Microsoft turns 2D images into 3D models with Copilot 3D

Turning a single 2D photo into a functional 3D mesh used to mean hours in Blender or paying up for specialized AI services. Microsoft is now trying to make that workflow as simple as dragging an image into your browser.

While the company was rolling out OpenAI’s GPT-5 into Copilot, it quietly slipped a brand new experiment called Copilot 3D into Copilot Labs. The pitch is dead simple: give it a standard 2D image, skip the text prompts altogether, and get back an interactive 3D asset in a single click.

Great on Ikea tables, nightmare on dogs

The feature is currently in preview on the web version of Copilot. It is free for all users globally rather than locked behind Copilot Pro, though Microsoft is rolling it out to a limited subset of accounts first.

You can upload JPG or PNG files up to 10MB. Once converted, your models sit in a creations hub for 28 days and can be downloaded freely as GLB files. That makes them immediately compatible with common 3D software, game engines, and AR tools, or easy to convert into STL files for home 3D printing.

According to hands-on testing by The Verge, the system shines on inanimate, clearly defined objects. Clean product shots of Ikea furniture, beach balls, and bananas converted into solid models that could be dropped straight into AR view. Even when an umbrella initially confused the system, swapping in an image with more depth produced a near-perfect result, minus a stray baked-in shadow on the fabric.

Biological shapes, however, are a different story. In tests on a pet dog, the model attempted to recreate the animal's anatomy and inexplicably placed a penis onto the dog’s back. Reconstructing a human face technically worked, but the visual output looked straight out of a horror film.

Microsoft also has strict guardrails in place. Uploads of Tim Cook and Taylor Swift were flatly rejected. While characters like Mario slipped through, Microsoft warned that uploading unlicensed images or unconsented human faces can get accounts restricted.

Why it matters

We think dropping single-image 3D generation into a mainstream, zero-prompt interface is a clever land grab. By stripping out text prompts entirely, Microsoft bypasses the prompt-engineering hassle and meets creators right where they are: holding a reference image.

For hobbyists, 3D printing enthusiasts, and indie game devs, this offers an unbeatable entry price of zero dollars. It is not ready to replace seasoned 3D artists, but as a rapid prototyping engine for household items or basic props, it cuts down substantial setup time.

Across the industry, this puts heavy pressure on niche AI 3D startups that charge monthly subscriptions for basic image-to-mesh conversions. If a free generalist assistant can deliver decent GLB files in seconds, specialized tools will have to prove their worth on topology and clean rigging rather than raw geometry extraction.

The clear bottleneck is organic comprehension. Copilot 3D leans heavily on shadow heuristics and rigid shapes. The minute it tries to infer unseen sides of living things, hallucination sets in fast.