Photo to 3D Model: Turn One Photo into a GLB

A photo to 3D model workflow takes one photograph of an object and returns a textured 3D mesh you can rotate, download, and open in Blender, Unity, or a 3D printing slicer. The result depends far more on the photo than on any setting, so this page is about the photo: how to shoot it, what the model cannot recover from it, and what you get back.

Ready to try it? Open the AI 3D Generator, upload one photo, and run the default Pixal3D job. New accounts get 220 signup credits, which is enough for one default job (details in what it costs).

What you get from one photo

You upload a single photograph and the hosted Pixal3D lane returns a GLB file: a mesh plus PBR texture maps such as base color and roughness. You can preview it in the page and download it.

Pixal3D is an image-to-3D model from TencentARC. In the project's own words it "explicitly lifts pixel features into 3D through back-projection" instead of loosely injecting image features, which is why the side of the object that faces the camera stays close to your photograph (official repository). The part the camera never saw is different: it is inferred, not measured.

That is the key difference from photogrammetry. Photogrammetry stitches dozens of overlapping photos taken around an object and measures its shape. A one-photo workflow trades that effort for speed and accepts that the hidden side is an informed guess. If your photo is a flat drawing or render rather than a photograph, the 2D image to 3D model guide covers that input instead.

How to shoot a photo that converts well

Five habits decide most results. None of them needs a studio.

Frame one subject, whole and centered

Photograph a single object with a little empty margin on every side. The model can only place the pixels it is given, so a cropped corner, base, or handle stays cropped. Overlapping objects blur where one ends and the next begins.

Pick a three-quarter angle, slightly above

A straight-on side view hides the top; a straight-down view hides the sides. A three-quarter view from a little above shows the front, one side, and the top in the same frame, which gives the model the most surface to work from.

Use a plain, contrasting background

Put the object on a sheet of paper, a table, or a wall that differs in color from the object. A busy background competes with the subject, and hard cast shadows can end up painted into the texture. The upstream research code mattes images that have no alpha channel with a background-removal model first; this site does not document what the hosted endpoint does, so a clean background is the safer bet.

Light it softly and evenly

Window light on an overcast day or a lamp bounced off a white wall works well. Avoid a direct flash, glare spots, and mixed warm and cool lights: whatever lighting is in the photo tends to end up in the texture, because the model cannot tell paint from reflection.

Keep it sharp and large enough

Focus on the subject, hold the camera steady, and shoot at a resolution where the object fills a good part of the frame. Blur and heavy compression read as soft geometry and noisy texture. The 3D pixel guide lists twenty checks, including a rough 512 px floor on the subject's shortest side. The upload accepts PNG, JPEG, WebP, AVIF, and GIF files up to 10 MB.

What fails, and what to do instead

Single-photo reconstruction has known weak spots. The workspace itself warns that cropped, occluded, or low-light subjects still need manual cleanup, and that a second model can read thin parts differently. Plan around these:

Treat the GLB as a generated asset to inspect and clean up, not a measured scan. A quick pass in the free GLB Inspector shows mesh and material counts before you commit to a pipeline.

Photo subjects that work well

Products on a table, mugs, shoes, toys, tools, small appliances, figurines, and decor items tend to produce the cleanest results because they have solid surfaces, a readable silhouette, and a clear front. Pets and people are harder: they move, they have fur or hair, and they rarely hold still against a plain backdrop. A sleeping pet photographed from a three-quarter angle can give a usable base mesh, but expect to clean up the fur.

Photos of artwork on a screen, game screenshots, and renders behave more like flat art than like objects, so read the 2D image to 3D model guide for those.

From GLB to the format you need

The download is a GLB. Two free browser tools handle the common handoffs without uploading your model:

• GLB to OBJ for Blender, Maya, and other tools that prefer OBJ.
• GLB to STL for slicers and 3D printing checks. STL stores geometry only, so textures are dropped.

If you want a second interpretation of the same photo, the 3D AI guide compares the Pixal3D and Trellis 2 lanes, and both run from the same workspace.

What it costs

• Signup credits: every new account gets 220 one-time credits, valid for 30 days. The default Pixal3D job costs 220 credits, so the signup grant covers exactly one default job.
• Sign-in is required. If you choose a photo before you sign in, the workspace keeps it and starts the job after you sign in.
• Other settings cost more. The Pixal3D lane ranges from 180 to 390 credits depending on the settings you pick; the price is shown before you run.
• More credits: see the pricing page for packs and plans.

Who runs this

pixal3d.ai is an independent hosted service. It runs the open Pixal3D model through the Fal provider and adds accounts, credits, and tools around it. It is not the official Tencent research demo; the AI wrapper disclaimer explains the boundary.

Ready when your photo is: open the AI 3D Generator and upload it.