A browser-based image-to-3D service built on Meta's open-source SAM 3D models, letting you click an object or a person in one photo and export a textured mesh as GLB or PLY without a local GPU.
Used this tool? Rate it
Used this tool? Rate it
SAM 3D, published at sam3d.org, is a browser-based image-to-3D service that turns a single ordinary photograph into a textured three-dimensional mesh you can download and open in standard 3D software. You upload one RGB image, indicate which object or which person in it you care about, and the service returns geometry, texture and pose — including plausible surfaces on the side of the object the camera never saw. There is no scanning rig, no turntable, no multi-angle capture session and no photogrammetry pipeline to babysit.
The first thing to be clear about is provenance, because the naming invites confusion. This website is not a Meta property. The site's own FAQ area carries an explicit disclaimer that the project is independent and not affiliated with Meta. What it is instead is an independent operator hosting Meta's openly released SAM 3D model weights behind a commercial web interface, billing and viewer of its own. The pitch is deliberately narrow: turn any photo into detailed, realistic 3D assets, built on Meta's open-source SAM 3D models, with no setup and no GPU. That sentence describes the entire value proposition honestly — the intelligence comes from upstream research, and the product being sold is convenient access to it.
Understanding that split matters for every practical decision you will make with this tool. Reconstruction quality, failure modes and licence terms on the generated geometry trace back to the upstream models and their documented behaviour. Pricing, credit accounting, upload limits, retention practice and customer support are entirely the operator's own. When you evaluate the service, evaluate those two layers separately: the model decides whether your mesh is any good, and the operator decides what you pay and what happens to your uploads.
Two distinct reconstruction pipelines. SAM 3D Objects handles scene-aware reconstruction, letting you select any object in a photo and generate a pose-aware 3D mesh even when the scene is cluttered. SAM 3D Body is the separate human pipeline, aimed at recovering body pose and shape from a single frame. They are exposed as different entry points because they solve genuinely different problems — one predicts arbitrary object geometry, the other fits a parametric human representation.
Promptable selection instead of manual masking. Selection is promptable in three ways — text prompts, point clicks, or box selection — so you never have to hand-paint a mask. Typing "wooden stool" is a legitimate way to isolate a subject; so is clicking it once. This inherits directly from the Segment Anything lineage, where interactive, open-vocabulary segmentation is the core contribution.
Occluded and back-facing geometry is inferred, not left hollow. A single photograph contains no information about the back of an object, so the model hallucinates a plausible reverse side rather than producing a flat card or an open shell. The result is a closed, viewable-from-any-angle asset. It is invention rather than measurement, which is exactly the trade you are accepting when you use single-image reconstruction.
A human rig that separates skeleton from soft tissue. The body pipeline builds on Meta Momentum Human Rig, a representation that keeps skeletal structure and soft tissue shape as separable components. For anyone who needs to animate the result, that separation is far more useful than an undifferentiated blob of vertices, because pose and body shape can be edited independently.
Standard, engine-ready export formats. Object exports come in GLB for games and web applications, PLY point clouds for technical workflows, or a complete ZIP bundle with all formats included. The exported files are described as ready for immediate use in Unity, Unreal, Blender and other standard 3D software, which is the practical reason to prefer GLB unless a downstream tool specifically wants point data.
An in-browser viewer before you spend a download. Generated meshes can be orbited, zoomed and inspected in an interactive viewer inside the page. This matters more than it sounds: single-image reconstruction fails in ways that are obvious from a three-quarter view and invisible from the original camera angle, so inspecting before committing saves real time.
Auxiliary format viewers. Alongside generation, the site hosts standalone online viewers for common 3D interchange files including GLB, PLY, OBJ, FBX, GLTF, STL and DAE. These are peripheral utilities rather than the main product, but they are convenient when you want to check an asset without launching a heavy DCC application.
E-commerce product visualisation. A catalogue photograph is often the only asset a small retailer has, and commissioning a 3D scan per SKU is prohibitive. Reconstructing a rotatable model from that existing photo enables "view in room" style placement and 360-degree product spinners. Treat the output as a presentation asset rather than a dimensionally accurate replica — it will look right long before it measures right.
Rapid game and prototype blockout. Level designers frequently need a chair, a crate, a scooter or a street bollard that reads correctly in silhouette while the real art is still being made. Generating a GLB from a reference photo and dropping it into Unity or Unreal fills that slot in minutes. Many of these placeholder assets are later replaced, which makes generation speed far more valuable than topology quality.
