*Measurement-grade only with a controlled capture, scale reference, and processing settings you have verified. A casual phone scan is not a metrology instrument.
Why this comparison matters now
Turning a photo into a 3D model used to mean one thing: photogrammetry. You walked around an object, took dozens or hundreds of overlapping photos, and let structure-from-motion software reconstruct camera positions and a point cloud. The result could be accurate, but the capture took time and the processing could take hours.
AI image-to-3D changed the input side of the equation. A single image, or a small set of views, now produces a textured mesh in seconds to minutes. The trade is straightforward: AI is faster and works with less input, but it invents the surfaces it cannot see. Photogrammetry is slower and needs a real object in front of the camera, but it records what is there instead of guessing.
Neither method replaced the other. In practice, the strongest 2026 workflow often uses both: photogrammetry or a scan for dimensional reference, AI generation for the visual asset, then cleanup in a 3D editor.
Photogrammetry vs AI image-to-3D at a glance
| Dimension | Photogrammetry | AI image-to-3D |
|---|---|---|
| Input | Many overlapping photos of a real object | One photo, or a few reference views |
| What it captures | Actual visible surfaces and texture | Visible surfaces plus inferred hidden geometry |
| Geometry accuracy | High when capture and scale are controlled | Low to medium; shapes are plausible, not measured |
| Hidden sides | Not captured if never photographed | Generated from learned priors |
| Capture effort | Minutes to hours; lighting and overlap matter | One upload; seconds |
| Processing time | Minutes to hours depending on photo count | Seconds to a few minutes |
| Texture quality | Faithful to the real surface | Convincing, but can invent or smooth detail |
| Reflective, clear, or textureless objects | Often fails or produces holes | Often handles better, still guesses |
| Output formats | Dense mesh, point cloud, textured mesh | GLB or mesh, sometimes STL after conversion |
| Best for | Real objects, measurements, documentation, heritage | Concept assets, game props, ecommerce visuals, single-photo requests |
The one-sentence rule
What photogrammetry actually does
Photogrammetry reconstructs geometry by finding the same feature points across multiple photographs. Structure-from-motion solves the camera positions, multi-view stereo densifies the point cloud, and a meshing step turns that cloud into a surface. Because every point is triangulated from real observations, the result is tied to the object you photographed.
The method has hard requirements. The object needs visible texture or geometric detail for feature matching. Lighting should be even and consistent between shots. Overlap should be high, typically 60–80% between adjacent frames. Glossy, transparent, reflective, or uniformly coloured surfaces break feature matching, which produces holes or drifting geometry. Moving objects, motion blur, and changing shadows also degrade the solve.
Scale is a separate step. A photogrammetry mesh is created in arbitrary units unless you place a known measurement in the scene or use a calibrated capture. That is why a scan can look perfect and still be the wrong size for a 3D print or a game engine import.
Where photogrammetry wins
You need dimensions you can trust
Fitted parts, jigs, replacement components, and any workflow where a millimetre matters.
The object has rich, stable surface detail
Carved stone, worn metal, fabric folds, and heritage objects where texture is part of the record.
You can photograph the whole object
A turntable or a walk-around capture with even light and generous overlap.
The deliverable is documentation, not decoration
Survey records, restoration baselines, insurance documentation, and museum archives.
Where photogrammetry disappoints
Photogrammetry cannot capture what the camera never sees. The bottom of a statue, the inside of a helmet, or the back of a wall-mounted object will be missing unless you move the object or the camera. It also struggles with surfaces that have no stable features: clear glass, polished chrome, black plastic, and plain white walls are common failure cases. And the processing step is not instant. A 200-photo capture can take long enough that it changes how you plan a project.
What AI image-to-3D actually does
AI image-to-3D models learn a prior over shape and appearance from large collections of 3D assets. Given an image, the model predicts a mesh, texture, and often a material set. Some pipelines use a single image; others accept multiple views to reduce ambiguity. The output is a generated object, not a measurement of the photographed object.
That distinction is the whole game. A single-image model has to invent the back, the bottom, and anything occluded in the photo. It usually produces a coherent, attractive result because it has seen many similar objects. It can also produce a confident result that is wrong: a chair with the wrong number of legs, a mug with no interior, a character whose back does not match the front. Multi-view input reduces the guessing, but it does not turn generation into measurement.
Where AI image-to-3D wins
You have one photo and no access to the object
Marketplace listings, archival images, concept art, and objects you cannot handle or re-photograph.
You need a usable asset today
Game props, blockouts, ecommerce previews, and social content where the model is a visual, not a survey.
The object is organic or stylised
Characters, creatures, food, plants, and decorative shapes tolerate plausible invention better than mechanical parts.
You want to iterate on shape language
Generate several variations from the same reference image and pick the one that reads best.
Where AI image-to-3D disappoints
AI is weakest when the job requires fidelity. Hidden geometry is a guess. Thin structures, hands, cables, and repeated patterns can collapse or fuse. Text and logos on the surface may be garbled. The mesh is often dense and unoptimised, with no clean topology for animation. And the model has no idea what units you intended, so scale is whatever the pipeline chooses until you set it.
