Air 3S photo quality is... disappointing

[Triggus]

I don't do pixel peeping. I shoot into the light A LOT, so shadow latitude is very important to me - that is what I was looking for. The observations about mushiness and purple fringing were an accidental discovery.

No amount of stitching will overcome poor performance in shadow performance. And the Air3S seems to be struggling in that department. I just don't understand how DJI can claim improved DR based on the images I have seen so far.

I would love to get my hands on more DNGs shot in contrasty situations, especially into the sun. I am surprised at the relative dearth.

Just a quick question, have you comapred both 50 mp and 12 mp modes ?
As its the 50mp that suffers terribly with noise and potential artifacts.
12 mp shots on 24 mm do not look much worse or possibly quite the same as air2s in terms of shadows and noise - just different aspect ratio of 4:2 instead of the "standard" 3:2 and lower pixel count as its 12 mp vs 20 mp (so possibly some minor details loss also when pixel peeping)
Still disappointing despite what the other guys here in the topic said about too high expecations etc ... the complaints arent about its not being better than mavic 3 pro, they are about its not being better then the predecessor air2s ..two upgrades later.. and the marketing gimmick of 50 mp they selling us adds to the confusions.

 

[BingBong]

Just a quick question, have you comapred both 50 mp and 12 mp modes ?
As its the 50mp that suffers terribly with noise and potential artifacts.
12 mp shots on 24 mm do not look much worse or possibly quite the same as air2s in terms of shadows and noise - just different aspect ratio of 4:2 instead of the "standard" 3:2 and lower pixel count as its 12 mp vs 20 mp (so possibly some minor details loss also when pixel peeping)
Still disappointing despite what the other guys here in the topic said about too high expecations etc ... the complaints arent about its not being better than mavic 3 pro, they are about its not being better then the predecessor air2s ..two upgrades later.. and the marketing gimmick of 50 mp they selling us adds to the confusions.

I was all set to get the Air 3S, but have got hold of some 12MP and 50MP JPG and DNGs. The 12MP shots seem to lack the detail you'd expect from 50MP pixel-binned to 12. The 50MP shots are sharp but noisy, and sometimes have colour artifacts like fringing and moire as you'd expect with less colour resolution from the Quad Bayer sensor. The Air 3S 24mm camera seems quite prone to a hard blob of flare into the light, much more than the Air 2S. I'm sticking with my Air 2S for now, as its native 20MP 3:2 sensor is still better for stills, which are my main interest. It's disappointing that all recent DJI drones have gone for 4:3 sensors. I think 3:2 works better for most single wide shots without the hassle of stitching.

 

[GroovyGeek]

I was all set to get the Air 3S, but have got hold of some 12MP and 50MP JPG and DNGs. The 12MP shots seem to lack the detail you'd expect from 50MP pixel-binned to 12. The 50MP shots are sharp but noisy, and sometimes have colour artifacts like fringing and moire as you'd expect with less colour resolution from the Quad Bayer sensor. The Air 3S 24mm camera seems quite prone to a hard blob of flare into the light, much more than the Air 2S. I'm sticking with my Air 2S for now, as its native 20MP 3:2 sensor is still better for stills, which are my main interest. It's disappointing that all recent DJI drones have gone for 4:3 sensors. I think 3:2 works better for most single wide shots without the hassle of stitching.

My exact observations from the files in the first link I posted in this thread. The binning reduces the noise as you would expect, but the loss of detail is - to me at least - unacceptable.

I have to go back and take a closer look at the 70mm binned images, I don't recall seeing any in the links I have access to. To me at least free pano is helpful but not critical, if I am going to stitch it does not matter if I have to crop later - just shoot that 180 degree pano and crop later. Or just do a manual pano series like you would do with a stills camera.

 

[RUFFSTUFF]

Any camera, be it a drone or your cell phone, that uses "technology" to create higher resolution images is a gimmick at best. Native pixels is what matters.

 

[wco81]

Well people are paying over a thousand on some of these phones, in large part for the newest cameras on those phones.

They're selling in the tens or hundreds of millions.

Apple, Google and Samsung though do the processing to make sure they look as good as possible. People don't pixel-peep phone camera photos as much but people would notice if there were obvious color artifacts.