Character reference and base meshes. The body pipeline turns a full-body reference photograph into a posed mesh that a character artist can use as proportional scaffolding. It substitutes for a sculpting blockout, not for a production character — but starting from an approximately correct human silhouette beats starting from a primitive sphere.
AR and VR scene dressing. Immersive scenes are hungry for props, and GLB is a first-class citizen in most WebXR and mobile AR pipelines. Photographing objects in the real environment you are recreating and reconstructing them individually produces a set that feels specific to that place rather than generically stock.
Education, archival and research visualisation. Museums, teachers and field researchers often possess photographs of specimens, artefacts or equipment they cannot physically re-scan. A rough three-dimensional model built from an archival image supports interactive teaching material and spatial discussion, provided the approximate nature of the geometry is stated plainly to the audience.
Computer vision experimentation without infrastructure. Researchers and engineers who want to see how these models behave on their own imagery, before deciding whether to self-host, can run a handful of representative images through the hosted service. Self-hosting the same open weights is documented as needing a Linux 64-bit environment and an NVIDIA GPU with at least 32 GB of VRAM, so a hosted trial is a cheap way to answer "is this worth provisioning hardware for" before you provision anything.
Choose the right pipeline for your subject. Send inanimate objects, furniture, products and scene props to the Objects entry point, and photographs of people to the Body entry point. Feeding a human into the object pipeline usually produces a stiff, statue-like mesh with none of the rig structure that makes a body model useful.
Prepare an input the model can actually read. Uploads accept JPEG, PNG and WebP files up to 10MB. Pick a frame where your subject is unobstructed, reasonably well lit, and large in frame. Because the reverse side is inferred rather than observed, a three-quarter view that reveals two faces of the object gives the model far more to work with than a flat, straight-on shot.
Isolate the subject. The documented object pipeline is a four-step loop: upload an image, select objects, generate the 3D model, then download it. At the selection stage, describe the object in words, click it, or draw a box around it, and confirm that the returned mask hugs the true silhouette. A mask that bleeds into the background or swallows an adjacent object will propagate that error into the geometry, so it is worth re-prompting here rather than fixing it later.
Generate and inspect from unfamiliar angles. Run the reconstruction, then immediately orbit the result in the viewer. Look specifically at the surface the camera never saw, at thin structures such as handles and legs, and at any region that was occluded in the source photo. These are where single-image reconstruction concentrates its errors.
Export in the format your destination wants. Choose GLB for game engines, real-time renderers and web viewers; choose PLY when a technical pipeline expects point data; take the ZIP bundle when you want to keep options open. For human reconstructions, GLB is the practical export target.
Clean up downstream before production use. Import into Blender or your DCC of choice and expect to do retopology, UV work, or decimation depending on where the asset is going. Treat the download as a strong starting block rather than a finished deliverable.
Photograph for reconstruction, not for aesthetics. If you control the camera, shoot the subject at a three-quarter angle, fully in frame, against a background that contrasts with it. Diffuse, even light is better than dramatic light — hard shadows get interpreted as geometry and baked into the texture.
Fix a bad mask at the mask stage. Almost every disappointing mesh traces back to a selection that was slightly wrong. Re-prompt with more specific wording, add a click, or tighten your box. Thirty seconds spent on the mask reliably saves a wasted generation.
Isolate rather than reconstruct whole scenes. Cropping to a single well-separated object generally outperforms feeding a dense composition and hoping the model resolves it. If you need a scene, reconstruct several objects individually and assemble them in your 3D application, where you keep full control over placement.
Budget generations, not just money. Because credits and segmentation requests are metered separately, a habit of segmenting several candidate selections and only generating from the best one uses your quota efficiently. Segmentation is the cheap step; generation is the expensive one.
Verify scale manually. Reconstruction from a single image carries no absolute scale reference. Once imported, set the size of at least one object against a known real-world dimension before you build anything around it.
Do not use the mesh as a measurement. For anything where dimensions carry consequences — fabrication, fitting, medical or safety-adjacent work — use a real measuring instrument. These models are visually plausible, which is not the same as being metrically correct.
Check the licensing chain before commercial release. Rights split into two layers: your right to use the service output, and your right to the content in your source photograph. A permissive upstream model licence does not grant you rights to someone else's product design, artwork or likeness.
Indie game developers and technical artists who need a steady supply of props and blockout geometry faster than a modelling pipeline can supply it, and who are comfortable cleaning up meshes afterwards.
E-commerce and product marketing teams sitting on large catalogues of flat product photography who want interactive, rotatable presentations without commissioning per-SKU 3D capture.