Do not use AI output as a measurement
Cost comparison using published prices
Prices below were read from each vendor's public pricing page on 2026-10-22. They are list prices, not quotes, and plans change. The point is the shape of the cost, not a promise that one option is always cheaper.
| Option | Published price | What that buys |
|---|---|---|
| Polycam | Free $0; Basic $12.50/month billed annually ($150/year) or $30/month billed monthly; Business $33/month per user billed annually | Photogrammetry captures: the free plan allows up to 150 images per capture, paid plans up to 2,000 images |
| Meshy | Free $0 with 100 credits/month; Pro $20/month with 1,000 credits/month; Premium $40/month with 3,000 credits; Ultra $100/month with 8,000 credits | Hosted AI 3D generation, including image-to-3D; the free plan outputs are CC BY 4.0 with attribution, Pro adds a private licence |
| Tripo API | 1 credit = $0.01; image-to-3D costs 20 credits without texture or 30 credits with standard texture | Pay-as-you-go API generation, roughly $0.20–$0.30 per image-to-3D task before optional add-ons |
| Photo3D | Credit packs from $9.90 for 250 credits; hosted lanes cost 70 credits (TripoSR draft), 220 (default pixel-aligned), or 800 (Meshy 6 pass) | Hosted image-to-GLB runs with the lane price shown before submission, plus free browser tools that never spend credits |
The cost structures are not directly comparable. Polycam sells a capture workflow, Meshy and Tripo sell generation, and Photo3D sells hosted runs across several model lanes. A buyer comparing them should price the whole job: capture time, processing, cleanup, and the number of retries needed to get a usable asset.
Photo3D's published lane prices are useful because they make retries visible. A 70-credit TripoSR draft is a cheap way to test framing before committing to a 220-credit default run or an 800-credit Meshy 6 pass. That tiered structure is closer to how photogrammetry users already work: a quick low-quality solve first, then a full-resolution run once the capture is known to be good.
Capture and workflow compared
Photogrammetry workflow
- Plan the capture. Decide which sides of the object must be recorded and whether you can move the object or must move the camera.
- Photograph with overlap. Keep lighting consistent, avoid motion blur, and include a scale reference if dimensions matter.
- Mask or trim the background if the solver struggles to separate the object from its surroundings.
- Run structure-from-motion and dense reconstruction. Inspect the point cloud for drift before meshing.
- Clean the mesh: remove floaters, fill holes, decimate, and UV unwrap if the texture needs editing.
- Verify scale and export the format the next tool expects.
AI image-to-3D workflow
- Choose a clean source image. A single subject, even lighting, and a plain background produce better results than a busy scene.
- Pick the model lane and resolution. Start with the cheapest lane that can answer the question: is the silhouette and pose right?
- Generate, inspect the mesh, and check the hidden sides. Rotate the model before trusting it.
- Re-run at higher quality if the shape works. Change the source image rather than stacking prompts if the silhouette is wrong.
- Clean up in a 3D editor: fix intersections, decimate, and retopologise if the asset will be animated.
- Export GLB for real-time use, or convert to STL or OBJ for printing and downstream tools.
Which method for which deliverable
| Deliverable | Better starting point | Why |
|---|---|---|
| Replacement part or fitted component | Photogrammetry or 3D scanning | Dimensions must match the physical object |
| Game prop or environment asset | AI image-to-3D, then retopology | Speed and shape language matter more than measurement |
| Ecommerce product visual | AI image-to-3D from a clean product photo | Fast iteration and consistent presentation |
| Museum or heritage record | Photogrammetry with a scale reference | The record must be tied to the real surface |
| 3D print of a decorative figure | AI image-to-3D, then mesh repair | Organic shapes tolerate plausible hidden geometry |
| 3D print of a functional bracket | Photogrammetry plus CAD verification | A generated guess can fail under load |
| Character for animation | AI or photogrammetry, then retopology | Neither raw output has animation-ready topology |
File formats and 3D printing
Photogrammetry tools typically export dense meshes, point clouds, and textured models in formats such as OBJ, PLY, or GLB. AI image-to-3D tools usually deliver GLB with baked textures. Neither output is automatically printable. A slicer needs a watertight, manifold mesh with a defined scale, and both methods can produce holes, non-manifold edges, or interior geometry that must be repaired first.
For printing, convert the model to STL or 3MF after cleanup and check wall thickness and scale in the slicer. Photo3D's browser tools can handle the local conversion and inspection steps without uploading the file: use the STL viewer, the 3D format converter, and the print readiness checker to catch problems before slicing.
Privacy note
Who this guide is for — and who it is not for
This guide is for buyers deciding how to turn photos into a 3D asset: product marketers, game artists, 3D printing hobbyists, and small teams that need a repeatable pipeline without a metrology lab.
It is not for surveyors or engineers who need certified dimensional accuracy. Those workflows require calibrated equipment, documented procedures, and verification against known standards. No consumer photogrammetry app and no generative model replaces that.
Sources
- Polycam pricing and capture limits: https://poly.cam/pricing (accessed 2026-10-22)
- Meshy pricing and plan limits: https://www.meshy.ai/pricing (accessed 2026-10-22)
- Tripo API credit pricing: https://developers.tripo3d.ai/en/pricing (accessed 2026-10-22)
- Photo3D credit costs and plan prices: https://photo3d.org/pricing
Photogrammetry and AI image-to-3D are complementary tools. The mistake is choosing one for every job. Match the method to the deliverable, and verify anything that has to fit.