Can One Photo Become a 3D Model? — Canon’s Dual Pixel 3D and How Other Methods Compare

September 26, 2026

I own an EOS R5 Mark II. I mostly use it to shoot photos for this blog, and it has never had anything to do with 3D.

Then Canon announced a Windows app that builds a 3D model from a single photo. No need to shoot dozens of frames from every direction, and no special measuring equipment. People have been making 3D from photos for years, so what is actually different about this one?

This is research as of September 2026. The app becomes available on October 8, so I have not tried it yet. Everything below comes from reading official pages and papers, not from testing the software myself.

What I Checked — Scope and Sources

I looked at three things: what this technology actually does, whether it works with my own camera and lenses, and how it differs from other ways of turning photos into 3D.

For numbers and specs, I only used what is written on Canon’s official pages and in its news release. For the other companies’ technologies, I read their official documentation and the published papers. I did not use estimates from review articles.

What Is Dual Pixel 3D Converter?

It is a Windows PC app that Canon announced on September 16, 2026. It becomes available on October 8 and is free at launch. (Dates and terms below are from Canon’s Japanese announcement; availability in other regions may differ.)

It does two things. One turns the whole frame, background included, into 3D. The other cuts out just the main subject and makes a 3D model of it. Canon says the cut-out models can be used in 3D editors such as Blender, and even in PowerPoint.

You can also adjust the strength of the depth afterward, choosing between a look that exaggerates the relief and a more natural one.

Which Cameras and Lenses Does It Support?

This was the part I cared about most. The official page has a list.

Supported cameras (8 models)

Supported lenses (12)

Two things in this list surprised me.

First, the EOS R3 and EOS R1 are not on it. The original EOS R6, the EOS R, and the EOS RP are left out too. So it is not a case of “the higher-end the body, the more likely it works." Instead, the EOS R7 and EOS R10 — entry-level to mid-range APS-C bodies — are included. I dig into why below.

Second, only 12 lenses are supported. Even if your camera is on the list, you cannot use it unless the lens you own is on the list as well.

Why Are the EOS R3 and EOS R1 Left Out?

Canon does not say why. So I opened the “Menu contents: still photo shooting" page in Canon’s product manual for each model and checked whether a DPRAW setting exists. Dual Pixel 3D takes DPRAW as input, so I wanted to see if that is where the line is drawn.

All 11 models I could check lined up the same way. The dividing line is not how high-end the body is, but whether the camera can record DPRAW. A body that cannot record it cannot produce the input Dual Pixel 3D needs in the first place. That also reads as the reason the APS-C EOS R7 and EOS R10 made the list: both have a DPRAW setting.

Then why don’t the flagship EOS R3 and EOS R1 have DPRAW? Both are high-speed burst cameras with stacked sensors. Canon writes that shooting in DPRAW lowers the continuous shooting speed, reduces the maximum burst, and disables the electronic shutter, multiple exposure, and HDR shooting. It is a feature that costs speed, so it seems plausible that it was dropped from the two bodies built for speed above all. Canon does not explain it this way, though — this part is my own guess. If that reading is right, Canon’s future high-speed burst cameras should not get DPRAW either.

One model does not fit neatly: the original EOS R. Canon’s FAQ has a page on how to turn on the DPRAW function on the EOS R, so DPRAW is available. Yet it is not supported by Dual Pixel 3D. It is a 2018 model, and a separate line seems to have been drawn for it alone.

Why the Spec Sheet Alone Can Mislead You About Support

One thing tripped me up along the way. On Canon’s product spec pages, the “recording image type" line lists DPRAW for the EOS R7, EOS R8, and EOS R10 — but not for the EOS R5 Mark II. It shows only JPEG, HEIF, RAW, and C-RAW, the same as the EOS R3 and EOS R1.

Even so, the EOS R5 Mark II does have a DPRAW setting and is supported by Dual Pixel 3D. Canon’s firmware update notes also mention the condition “when DPRAW is set." The spec sheets are not written consistently across models, so judging support from one line of a spec sheet will lead you astray. If you want to check your own camera, opening the camera’s menu is more reliable than reading the spec sheet.

My gear is the EOS R5 Mark II with the RF24-105mm F4 L IS USM and the RF50mm F1.8 STM, and all of them are on the list. The camera, however, comes with a condition: firmware Ver.1.3.1 or later. That is the first thing to check before trying it.