AR/VR and WebXR builders who need GLB props for immersive scenes and value turnaround speed over polygon-level craftsmanship.
3D generalists and hobbyists exploring what current single-image reconstruction can and cannot do, who would rather pay a small subscription than build a workstation for it.
Computer vision engineers and researchers evaluating the SAM 3D family on their own imagery before deciding whether to self-host the open weights.
Educators and content creators who want interactive three-dimensional illustrations built from images they already have, for lessons, explainers or video work.
Character artists looking for a proportionally sensible human base mesh from a reference photograph, as a starting point for sculpting rather than a finished character.
Prototypers and designers who need a quick spatial stand-in for an object during concept and layout work, where approximate is entirely sufficient.
Web browser, as the entire product. The service runs in the browser with generation performed server-side. There is no advertised desktop application, mobile app or browser extension, which is the intended design — the point of a hosted service is that the heavy computation is not on your machine.
Output that travels into standard 3D software. The exported files are described as ready for immediate use in Unity, Unreal, Blender and other standard 3D software. GLB in particular is broadly supported across engines, web viewers and modern DCC applications.
Format viewers hosted alongside generation. Online viewers are available for GLB, PLY, OBJ, FBX, GLTF, STL and DAE files, useful for quick inspection without opening heavier software.
Self-hosting as the alternative route. Because the underlying models are openly released, running them yourself is a genuine option. Self-hosting the same open weights is documented as needing a Linux 64-bit environment and an NVIDIA GPU with at least 32 GB of VRAM. That hardware requirement is the clearest argument for why a hosted front end exists at all.
No published public API tier. The site does not advertise a documented developer API or SDK as part of its plans. Teams needing programmatic batch reconstruction should evaluate self-hosting or an inference platform that exposes these models as an endpoint. (To be verified)
Access is subscription-based across three published tiers — Basic, Pro and Max — with an annual billing option advertised at half the monthly rate. The tiers differ by volume rather than capability: all three list the same feature set, and what changes as you move up is how much you can process each month. Each tier meters two separate budgets: a pool of monthly generation credits and a distinct monthly allowance of image segmentation requests. That dual accounting is worth internalising before you subscribe, because a workflow that explores many candidate selections consumes segmentation allowance quickly while barely touching generation credits, and a batch conversion workflow does the opposite. The published tiers also state an effective per-credit rate, which falls as the tier rises — the usual volume-discount shape.
Paid tiers advertise watermark removal, priority processing speed and commercial usage rights as plan-level entitlements. The presence of watermark removal as a paid feature implies that unpaid or trial output carries a watermark, which is the conventional arrangement for this category. The site does not publish a detailed free-tier allowance, a refund policy or an enterprise contract track on its pricing page, and prices for any subscription service change over time. Verify the current rates, the exact credit cost of a single generation, and the commercial-use terms on the pricing page itself before committing to a plan, particularly if your work depends on the commercial rights. (To be verified)
If single-image reconstruction is what you need, several routes exist and they differ mainly in who operates them. The most direct alternative is going upstream: the SAM 3D weights are openly published by Meta, so running them yourself removes the middle layer entirely, at the cost of provisioning a substantial GPU. Meta also operates its own hosted playground for trying the models, which is the natural first stop if you simply want to see the research output without a subscription.
Beyond this model family, the broader image-to-3D and text-to-3D generation space includes dedicated 3D generation products aimed at game-ready assets, and general inference platforms that expose many open models — including this one — behind a pay-per-call API, which suits automated batch work better than a seat-based subscription does. For work where accuracy genuinely matters, the honest alternative is not another generator at all: classical photogrammetry from a proper multi-angle capture, or a real 3D scanner, still produces measurably correct geometry that no single-image method can match. Choose by requirement — speed and convenience point to hosted generation, fidelity and dimensional trust point to actual capture.
Resolution and detail are bounded. Meta's own model documentation is candid that moderate output resolution limits detail, and that the object model cannot reason about physical interactions such as contact or interpenetration. Fine surface texture, crisp mechanical edges and small features will be smoothed or lost. Objects reconstructed separately may intersect each other when assembled, because nothing in the pipeline enforces physical plausibility between them.
The unseen side is invention. Everything facing away from the camera is generated from learned priors, not observed. For familiar, symmetric objects the guess is often convincing. For asymmetric objects, or anything with distinctive detail on the hidden face, the reverse side will be confidently wrong. Always inspect from behind before you trust an asset.
