3Dnatives Lab: Testing Meshy, the AI 3D Model Generator, for 3D Printing

Meshy, the AI 3D model generator, did not start as a 3D printing tool. When Yuanming (Ethan) Hu launched Meshy-1 in 2023, the pitch was aimed at game developers and animators: characters, props and environments, generated in a minute and dropped into Unity or Unreal. Hu, a Tsinghua graduate who did his PhD at MIT’s Computer Science and Artificial Intelligence Laboratory, had already created Taichi, a programming language for graphics and physics simulation. At the time, generative 3D forced a choice between slow methods that lifted 2D image models into three dimensions and fast native 3D methods that produced blobs. Meshy set out to close that gap and put academic research in a browser, where an artist could actually use it.

Meshy has released seven major models in under three years. Meshy-2 arrived in February 2024 with polycount control, and Meshy-3 followed in April 2024 with a new image-to-3D pipeline and physically based rendering (PBR) textures. Meshy-4, released in August 2024, promised hard surfaces “free from the bumps and dents” of earlier versions. Meshy-5 added multi-view input in August 2025, and Meshy-6 brought multi-color 3MF export and a low-poly mode in January 2026. Meshy-7 came out in August 2026, tuned to match the input image more faithfully. Meshy-7.1 followed on September 10, 2026, with an “Ultra 4K” geometry mode, which the company says produces the richest surface detail of any generator on the market.

The money followed too. In July 2026, Meshy announced a Series B of nearly $400 million at a $1.5 billion valuation, which it describes as the largest round ever raised by a company built specifically for AI 3D. IDG Capital, Matrix Partners China and Monolith led the round, joined by existing backers including HongShan (formerly Sequoia China) and Source Code Capital. At the time, Meshy claimed more than 12 million registered users and over 100 million generated models.

Games remain Meshy’s core business. It auto-rigs characters, offers several hundred animation presets and has plugins for Unity, Unreal, Godot, Roblox and Blender. Nexon and NetEase Games are among its customers. Since Meshy-6, the company has also turned toward our world, adding one-click export to eight slicers and tools to check, repair and split models for printing. Bambu Lab, Creality, Elegoo, FlashForge and xTool are named customers, and a “Print with Form Now” button sends models straight to Formlabs’ on-demand printing service.

While our previous coverage tracked Meshy’s funding, its 3D Agent and its published specifications, we wanted to look beneath the surface and examine the export files themselves. So we put Meshy-7.1 through the 3Dnatives Lab to answer one practical question: when you upload a photo of a real object and click Export, what does the slicer actually receive? We deliberately evaluated the exported files rather than physical prints. A physical print reflects hardware calibration and slicer settings as much as software quality. Meshy delivers a digital file, and that file is what we evaluated.

1. Getting started, plans and credits

Meshy runs in the browser. It also offers a mobile app for capturing photos on the go, plus plugins for Blender, Unity, Unreal, Godot and Roblox. Creating an account takes a minute and requires no credit card. The free plan gives 100 credits a month and one generation at a time in a low-priority queue. It also comes with a condition that matters for makers: Meshy owns everything you generate on the free plan and licenses it back to you under CC BY 4.0. You can use the models and even sell prints of them, but you must credit Meshy, and the models are public.

Paid plans change that. Pro costs $20 a month and includes 1,000 credits, private ownership of your models, 10 simultaneous tasks and API access. Premium ($40, 3,000 credits) and Ultra ($100, 8,000 credits) add volume, and a Studio plan covers teams. Monthly credits do not roll over. In our test, a generation at the Ultra 2K setting we used cost 25 credits (20 at Standard, 30 at Ultra 4K) and took about a minute, with a further 10 credits for the texture and 5 more for 8K. Auto Repair, Auto Split and the multi-color step cost 10 credits each, and the printability check is free. On paper, Pro buys about 30 textured models a month, fewer once you start retrying.

The fine print deserves a look. Meshy’s Terms of Use (section 2.9) allow the company to use both the images non-Enterprise customers upload and the models they generate to train and improve its services. The FAQ on its own pricing page says the opposite: Meshy will not use your data “for any training purpose without your consent”. The help center sits in between, stating that uploaded images are not currently used for training and that non-Enterprise data may be used in anonymized form later. Enterprise customers are excluded by contract.

For this test we used Pro plan starting with 1,000 credits.

2. The software

The workspace is uncluttered, and its menu shows where Meshy comes from. The generation modes sit on the left: Text to 3D, Image to 3D, AI Texturing, an image generator, auto-rigging and animation, and the 3D Agent, a chat window that suggests directions, generates variants and answers printing questions in the same conversation. Image to 3D accepts a single photo, up to four views of the same object, or a batch of up to 10 different images. Each generation returns four candidate previews; you pick one and refine it. Meshy-7.1 is the default model, and Meshy-6, Meshy-6 Lite and the faster Smart Topology models remain selectable. For a game artist, this is a complete pipeline: generate, retexture, remesh to a target polygon count, rig, animate and export as FBX. For a maker, most of it is beside the point.

The 3D printing tools appear in a separate panel once a model exists. The printability check looks for holes, non-manifold edges and other watertightness problems. Auto Repair promises to fix broken meshes, and Auto Split cuts a model into watertight parts and lays them out on a build plate, with connectors announced for a later update. The multi-color tool writes a 3MF with color data for Bambu Lab’s Automatic Material System (AMS), Creality’s Filament System (CFS) and OrcaSlicer. Exports include STL, 3MF, OBJ, GLB, FBX, USDZ and .blend, and on paid plans a single click opens the model in your slicer.