An even newer Ver.1.3.2 came out on September 8, 2026, and its change notes say it “fixes a phenomenon where the camera may restart during shooting when DPRAW is set." Dual Pixel 3D depends on shooting in DPRAW, so I plan to update without waiting for October 8.

How Does It Get Depth From One Photo?

It uses Canon’s autofocus technology, Dual Pixel CMOS AF.

An image sensor is a grid of pixels. Each pixel contains two light-receiving parts (photodiodes). The images those two receive are very slightly offset, and from the size of that offset you can estimate how far away the subject is. This is the information the camera has always used to focus.

The key point here is that only what was in the frame when you shot becomes 3D. You took just one photo, so there is no information about the back side or anything hidden behind other objects. The official page also says areas in blind spots cannot be made into 3D, and the subject may come out stretched or distorted.

How Many Other Ways Are There to Make 3D From Photos?

This technology is not unique in itself. There have been several ways to make 3D from photos for a while, and they fall roughly into three groups.

The first builds from many photos. This is called photogrammetry. Apple’s Object Capture for macOS and iOS is one example. According to the official documentation, you feed it many photos taken from different angles, and it matches landmarks in the overlapping parts of the images to build a 3D model. It needs a Mac with a ray-tracing-capable GPU with 4GB or more, or certain iOS devices with LiDAR.

The second measures with a dedicated sensor. The LiDAR on some iPhones and iPads is the best-known case: it shines light and measures distance from how long the light takes to bounce back. That is direct distance measurement, not photo analysis.

The third reproduces appearance from many viewpoints. These are the techniques called NeRF and 3D Gaussian Splatting, and they also need many photos. Rather than building a shape, the idea is to compute how the scene looks from any given angle.

Did Single-Photo Methods Already Exist?

Methods that need only one photo, like this one, came earlier too. There are two lines of them.

One is having generative AI guess. TripoSR, released by Stability AI and Tripo AI in March 2024, is an example; its paper says it builds a 3D mesh from a single image in under 0.5 seconds. Even the unseen back side is filled in from what the model learned — “it probably looks like this."

The other derives depth from phase difference on the image sensor. And Canon was not the first to do this.

A 2019 paper by Google researchers says:

We estimate depth from a single camera by leveraging the dual-pixel auto-focus hardware that is increasingly common on modern camera sensors.
(from the abstract of arXiv:1904.05822)

That is exactly the same idea. What this research was aiming for was depth estimation to blur the background in smartphone portrait mode.

How Is It Different From the Earlier Methods?

Here they are side by side.

Laid out like this, Dual Pixel 3D’s position becomes clear.

Its first strength is that one photo is enough. You don’t have to walk around the subject shooting dozens of frames as with photogrammetry, and you don’t need measuring gear like LiDAR.

Next, the depth is not a guess. Among single-photo methods, generative AI fills in unseen parts from what it has learned. This one uses the phase-difference information the camera actually captured. An article on ASCII.jp, a Japanese tech news site, also described it as an approach only a camera maker could take.

And the surface texture is a real EOS photo. The photo you took is applied directly to the surface of the 3D model, so the original image quality carries straight over.

The difference from Google’s research is where it ends up. Google’s research stopped at a depth map — depth information for blurring the background. Canon carried it further, to a 3D model (OBJ / GLB) that other software can open. Same starting idea, different destination.

How Strict Are the Shooting Conditions?

On the other hand, something has been given up: freedom in how you shoot.

Two conditions are written not as recommendations but as “no 3D is generated": aperture and ISO. A photo shot stopped down in a dark place simply won’t be accepted.

The official page also lists subjects it handles poorly: those with heavy background blur (bokeh), many blind spots, thin shapes, or strong reflections.

That made me notice something. These overlap with what Apple’s photogrammetry struggles with. Apple’s official documentation says to avoid objects that are extremely thin in one direction, highly reflective, transparent, or single-colored and glossy. The way depth is obtained is completely different, yet the weak spots are similar.

Why are the conditions so detailed? The Google paper mentioned earlier offers one clue. The same paper says that depth estimated from dual-pixel cues has an inherent ambiguity. Narrowly specifying aperture and distance may be a way to stay within a range where that ambiguity is less likely to show up. Canon does not explain it this way, though — this part is my own guess.