Human reconstruction has documented boundaries. For human reconstruction, the underlying model processes each individual separately, without considering multi-person or human-object interactions. Group photographs will not yield a coherently interacting group, and fine hand articulation is a known weak point of general body models.
Meshes are not manufacturing-ready. One independent hands-on review aimed at additive manufacturing concluded the meshes were not suitable for direct processing in slicers, citing surfaces that were neither smooth nor watertight. Anyone planning to 3D print output should expect a repair and remeshing pass, and should not assume printability.
Topology is generated, not authored. Expect dense, irregular meshes without artist-planned edge flow or clean UV layout. For animation, deformation or a tight polygon budget, retopology is not optional.
Input constraints are real. Uploads are limited to JPEG, PNG and WebP at up to 10MB, and body reconstruction works best from full-body shots. Low-resolution, heavily compressed, motion-blurred or extremely oblique photographs will degrade results before the model has a chance.
Rights are layered, and marketing language flattens them. The upstream code carries a permissive license similar to Apache 2.0 that is free for commercial use, which is precisely what makes independent hosting of these weights lawful. That licence governs the model, not your input image. Reconstructing a copyrighted product design, a piece of artwork, or a recognisable person creates obligations that no model licence resolves.
Privacy deserves a deliberate decision. Uploading a photograph to any hosted service means transmitting it to a third-party operator. Photographs of identifiable people are personal data in many jurisdictions, and body reconstruction is inherently an operation on images of people. The site does not prominently publish a data-retention schedule, so confirm current terms before uploading anything sensitive, confidential or unreleased. (To be verified)
Operator continuity is an unknown. This is an independent service built on someone else's open models, and this category churns. Download and archive assets you depend on rather than treating the service as durable storage.
No. This is an independent commercial service that hosts Meta's openly released SAM 3D models behind its own interface and billing. The site's own FAQ area carries an explicit disclaimer that the project is independent and not affiliated with Meta. If you specifically want Meta's own hosted demonstration or the original research materials, go to Meta's own AI properties instead. The distinction matters for support, billing, privacy terms and any commercial commitment.
One image. Uploads accept JPEG, PNG and WebP files up to 10MB. Unlike photogrammetry, no multi-angle capture set is required — which is the whole point of the technique, and also the source of most of its limitations, since everything not visible in that single frame has to be inferred.
Object exports come in GLB for games and web applications, PLY point clouds for technical workflows, or a complete ZIP bundle with all formats included. Human body reconstructions export as GLB. GLB is the right default for anything heading into a game engine or a web viewer; PLY suits technical pipelines that expect point data.
Paid tiers advertise watermark removal, priority processing speed and commercial usage rights as plan-level entitlements, and the upstream model licence permits commercial use. But that only covers the tool and the model. It does not grant rights over what is depicted in your source photograph — a copyrighted product design, an artwork, a logo or a recognisable person all carry separate obligations. Read the current terms of service and clear your source imagery independently.
Good enough for visualisation, prototyping, props and reference; not good enough to be treated as measurement or as final production geometry. Meta's own model documentation is candid that moderate output resolution limits detail, and that the object model cannot reason about physical interactions such as contact or interpenetration. Plan for a cleanup pass.
Not reliably without repair work. One independent hands-on review aimed at additive manufacturing concluded the meshes were not suitable for direct processing in slicers, citing surfaces that were neither smooth nor watertight. You would need to run the mesh through repair and remeshing tools, and verify watertightness, before slicing.
Because the hardware is the barrier. Self-hosting the same open weights is documented as needing a Linux 64-bit environment and an NVIDIA GPU with at least 32 GB of VRAM. If you already have that hardware and the operational appetite, self-hosting is cheaper at volume and keeps your images on your own infrastructure. If you do not, a subscription converts a capital and operations problem into a monthly fee.
They are two separate meters. Each tier meters two separate budgets: a pool of monthly generation credits and a distinct monthly allowance of image segmentation requests. Selecting and refining a mask draws on the segmentation allowance; producing the actual 3D mesh draws on generation credits. In practice this rewards a workflow where you segment freely to find the right selection, then generate only once.
Each person is handled independently. For human reconstruction, the underlying model processes each individual separately, without considering multi-person or human-object interactions. You can reconstruct multiple figures from one photograph, but their relative positions, contact and interaction will not be modelled coherently — you will need to arrange them yourself in a 3D application.
The service is built around single-image reconstruction, so the practical approach is to extract a representative frame from your video and use that as the input. Choose a frame with minimal motion blur where the subject is unobstructed and large in frame. (Based on public information)
