Looking for 3x1 samples (any camera)

[EvanMcG]

Hello,
I'd like to experiment around with 3x1 binning mode and desqueezing methods vs the more typical 1x3 binning. While it may seem like 1x3 binning would give a preferential result, it leaves horrendous stretch artefacts from object edges and small lines being stretched over three pixels.

3x1 will have the same issue vertically I suppose, but from sample imagery, it sometimes seems less so:

- Conflicting results here: https://www.magiclantern.fm/forum/index.php?topic=16516.25.
- A 3x1 vs 1x3 sample here: https://www.magiclantern.fm/forum/proxy.php?request=https%3A%2F%2Ffarm5.staticflickr.com%2F4825%2F32682240208_5f5a3c0415_o.png&hash=0631f7332d4edcc36621dba971252555, shows the 3x1 imagery on the right being vastly superior, needing only some antialiasing and a more advanced destretch algorithm to make it visually appealing).
- This footage which is 3x1 looks amazing with minimal aliasing: https://www.youtube.com/watch?v=l9u-nfAwaV8

Then again, there IS some good 1x3 footage out there that lacks the 'horizontal stetching' weirdness - I wonder what's different about this footage, for example: https://www.youtube.com/watch?v=3W52N9GVHqU

With that in mind, I was wondering if anyone had good quality (ideally >3.5k horizontal resolution when de-stretched) footage of any aspect ratio of 3x1 footage? Best case scenario would be a direct comparison of the same scene compared in 1x3 vs 3x1, but just 3x1 would be fine.

Thank you for any assistance!

[masc]

It does not make much sense using 3x1. All ML cams (but 5D3) do line skipping and column binning. The line skipping produces moiree and tons of artifacts. With 1x3 you can avoid this. With 3x1 you can't, because you through away the binning feature, but use the worse skipping.

[EvanMcG]

The third link I posted was shot on a 5D MkII and uses line skipping. It seems to look great if I'm honest.

I did also post a link of a more recent 1x3 binning video shot in 16:9 at 4.5K - the last link - but I'm not sure how they got around the muddiness issue that plagues 1x3. I'd love to know. Their video looks amazing, and has no clear horizontal artefacting. So 1x3 can look good but I'm not sure how they made it look passable.

I still believe that 3x1 could look good at a high resolution e.g. 4600x826, IF that video I sent is indeed using a 5dmkii in 3x1. But you know better than me, feel free to enlighten me here!

[reddeercity]

I have a sample for you , 3x1_5DMrk2.zip
Its a single frame Dng 4064x590 3x1 .
I'm the first one to implement 3x1 in crop_rec .mlv 10,12,14 bit uncompressed after A1ex of course in the 5D2 D4 cam
i write the code for 3x1 @24fps (4064x590) in my crop_rec for 5D2 D4 here are the builds 3K/UHD 5D2 Raw development and Other Digic IV Cams

In the Zip there is 1 dng(4064x590) ,2 Tiff created in After Effect with Camera Raw , 1 is the original size 4064x590 & the other is full size 4064x1770
3x1 is as good as 1x3 , its just how you used it ! it shines at close ups where sharp angles don't give artifacts , but if you do record strips etc. ....
then it breakdown a bit . No different then using 3x3 on all the cam ml supports (not including 5D3 it doesn't do line skipping)
because the 5D2 has a very big sensor 5.7k is does a superior job over the smaller 4.5-5k sensors , full frame is always better
and the color science is different then the D5 cameras 5D2 is more Log like . So color science play a part in 3x1 , anyway that my 2cents 
 
edit: its also how you process the file , i only use Adobe After Effects with Camera Raw processing , it the only way to go
with the 5D2 files, all the other apps don't reproduce the full range of colors , there always is something off like a green tone WB shift but not
in Adobe products ,

[EvanMcG]

Thanks! I was wondering if we could push it up to something like 4800x680 - or at a stretch, 5184x736 (=3,815,424 pixels, which is similar to some of Bilal's Crop Mood 1x3 presets)?

