Monday, April 3rd, 2023 Posted by Jim Thacker

3Dpresso creates 3D models from videos of real objects

Originally posted on 3 April 2023, and updated with details of current pricing.

Korean start-up Recon Labs has launched 3Dpresso, a new AI-powered online service that generates 3D models from uploaded video clips of real-world objects.

The resulting models can be downloaded in OBJ or GLB format for use in DCC applications and game engines.

Augments photogrammetry with newer AI techniques to generate 3D models
Although there are a number of existing tools for reconstructing 3D models from photos of real-world objects, several of them available for free, 3Dpresso differs in a number of ways.

Firstly, its MetaRECON engine combines traditional photogrammetry with Neural Radiance Fields (NeRF) to generate and optimise a volumetric representation of a 3D object from a set of source images.

The engine generates “more accurate and stable reconstructions” than conventional tools, particularly for textureless or reflective surfaces that photogrammetry struggles to recreate accurately.

Secondly, 3Dpresso uses generative AI to create variant texture maps for the 3D models, with users able to enter plain text prompts to guide the look of the textures generated.

Generates a textured 3D model in OBJ and GLB format suitable for use in real-time work
The process of generating a 3D model with 3Dpresso begins by uploading a video of a 3D object captured on a smartphone or digital camera, moving around the object to show it from all sides.

It requires around a minute of video showing two complete 360-degree tours of the object, one captured face on, and the other at an angle to show the top of the object, as set out in this video guide.

Higher-resolution video results in better results – Recon Labs recommends uploading 4K footage – and the 3Dpresso YouTube channel has examples of suitable clips for reconstructing props, furniture and clothing.

The video is then processed on Recon Labs’ servers, which takes “less than 30 minutes” on average, with 3Dpresso sending out an email with a download link once processing is complete.

The resulting model can be downloaded in GLB or OBJ format with a 1,024 x 1,024px diffuse texture.

The quality looks suitable for use in real-time applications, including games and augmented reality projects, after some manual clean-up: the video embedded above shows the workflow inside Blender.

AI-based texturing tools and mobile app to come
3Dpresso also has an AI texturing feature, which enables users to generate variant texture maps by typing in plain-language prompts like ‘blue crystal’ or ‘hyperrealism’ to guide the visual style of the output.

Recon Labs also plans to launch a mobile app, making it possible to upload video directly from your phone.

Price and system requirements
3Dpresso is browser-based, so it should work with any standard desktop web browser.

Pricing is currently credit-based, with processing requests on faster servers, or downloading generated models or textures consuming ‘beans’.

Free accounts get 30 beans per month, enough for one video-to-3D download.

Lite subscriptions usually cost $12/month and include 150 beans/month; Pro subscriptions usually cost $28/month and include 500 beans/month. Find details of current pricing here.

The copyright policy gives Recon Labs a non-exclusive licence to use anything uploaded to 3Dpresso in its promotional material, so it isn’t suited to anything commercially sensitive.

Generate a 3D model from a video of an object with AI-based online service 3Dpresso