What is Foveon X3 Quattro and is it any good?

Conventional image sensors cannot perceive color, so have a pattern of colored filters placed in front of them. That way, the processing software knows which color (Red, Green or Blue) would have got through the filters to reach each photosite. Complex processing called demosaicing takes information from neighbouring photosites to interpolate the two missing color values for each position on the sensor. This processing usually generates one pixel for every photosite, even though the amount of captured color information was much lower than that.

A typical 24MP Bayer sensor will capture 24 million pixels of spatial resolution but only 12 million pieces of green information, 6 million pieces of red information and 6 million pieces of blue. On top of this, by filtering two of the three colors out, for each pixel, the filters end up absorbing around half of the light shone on the sensor. It sounds terrible but works remarkably well.

The original Foveon X3 design worked very differently. It layered three photosites on top of one another at each position on the sensor, capturing light at three different depths in the silicon. This took advantage of the fact that photons of different colors have different amounts of energy and more energetic light gets absorbed sooner. The promise was greater color resolution and, with no color filters, no light was lost to them. An undeniably brilliant idea.

Although Foveon's diagrams show different colors being detected at different layers, the truth is a little more complicated. Consequently, so is the computation required to distinguish which light was what color.

The problems were manifold: for a start, the sensors themselves weren't as efficient as their Bayer contemporaries (not helped by an entire industry pushing to optimize Bayer sensors, while just a single company worked on the X3 design). This meant that, while you weren't throwing away half of your light to a color filter, you weren't able to capture as high a proportion of that which was available. The much greater complexity of the chips also produced more electronic noise than the more simplistic Bayer pixels did.

Early Foveon sensors struggled to keep up with the megapixel race that Bayer sensors experienced, and struggled with color, dynamic range and noise. They produced excellent levels of detail, though.
Sigma DP2 | 1/100th | F8 | ISO 50
Photo by Richard Butler

Other problems stemmed from the processing that was needed. Although Foveon's diagrams show blue light being captured at the top layer, green in the middle and then red light at the lowest layer, the truth is more complex. Our understanding is that red light would be absorbed to varying degrees at all layers, green would be captured at both the top layer and the middle and blue would be captured in the top layer. Consequently, to get to a Red, Green, Blue image, you needed to understand the way red and green light were absorbed, then mathematically deduce how much of the top layer's signal had come from red light, and how much was from green, based on what had made its way further down into the sensor.

So, while you didn't have the blurring effect of demosaicing, you had a complex calculation step that amplified error. Or, to put it another way, generated noise. Add in problems with color accuracy that came from the color separation being based on the read-out depths of each layer, and it's fair to say that there was a price to be paid for the very crisp, detailed images that X3 could produce.

Foveon X3 Quattro

The X3 Quattro (the numbers here represent the APS-C version of this sensor), captures the lower two layers at lower resolution but with more light per photodiode, to improve noise.

The X3 Quattro design attempts to address many of these shortcomings, while retaining some of the original promise. It still functions on the same fundamental principle but it captures the image at lower resolutions in its two lower layers. This means you lose some of the color resolution benefits of the original design (that humans aren't great at perceiving) but with reduced silicon complexity and hence better noise performance.

So, while you still have some of the noise that comes from the complex mathematics involved, you're starting with cleaner, stronger signals for the lower layers and you're still capturing luminance (detail) information at every pixel.

It's hard to argue with the level of detail the X3 Quattro sensor can capture.

The result is 25 million pieces of luminance/blue information, 6 million pieces of green information and 6 million pieces of red information and no blurring of the luminance data by interpolating values from neighboring pixels. As you can see, in numerical terms alone, this is still more information being captured than a 24MP Bayer sensor.

The capture of comparable luminance data at every pixel and the lack of demosaicing means the sensor also lends itself well to mono capture, with the added bonus that you've also recorded color information if you wish to simulate a color filter or emphasize certain colors from the scene. The lower dependence on the noisy, red channel means it works better in low light, too.

The latest chip is also a newer generation of silicon design, meaning a further generation's improvement in read noise and further opportunity to fine-tune the spacing between capture depths to improve color separation.

To cap it off, the latest Sigmas are able to do all the complex deconvolution calculations onboard, then output the results as DNG files, greatly broadening the range of Raw conversion software that can be used, rather than leaving you to the delights of Sigma Photo Pro, a powerful but computationally challenging piece of software.

Is it any good?

Conceptually, at least, the idea of sacrificing some of the original Foveon's pixel-level detail in return for better noise performance might sound dangerous. After all, Bayer already does extremely well at striking a balance between luminance and color capture. However, the Quattro seems to suggest there is a middle path to be struck.

With a little care to protect the highlights, the Quattro H can capture a lot of detail and DNG support gives you access to familiar tools. This image has a gradient and minor localized adjustments that wouldn't be available in Sigma Photo Pro

It captures detail levels well beyond its nominal pixel count: Sigma's claims that it rivals 50MP Bayer sensors are absolutely true, in terms of detail capture. However, any hopes that it somehow rivals medium format are over-optimistic. The supposed 'medium format look' stems from its greater light capture and, while some full frame sensors can close some or all of this gap, it's a leap too far for an APS-H sensor, especially one with lower overall efficiency.

That said, the Quattro H doesn't cost anything like medium format money. Nor, for that matter, does it cost full frame money. So, while you don't magically get medium format image quality, you still get seriously impressive results for your cash. And, while there are still significant drawbacks in terms of processing latitude (and we would strongly recommend against leaving ISO 100), the SD Quattro's image quality is in many respects very attractive. Its way of rendering detail is as unique as its method for capturing it.