Where Do You Open the 3D Model It Makes?

What comes out is an OBJ or GLB 3D model, or an MP4 video. OBJ can be read by Blender and most other 3D software. GLB is a format that packs shape and color into a single file, and it opens directly in web browsers and AR viewers. MP4 can go straight onto a blog or social media.

The uses Canon lists fall into two tiers. For individuals, you can make two kinds of 3D model — the whole scene including the background, or the main subject alone — and change how it looks by strengthening or softening the sense of depth. For businesses, the “Dual Pixel 3D SDK," planned for release within 2026, is meant to be built into photo-spot services, e-commerce, and advertising.

There is no official word on 3D printing. But since blind spots are not reproduced, the result should be a shape with relief only on the front side, not a full object you could walk around. A print would probably be closer to a relief sculpture. Whether it can be 3D printed is my guess, and I have not tried it. If this reading were wrong and a full all-around object came out, Canon would have no reason to write that “blind spots are not reproduced."

What I Haven’t Confirmed in This Article

• I have not tried it on real hardware. It becomes available on October 8, 2026, and this article was written before that by reading the official pages
• How long it takes to generate a 3D model is not stated on the official pages
• The polygon count and accuracy of the output 3D model are also not stated on the official pages
• No reason is given for C-RAW not being supported. Canon only says it is “not supported"
• For the other companies’ technologies in the comparison, I only read official documentation and papers; I have not run any of them
• The business-oriented “Dual Pixel 3D SDK" is planned for release within 2026, but Canon says its specs, such as recommended subjects, are different, and I don’t know what it contains
• The presence of a DPRAW setting was counted from the menu contents in the official manuals; the only body I checked in person is the EOS R5 Mark II. I have not been able to confirm the EOS R6 Mark III and EOS R8 Mark II
• There is no official explanation of why the EOS R supports DPRAW but is not supported

Summary — What Do You Give Up to Get Away With One Photo?

• Making 3D from photos is not new, and the methods fall roughly into three groups: building from many photos, measuring with a dedicated sensor, and reproducing appearance from many viewpoints
• Single-photo methods came first, too, in two lines: generative AI that guesses (TripoSR, 2024) and depth from phase difference on the image sensor (Google research, 2019)
• Dual Pixel 3D’s strengths are that one photo is enough, the depth is not a guess, and the surface is a real EOS photo. It takes what Google stopped at a depth map all the way to a 3D model
• The downsides are that support is limited to 8 camera bodies and 12 lenses, that it runs only on Windows 11 with no Mac version, and the strict shooting conditions. Stop down beyond F8 and nothing is generated; go above ISO3200 and nothing is generated either
• Blind spots cannot be made into 3D. You took only one photo, so there was never any information about the back

The convenience of “one photo is enough" seems to rest on a trade: you accept being told precisely how to shoot. That plants and food are the recommended subjects is probably because they make the conditions easy to meet.

My own camera and lenses are on the supported list. Once October 8 has passed, I plan to update the firmware and actually try it.

This article is research as of September 2026, based on reading official pages, the news release, and published papers. Specs may change after the app becomes available.

References

• Canon news release: launch of “Dual Pixel 3D Converter," which generates 3D models from photos (September 16, 2026, in Japanese)
• Dual Pixel 3D official site (supported models, shooting guide, details)
• ASCII.jp: Canon app generates 3D models from photos using EOS AF technology (in Japanese)
• Apple Developer: “Creating 3D objects from photographs" (RealityKit Object Capture)
• arXiv:1904.05822 “Learning Single Camera Depth Estimation using Dual-Pixels" (2019)
• arXiv:2403.02151 “TripoSR: Fast 3D Object Reconstruction from a Single Image" (2024)
• TripoSR model card (states it was jointly developed by Stability AI and Tripo AI)
• Canon product manual: “Menu contents: still photo shooting" (the same check was done for 11 models: EOS R5 / R5 Mark II / R6 / R6 Mark II / R7 / R8 / R10 / R3 / R1 / R50 / 1D X Mark III)
• Canon support: EOS R5 Mark II firmware Version 1.3.2 (September 8, 2026, in Japanese)