75 to 1: The 3D Conversion Revolution

How Technology Changed 2D to 3D Conversions

What once required a team of dozens of specialists can now be achieved by a fraction of that workforce, in considerably less time. And the technology is still improving.

Back in June 2023, I wrote an article comparing the 2D-to-3D conversion of a collection of Three Stooges short films in 2012 with the conversion of a feature-length film using EYEPOP-3D's nascent conversion techniques in 2023.

The comparison was quite remarkable.

The Three Stooges conversion, undertaken by the pioneering conversion company Legend3D, appeared to involve somewhere in the region of 70–100 credited individuals across the various conversion roles. By comparison, in 2023, I had been able to convert an entire feature-length film largely as a one-person operation, with some additional assistance from Alex and Phil for feedback and quality control.

That represented a dramatic change in how stereoscopic conversion could be approached.

Three years on, I think it is worth revisiting the subject.

The technology hasn't stood still. In fact, developments in AI-assisted depth estimation, image processing, inpainting and computational power have continued to improve both the speed and quality of the conversion process.

What was already a remarkable improvement in 2023 can now be achieved even more efficiently.

2012 – The era of large conversion teams

To understand just how much things have changed, it is worth looking back at the techniques used during the 3D cinema boom.

Companies such as Legend3D were pioneers in developing high-quality stereoscopic conversion for major Hollywood productions. Their work included titles such as Alice in Wonderland, Transformers: Dark of the Moon, The Avengers and Life of Pi.

The fundamental principles of conversion were much the same as those used today. The original 2D image had to be analysed, objects separated into different depth planes, depth maps created, and the resulting images manipulated to produce the left-eye and right-eye views. However, much of this work was performed manually. Artists had to identify objects, create mattes, establish depth relationships, construct depth maps and deal with the inevitable problems that occur when an image is displaced to create a second viewpoint. Every object needed attention. Hair, transparent materials, fine edges, moving subjects and complicated backgrounds could all present significant challenges. The process was labour-intensive, and even relatively short pieces of footage could require substantial teams.

Looking at the credits for Legend3D's conversion of the Three Stooges shorts, I was surprised to find approximately 70 or more people associated with the conversion work.

This is not necessarily a precise headcount of artists working simultaneously on the project, but it gives an indication of the scale of the operation. At the time, this was understandable. High-quality conversion required a considerable amount of skilled manual intervention.

2023 – A very different approach

By 2023, things had changed considerably.

When I converted Nightmare Castle, I was able to undertake the majority of the conversion work single handedly, using a somewhat different approach supported by early, and fairly rudimentary AI depth mapping technology.

Instead of a large team of artists manually constructing every depth map, I was able to use AI-assisted depth estimation as a starting point. This did not eliminate the need for artistic intervention. Far from it. AI-generated depth information is not always accurate, and was even less so in 2023. It can misunderstand objects, produce inconsistent depth around edges and struggle with complicated visual relationships.

However, it provided a substantial head start.

Rather than creating everything manually, the process became one of generating, reviewing, correcting and refining. This was a significant change in productivity.

The fundamental conversion techniques were still recognisable from those used in 2012, but the amount of manual work required to achieve a result had been reduced considerably.

I also felt that Nightmare Castle 3D compared reasonably well with the Three Stooges conversion, particularly in the naturalness of its depth and the level of stereoscopic detail. Of course, that is a subjective assessment, but it demonstrated to me that a much smaller operation could produce results that stood up well against work created by a much larger team.

2026 – The next generation of conversion technology

So what has changed since 2023?

Quite a lot.

The most significant developments have been in the tools available to assist with the most time-consuming aspects of conversion.

1. AI depth estimation has improved

Depth estimation is arguably the foundation of modern 2D-to-3D conversion.

The original image needs to be interpreted in terms of the relative distance of objects from the camera. This information is then used to determine how much each part of the image should be displaced when creating the second eye.

Earlier AI depth estimation systems were useful, but could produce inconsistent results, particularly when dealing with complicated scenes, similar-looking objects or moving subjects.

Newer models have become considerably more capable.

They can provide more convincing depth information, better distinguish foreground and background elements, and offer improved results across a wider variety of imagery.

Video-based depth estimation is also becoming increasingly important because it considers information across multiple frames rather than treating every frame as an entirely separate image.

This can help reduce the distracting changes in depth that can occur when individual frames are processed independently.

The practical benefit is that less time needs to be spent correcting basic depth information before the more creative work begins.

2. Better object recognition and segmentation

Separating objects from their backgrounds is another major part of conversion.

In the past, creating accurate mattes and rotoscoping complicated objects could consume a significant amount of time.

AI-assisted segmentation and object recognition have continued to improve, making it easier to identify and isolate individual elements for clean plate creation and basic inpainting tasks.

This doesn't mean that every object can be extracted perfectly at the press of a button, but it can significantly reduce the amount of manual preparation required. For a conversion artist, that means more time can be spent refining the stereoscopic result rather than performing repetitive isolation work.

3. Improved inpainting and background reconstruction

One of the unavoidable problems with 2D-to-3D conversion is that moving an object to create a second viewpoint reveals areas that were hidden behind it in the original image. These missing areas need to be reconstructed. This is known as inpainting. It has traditionally been one of the more demanding aspects of conversion, particularly when dealing with complex backgrounds, moving subjects and detailed textures.