 

[GroovyGeek]

Any camera, be it a drone or your cell phone, that uses "technology" to create higher resolution images is a gimmick at best. Native pixels is what matters.

That is incorrect. If you have tons of tiny noisy pixels and enough compute power you can do some amazing things - choose focus points after the fact, apply a broken of your choosing, etc etc. Except for readout noise, which is usually small,here is nothing inherently different between the signal from 1 large pixel vs the sum of 4 small pixels. The mushiness in the 3S images likely comes from post processing corrections for distortion, diffraction, etc.

 

[Droning on and on...]

Any camera, be it a drone or your cell phone, that uses "technology" to create higher resolution images is a gimmick at best. Native pixels is what matters.

What is a "native pixel" so I can understand what you mean by "gimmick".

 

[Droning on and on...]

The widespread ignorance among so many that should know better is sad.

So many think Quad Bayer is some sort of "gimmick" where simple Bayer sensors produce faithful, precise detection of R, G, and B values at each pixel.

Every time you hear someone talking about a "true" or "native" image produced by a Bayer sensor, you know you're listening to someone that doesn't actually know what they're talking about, but is simply parroting wrong information they read on the internet, thinking they sound knowlegable.

 

[GroovyGeek]

The widespread ignorance among so many that should know better is sad.

So many think Quad Bayer is some sort of "gimmick" where simple Bayer sensors produce faithful, precise detection of R, G, and B values at each pixel.

Every time you hear someone talking about a "true" or "native" image produced by a Bayer sensor, you know you're listening to someone that doesn't actually know what they're talking about, but is simply parroting wrong information they read on the internet, thinking they sound knowlegable.

That is generally true about any color sensor. The fact that the sensor itself is color-blind, and any color it produces is a mathematical construct does not seem to widely appreciated.

But quad Bayer sensors are, in a sense, a marketing gimmick. At least their high MPx output is. I think earlier in this thread someone asked "how is it possible that a 48Mpx image is worse than a 12 mpx". Camera makers are now playing the MPx game the same way CPU vendors played the MHz game in the early 2000s.

 

[Droning on and on...]

That is generally true about any color sensor. The fact that the sensor itself is color-blind, and any color it produces is a mathematical construct does not seem to widely appreciated.

But quad Bayer sensors are, in a sense, a marketing gimmick. At least their high MPx output is.

Please explain. How is a QB sensor any more of a "gimmick" than a standard bayer sensor? Do you believe a 12MP simple bayer capture produces a 12MP image?

Or are there color errors in 2 channels at every pixel, just as with a QB capture, and the missing two channels are computed?

I'm objecting to the idea that QB is some sort of phony gimmick, while Bayer sensors are accurate and true. This is completely mistaken, and this opinion exposes an ignorance about sensor technology.

Functionally, Bayer and Quad Bayer are essentially the same. Every pixel accurately captures the intensity of one of red, green, or blue at that location. The other two channels are interpolated from neighbor pixels that captured that color in their location.

As such, the accuracy, fidelity, and resolution is highly dependent on computational algorithm. What is the resolution of a 12MP bayer capture? Less than 12MP. What is the resolution of a 48MP QB capture? Less than 48MP.

The QB filter pattern results in larger errors than Bayer using the same interpolation algorithm, because the neighbor pixels are further away. However "greater error" by no means renders it a "gimmick", and using more sophisticated and complex demosaicing algorithms can reduce those errors close to a Bayer capture.

A drone imaging SoC doesn't have the horsepower (yet) to do this, so we get more color errors out of the drone for 48MP captures.

If resolution is important, take 48MP raw, and demosaic in post with a sophisticated QB algorithm. Stunning results most of the time that blows 12MP out of the water.

 

[GroovyGeek]

Please explain. How is a QB sensor any more of a "gimmick" than a standard bayer sensor? Do you believe a 12MP simple bayer capture produces a 12MP image?

Or are there color errors in 2 channels at every pixel, just as with a QB capture, and the missing two channels are computed?

I'm objecting to the idea that QB is some sort of phony gimmick, while Bayer sensors are accurate and true. This is completely mistaken, and this opinion exposes an ignorance about sensor technology.