Going to try on a 700D at some point using custom crop_rec, if I can't get those resolutions to work, even 4.6k would be very nice. We're basically getting QHD images equivalent.
What I've noticed is that it is much harder to actually tell the vertically stretched images have less detail! I'm not quite sure why.

[EvanMcG]

Looked at the sample - it looks great! I wonder how far this could be pushed; at 1.85:1, we could probably get away with 4600x828, if we used 12/10 bit lossless.

At 2.35? Honestly, from a data rate perspective, it may be possible to do full sensor readout 5208x736 (=3,833,088px), but I'm not sure about computational intensity. It would look incredible, let's be honest here. A DCI 2K 2.39 aspect ratio file is 2048x858 - we're pretty much there vertically, and well over horizontally. I'd reckon it would be approximately equivalent in resolving power to 2880x1206 in 1:1 crop mode - a claim that will be interesting to test.

[Skinny]

Just my two cents - I experimented with 3x1 on 5d2 a lot (because there is no 1x3 preset for it) and honestly I still prefer regular 3x3..
Why? Because maybe 3x1 gives more details, but it also introduces a lot of artefacts. And they are really annoying..
It is almost exactly as illustrated here:

3x3 Readout 1736x1160


3x1 Readout 5208x1160 Upscaled to 5208x3478



So as you can see, while 3x3 is much less detailed, it is also more organic. And I prefer 3x3 in this case.. It is also much easier to use, no bugs and real-time preview, takes less card space and provides wider angle because 3x1 is still somewhat cropped..
This is why I prefer it although in theory 3x1 should be better.

[theBilalFakhouri]

Note: these pictures in Skinny reply were taken from my old post, and at that time these jagged lines were an issue in 3x1 Binning mode, if you follow the later replies you will find that we were able to fix it.
So this 3x1 image doesn't represent 3x1 actual quality (details).

I will try to capture new frames in 1:1, 1x3 and 3x1 modes tomorrow using my 700D with a static scene.

[Skinny]

It would be great to see some more recent tests.. I was constantly getting these green and red vertical stripes everywhere with 3x1 mode on 5D2, maybe it was not that bad as illustrated above but they were kind of annoying anyway


By the way I tried to switch the camera in 3x1 now, to make some samples, and can't do it for some reason.. Well, it switches but there is some kind of framerate problem, I spent maybe an hour right now already and it just doesn't work.. Who knows why. These modes are really unreliable..

[ShittyWebsite]

Is it possible with 5D3? i see anamorphic 1920x2300 @24fps and 1920x3760 @15fps
My question about it from January 03, 2022:
https://www.magiclantern.fm/forum/index.php?topic=23041.msg237597#msg237597

[Danne]

3x1 is worse than 3x3. It was never good enough for usage imo. Logically you get stronger aliasing than 3x3, jagged edges etc. Please prove me wrong  .

[theBilalFakhouri]

Here is new sample in 1:1, 1x3 and 3x1 modes (two frames for each MLV):
https://drive.google.com/drive/folders/1sxGvIITDLJt3Bp1tttwG9nlq02rxPaEx

To me, it seems 3x1 is sharper than 1x3? but 1x3 looks cleaner (due to vertical line skipping in 3x1),
both 1x3 and 3x1 has "stairs" artifacts due to stretching (it become more visible depends on objects in the scene).

It doesn't make sense to use 3x1 in models which does vertical pixel skipping (has a lot more aliasing compared to 1x3) IMO.
Will make try to make 3x1 preset on 5D3 in future and get some samples (it does vertical pixel binning instead of skipping).

Good luck!

[andy kh]

Here is new sample in 1:1, 1x3 and 3x1 modes (two frames for each MLV):
https://drive.google.com/drive/folders/1sxGvIITDLJt3Bp1tttwG9nlq02rxPaEx

To me, is seems 3x1 is sharper than 1x3? but 1x3 looks cleaner (due to vertical line skipping in 3x1)

yeah 3X1 is sharper than 1X3

[EvanMcG]

It would be great to see some more recent tests.. I was constantly getting these green and red vertical stripes everywhere with 3x1 mode on 5D2, maybe it was not that bad as illustrated above but they were kind of annoying anyway


By the way I tried to switch the camera in 3x1 now, to make some samples, and can't do it for some reason.. Well, it switches but there is some kind of framerate problem, I spent maybe an hour right now already and it just doesn't work.. Who knows why. These modes are really unreliable..

