be interested.
This is from TuFuse'ing groups of 15 frames with an overlap of 8 frames, then running CombineZM with a Do Stack macro modified to eliminate alignment, Find Detail 100,3,0, and skip the final sharpening and contrast adjustment. The result matches pretty closely what I see in the in-focus portions of individual images.
This image has had no manual editing, and post-processing is limited to levels adjustment and last-minute sharpening. I see some harsh halos around one antenna and around a few bristles near the top of the head. There is a small swirly transparency defect on one antenna, and a "flash" test shows that one bristle on a mandible got lost. But overall, this looks pretty good to me. ("Pretty good" is a moving target, of course. Four years ago this quality would have been jaw-dropping.
)
--Rik