Functionally, Bayer and Quad Bayer are essentially the same. Every pixel accurately captures the intensity of one of red, green, or blue at that location. The other two channels are interpolated from neighbor pixels that captured that color in their location.

As such, the accuracy, fidelity, and resolution is highly dependent on computational algorithm. What is the resolution of a 12MP bayer capture? Less than 12MP. What is the resolution of a 48MP QB capture? Less than 48MP.

The QB filter pattern results in larger errors than Bayer using the same interpolation algorithm, because the neighbor pixels are further away. However "greater error" by no means renders it a "gimmick", and using more sophisticated and complex demosaicing algorithms can reduce those errors close to a Bayer capture.

A drone imaging SoC doesn't have the horsepower (yet) to do this, so we get more color errors out of the drone for 48MP captures.

If resolution is important, take 48MP raw, and demosaic in post with a sophisticated QB algorithm. Stunning results most of the time that blows 12MP out of the water.

Agree on the color and demosaicing algorithm. Don't agree on the acuity. Tiny pixels suffer from diffraction effects much sooner than large pixels.

It is a marketing gimmick in the sense that the noise and dynamic range of the 48Mpx sensor is notably worse than that of the "12 Mpx equivalent ", and the Mpx count is used in a way that intentionally obfuscates that fact and is deceiving to those who are unaware of the tradeoffs. Just like Intel in the Pentium 4 era was obfuscating their product deficiencies relative to AMD by promoting GHz, quad sensors these days promote Mpx over other other metrics that matter more to the vast majority of users.

Specifically for the case of the Air3S, the 48 Mox images also suffer from HORRIBLE color fringing, not sure if that is an optics defect or something in the post processing algorithm. Last but not least, the AI denoise algorithm in PS is quite ineffective on Air3S images relative to Air2S and Mavic 3 pro.

 

[GroovyGeek]

Here are some comparisons from the links I posted in #1 and #2. All images have run through Photoshop's AI denoising and the shadows have been brought up 1 stop. There has been no other processing outside of what the PS profile may be doing.

First set is Air 3S





Second set is Mavic 3 Pro






The Mavic 3 Pro outperforms in every possible metric - shadow noise, color noise, acuity, color fringing, color fidelity. These are images taken in broad daylight, which makes the Air3S particularly disappointing and not on par with most modern mid-range cell phones.

Yes, the M3P a drone that costs 3x more. But you don't see that in still cameras, where the performance of the sensor in the $2000 camera is, to first order, indistinguishable from that of the $6000 camera. The differences are mostly in frame rates, buffer sizes, AF capabilities, not in "core" sensor performance.

The Air 3S may or may not be a fine video drone, I don't do video and can't express an informed opinion. It is a poor choice for those whose primary focus is stills.

 

[Occams Razor]

Here are some comparisons from the links I posted in #1 and #2. All images have run through Photoshop's AI denoising and the shadows have been brought up 1 stop. There has been no other processing outside of what the PS profile may be doing.

First set is Air 3S

View attachment 180646
View attachment 180647
View attachment 180648

Second set is Mavic 3 Pro
View attachment 180649

View attachment 180650

View attachment 180651

The Mavic 3 Pro outperforms in every possible metric - shadow noise, color noise, acuity, color fringing, color fidelity. These are images taken in broad daylight, which makes the Air3S particularly disappointing and not on par with most modern mid-range cell phones.

Yes, the M3P a drone that costs 3x more. But you don't see that in still cameras, where the performance of the sensor in the $2000 camera is, to first order, indistinguishable from that of the $6000 camera. The differences are mostly in frame rates, buffer sizes, AF capabilities, not in "core" sensor performance.

The Air 3S may or may not be a fine video drone, I don't do video and can't express an informed opinion. It is a poor choice for those whose primary focus is stills.

The Mavic 3 Pro outperforms in every possible metric - shadow noise, color noise, acuity, color fringing, color fidelity. These are images taken in broad daylight, which makes the Air3S particularly disappointing and not on par with most modern mid-range cell phones.