That's really interesting, because that seems to be the model that Reddeercity and a lot of other people filmed 4096x590 3x1 footage on. I wonder what modifications made it look so clean?

Also - thanks for the samples, Bilal! I'll play around with them. I appreciate the work you've done for the community, your builds are great!

[Walter Schulz]

Typo?
I hang around here a bit and "lot of people" using this mode I kind of fail to remember.

[Skinny]

If you are talking about this video, https://m.youtube.com/watch?v=l9u-nfAwaV8

It doesn't have a lot of things where you can evaluate the amount of artifacts after the vertical stretching...
And also he moves the camera like crazy and there is not enough time to even notice anything.

Just remember, 590 is a very small amount of pixels. So it is not ideal by any means..

[EvanMcG]

Some screenshots of Bilal's recorded MLV files with some chroma smoothing applied are below.

To serve as a comparison, in this album you can ALSO see what 1x1 looks like at a 60% sample, but crucially using pixel binning to prevent downscaling sharpness.

https://ibb.co/album/MBwDgB

There will never be competition with 1x1. It looks night and day better – double, triple as good – because it literally is 3x the number of pixels. An unobtainable pixel count.

On the other hand, continuous video at narrower aspect ratios than 3:2 would be theoretically possible using 3x1 or 1x3 at full sensor resolutions.

And it's pretty clear to me that the clarity scale goes 1x3 at the worst, then 3x1, and then lightyears out in front, 1x1.
I believe that if we clean it up with antialiasing and use the right settings for MLV capture / processing, we could possibly get this to look the same as a 60% resolution 1x1 file – e.g. 5120x2760 3x1 would be approximately equal in acuity to 3072x1656 1x1, while using a lower bitrate and full sensor width.



IMAGES:

1x3

3x1

1x1




(I'd appreciate it if someone could tell me how to shrink these massive images)

[Skinny]

So, you can see now these green and red pixels everywhere.. And what could be done to remove them? I don't even know, things like CA desaturation will work at extreme settings, at the same time significantly reducing the image quality...

[a.sintes]

maybe something like the Chromatic Aberration fixing tool of CameraRaw ?

[a.sintes]

After a quick test, it seems the Camera Raw's "Color Noise Reduction" module is clearly enough to quickly get rid of these red/green pixels when dealing with 1x1, without compromising the original image quality.

[Skinny]

I was talking about 3x1 image sample because this is a thread about 3x1 I believe?
 
1x1 looks good by default... Sorry if I wasn't clear enough

[a.sintes]

Ho ok! I personally don't get the point of these 1x3 / 3x1 / anamorphic things supposingly being more "cinematic" in the end, they all just look ugly to me compared to 1x1 :-(

[Skinny]

It's just because you can use more sensor area with them...

And it is just more convenient with your usual full frame lenses, so these modes are desirable.


I think they are not cinematic by themselves, (1x3 looks worse than 1x1) but the combination of some lenses and large sensor area could look more cinematic or just pleasant to one's eyes.

[a.sintes]

I get the point of using anamorphic lenses (to get specific bokeh, lens-flares...), but all I see is people recording with spherical lenses using ML anamorphic presets, which I don't really understand as the result after the unsqueezing process just look more blurry to me... maybe I'm missing something.

[vastunghia]

Totally aligned with a.sintes on this. By the way the wonderful1x1 3.5K preset has a 1.6x crop factor, so it even allows you to mount those cheap third-party APS-C lenses on your FF body (think of the Sigma 18-35mm f/1.8 Art for instance).

IMO 1x3 anamorphic is desirable only when you need to really maximize your field of view. That is, if you want a 24mm equivalent FOV, way better going for 1x1 3.5K with a 16mm lens (yielding ~25mm equivalent) than 1x3 "5.7K" with a native 24mm lens. But if you need to go as wide as 16mm, 1x3 anamorphic may be your only choice.