Improvements in image reconstruction and technology assisted inpainting are making this process considerably more manageable. Newer techniques can use information from surrounding frames to help reconstruct missing areas, rather than relying entirely on the information available in a single image.

Research published by Disney Research Studios in 2024 demonstrated significant advances in disparity-aware image warping, foreground-background compositing and background-aware inpainting.

Their approach also incorporated information from neighbouring video frames to improve temporal consistency.

This is important because a reconstruction that looks convincing in one frame is not necessarily convincing when played back as part of a moving sequence.

4. Better displacement and image processing

The actual creation of the stereoscopic image has also benefited from improvements in processing techniques.

Displacement, compositing and image reconstruction can now be handled using increasingly sophisticated tools.

At EYEPOP-3D, we have also developed our own bespoke displacement and inpainting systems to improve the conversion process.

Our custom-developed application, Cyclopz, is used alongside established conversion tools to provide additional control and improve the quality of the resulting stereoscopic images.

The advantage of developing our own tools is that we can address specific problems encountered during conversion rather than being restricted entirely to the functionality of off-the-shelf software.

This is an important part of the evolution of our workflow.

5. Processing power and workflow efficiency

We shouldn't overlook the improvements in computer hardware.

Modern GPUs, faster memory and increasingly capable image-processing systems have made it possible to handle demanding operations more efficiently.

Tasks that once required substantial processing time can now be completed much faster.

But I would argue that the biggest improvement isn't simply that computers are faster.

It is the combination of better algorithms, more capable AI models, improved software and more refined workflows.

The technology is doing more of the repetitive work, leaving the conversion artist to concentrate on the decisions that actually determine the quality of the final 3D.

Does this mean conversion is now fully automated?

No. And I think this is an important distinction.

There is an increasing number of automated 2D-to-3D solutions available, and some can produce surprisingly impressive results.

However, there is a considerable difference between generating a stereoscopic image and creating a professionally finished stereoscopic film.

AI can estimate depth, but it doesn't necessarily understand how depth should be used to support the storytelling.

It can identify objects, but it doesn't always understand the relationship between those objects.

It can reconstruct missing image information, but that doesn't guarantee that the reconstructed material will remain convincing throughout a moving sequence.

There are also artistic considerations.

How much depth should a particular shot have? Should the foreground subject sit close to the screen or further back? How should the depth transition between shots? Should a particular object be allowed to pop out of the screen, or would that distract from the composition?

These are decisions that require an understanding of stereoscopic photography and visual storytelling.

The technology is becoming better at doing the technical groundwork, but the artist remains responsible for the result.

In fact, as the technology becomes more capable, I would argue that the importance of artistic judgement becomes even greater.

So how much faster is it now?

This is where things become particularly interesting.

In 2023, the ability to convert a feature-length film with a small team was already a considerable improvement on the large-team approach of 2012. But it took a long, long time. Nightmare Castle which seemed relatively impressive at the time took nearly six months of manual tweaking and fixing to create what would generally be considered a mild 3d conversion.

Since then, the workflow has simply become more & more efficient.

The improvements in depth estimation, object separation, displacement, inpainting and processing mean that many of the repetitive tasks can be completed more quickly, with fewer manual corrections required.

Consequently, a project that might have taken 6 months in 2023 can now be approached much more efficiently.

The exact saving depends on the source material. A film with straightforward compositions, relatively static backgrounds and clear subject separation will naturally be quicker to convert than one containing complex action, fine detail, smoke, transparency and frequent occlusion.

I wouldn't suggest that every film can be converted in a fraction of the time, or that the same workflow is suitable for every production.

But the overall direction is clear: the amount of time required to produce high-quality stereoscopic conversion continues to fall.

And importantly, this is not necessarily at the expense of quality. In fact quite the opposite. The modern pipeline deployed by EYEPOP-3D has consistently improved to the level we have just delivered our first feature film conversion for theatrical release in a little over 6 weeks.

The real opportunity

For me, the most exciting aspect of these developments is what they mean for the future of 3D film production for smaller, more creatively led indie film projects.

During the original 3D cinema boom, high-quality conversion was an expensive undertaking, generally reserved for major studio productions with substantial budgets.

The technology and the economics have changed.

This creates serious opportunities for independent filmmakers, rights holders and distributors who might previously have considered 3D conversion financially impractical to create premium quality stereo 3d experiences within indie level budgets.

It also opens up the possibility of revisiting existing film catalogues and creating new stereoscopic releases for theatrical, physical media and emerging VR and AR platforms.

The technology is still evolving, and I suspect we have not yet seen the full extent of what will become possible.

From 75 people to one – and still evolving

Looking back at the Three Stooges conversion from 2012, and comparing it with what was possible in 2023, was a useful reminder of just how quickly the industry has changed.

Revisiting that comparison in 2026 makes the progress even more apparent.

At EYEPOP-3D, we continue to develop our own tools and refine our workflow to take advantage of these advances.

Our objective remains the same: to make high-quality 2D-to-3D conversion as efficient and cost-effective as possible, without compromising the artistic and technical standards that make a convincing stereoscopic film.

The technology has moved on considerably since 2012. And we're not finished yet.

References:

• Disney Research Studios (2024): Stereo Conversion with Disparity-Aware Warping, Compositing and Inpainting

  
  

• M2SVid (2026): End-to-End Inpainting and Refinement for Monocular-to-Stereo Video Conversion