This is comparing a 12MP 1" sensor with a 20MP M43 sensor. A higher-resolution, larger sensor will always outperform a lower-resolution, smaller sensor assuming similar sensor and processing technologies. A drone is a tool that should be chosen based on your shooting requirements and budget. If IQ is your priority then I would buy a drone with the best sensor that you can afford. I travel quite a bit so one of my priorities is smaller drone size/weight and IQ is a lower priority.

Yes, the M3P a drone that costs 3x more. But you don't see that in still cameras, where the performance of the sensor in the $2000 camera is, to first order, indistinguishable from that of the $6000 camera.

You will see a difference in still cameras if different sensor sizes and resolutions are involved as in your above drone comparison. Your examples assumes a comparison between current generations of FF sensors where image quality differences are minimal. However, if the comparison was between a Fujifilm GFX 100S ($5,000) and an Olympus OM-1 II ($1800), you would see a significant difference in detail, noise, tonal gradation, etc. However, they are different tools with different purposes.

 

[Dakrisht]

Following this thread. Lots of good info.

Just picked up an Air3S and don’t have much more than 15 minutes of flight time on it.

I’ve shot on an M2Pro for years and found the image quality to be generally pretty stellar.

My reasons for upgrading to the Air3S were the 3x 70mm focal length camera + the ability to shoot 4K 120 for slow mo surf and sport photography.

I haven’t had enough time to shoot stills in either 12mp or 48mp modes but I’m wondering, after reading a ton of complaints so far in this thread, if the Air3S still IQ will be significantly inferior to the M2Pro at either the wide focal length or the 70mm 3x… Unless we’re nitpicking and comparing to M3Pro, which will certainly be better.

What does better mean… I assume the M3Pro will have better dynamic range, color data, sharpness and generally just more info available to store given the larger native sensor.

But I’m curious what M2Pro >> Air3S pilots are seeing for stills.

Video is probably better or the same. The added lens and 4K 120 are certainly upgrades.

 

[GroovyGeek]

The Mavic 3 Pro outperforms in every possible metric - shadow noise, color noise, acuity, color fringing, color fidelity. These are images taken in broad daylight, which makes the Air3S particularly disappointing and not on par with most modern mid-range cell phones.

This is comparing a 12MP 1" sensor with a 20MP M43 sensor. A higher-resolution, larger sensor will always outperform a lower-resolution, smaller sensor assuming similar sensor and processing technologies. A drone is a tool that should be chosen based on your shooting requirements and budget. If IQ is your priority then I would buy a drone with the best sensor that you can afford. I travel quite a bit so one of my priorities is smaller drone size/weight and IQ is a lower priority.

Yes, the M3P a drone that costs 3x more. But you don't see that in still cameras, where the performance of the sensor in the $2000 camera is, to first order, indistinguishable from that of the $6000 camera.

You will see a difference in still cameras if different sensor sizes and resolutions are involved as in your above drone comparison. Your examples assumes a comparison between current generations of FF sensors where image quality differences are minimal. However, if the comparison was between a Fujifilm GFX 100S ($5,000) and an Olympus OM-1 II ($1800), you would see a significant difference in detail, noise, tonal gradation, etc. However, they are different tools with different purposes.

No. Compare a modern APS-C to a FF sensor and the differences will be pretty mild.

Read Noise in DNs versus ISO Setting

That is far more difference than a M4/3 to 1". Furthermore the Air 2S is also 1" and it exhibits none of the issues that the Air 3S does. Not only that, but a 1" 20Mpx sensor from 3 years ago should be inferior to a 12Mpx 1" sensor today in terms of color noise, shadow noise, color fidelity, etc. That is absolutely not the case when comparing THESE IMAGES to what my Air 2S does.

It could all be software issues that get fixed, or it could be something I hererent to the hardware that is with us till the Air 4.

The examples I posted are dramatic. Yes, they are from a single set of photos that may be abnormal. Unfortunately I have not found any other DNGs from Air 3S.

You have a valid reason for preferring the Air 3S over the M3P. However the topic of this thread is image quality, not overall drone performance. Imo the Mini 3 pro has better IQ than the Air 3S. Even normalized for price the Air 3S underperforms... in my eyes.