Meshy’s own figure for all this, published in a May 2026 blog post that sliced 75 models from five generators in Bambu Studio, 15 per tool, is that around 97% of figurine-type models pass validation. That is an internal benchmark (and 97% of 15 models is not a whole number), and it is the claim we most wanted to check.

3. Test method

We chose nine inputs to cover the cases a maker is likely to throw at the tool. Two were clean product shots: a photo of a 3D-printed figure of our mascot, Leo, and a studio photograph of a Jacob & Co. Astronomia, a watch with a notoriously complex dial. Two were multi-photo tests of a real object, a Seiko 5 Sports GMT: three phone photos of the watch lying on a desk, then three photos of the same watch on the wrist. One was a car, the Ligier JS2 R, from three press photos taken from the front, side and rear. One was a wide tourist photo of Saint-Julien Cathedral in Le Mans, buttresses, city wall and neighboring houses included. The last three were images we generated ourselves with Gemini: a bas-relief plaque of the 3Dnatives team and two poses of a mecha figurine.

All nine were generated with Meshy-7.1, and every model took about a minute or less, textures included. We exported each one as a 3MF file, the format Meshy recommends for printing. We then analyzed every file before slicing: polygon count, dimensions, whether the mesh was watertight, how many separate bodies it contained, and how many edges were shared by more than two triangles, the non-manifold junctions that trip up slicers. We also compared each model with its photograph from seven angles.

We then opened each 3MF in Bambu Studio to see what a maker sees: the model on the plate, the size it arrives at and any warning the slicer raises.

4. What came out

Leo, the 3Dnatives mascot

The simplest input produced the cleanest result. From one PNG on a white background, Meshy returned a single closed solid of 385,000 triangles: the egg-shaped body, the waving hand with separate fingers, the two feet, raised domes for eyes and a groove for the mouth. The file also contains two tiny slivers, which Bambu Studio did not flag. The back of the model, which the image never shows, is smooth and plausible.

One detail made us smile. The source image is a photograph of an FDM print, and Meshy reproduced its visible layer lines as ridges in the geometry. It copies what it sees, including surface texture that should not be there.

Leo: the source image and the model from the front.

The model seen from behind: the side the image never showed.

In the slicer, Leo needs no work: there are no warnings, and supports are needed only under the raised arm. At 50 mm, the ridges from the layer lines are about the size of a 0.2 mm layer, so they would print as a faint texture rather than a defect.

Ligier JS2 R, from three photos

The car was the most convincing complex result. Three photos from three angles gave Meshy enough to build a single closed body of 778,000 triangles, plus one tiny sliver that Bambu Studio did not flag. The detail is striking: the rear wing sits on its supports, both mirrors are on their stalks, and the roof scoop, side vents, splitter, diffuser and spoked wheels are all there. Against the real car’s dimensions of 4,415 by 1,900 by 1,180 mm, the model comes out about 7% too wide and 3% too tall once the length is matched. The 8K texture keeps the blue, white and red livery and the number 26 on the doors, but all the lettering (the JS2 R on the sills and the banner across the windshield) has turned into unreadable shapes.

Ligier JS2 R: one of the three source photos and the textured model.

At the exported length of 182 mm, the wheel spokes, wing supports and mirror stalks are around a millimeter thick, which is printable on FDM only with a fine nozzle and patience. Scaled to 250 mm, they grow to about 1.4 mm. The underbody is flat enough to print directly on the plate, wheels down.

The front of the model: the windshield banner has turned into shapes.

Two watches, and the difference three photos make

The Jacob & Co. Astronomia Solar Zodiac (reference AS310.40.SP.ZK.A, 43.4 mm, rose gold) is a stress test by design: a skeletonized movement with planets on stalks, a tourbillon (a tiny rotating cage in the movement), a zodiac disc and gears, all under a domed crystal. On the dial side, Meshy’s geometry is remarkably detailed. The gears are there, the tourbillon is there, and the planets are separate spheres. The problem is what to do with it. The file contains 35 separate bodies, and several of the planets have lost their stalks and float above the dial with nothing holding them. The finest parts are well under a millimeter thick at the exported size. The strap, which leaves the frame in the photograph, has been completed into a full loop with a buckle and a keeper that Meshy invented from nothing. With the strap cut off, the model could work as a resin display piece, but not as a print straight from the file.

Jacob & Co. Astronomia Solar Zodiac: the studio image and the dial side of the model.

The strap Meshy invented, with buckle and keeper, seen from behind.

The Seiko told us more, because we photographed it ourselves. Three phone photos of the watch lying on a desk produced a very good model. The serrated bezel, the crown, the hands, the hour indices and the date window are all there, along with a jubilee bracelet with individual links, its ends curled upward exactly as they were on the desk. The numerals on the bezel are half legible, the SEIKO 5 logo is a blob and the date is a bump. The case back, which no photo shows, has been given an engraved ring of text that does not exist on the real watch. Under the surface, this was the roughest file of the set: 51 bodies and 99 non-manifold junctions, although not a single open edge.

Seiko on the desk: the phone photo and the model from above.

The invented case back, with its ring of text.

Then we photographed the same watch on the wrist, and Meshy modeled the wrist instead. The result is a forearm and a hand, knuckles and hair included, with a small watch on it. The tool removes a plain background well enough, but it cannot tell which object you want. Anything touching the watch becomes part of the model. So if you want the watch, be sure to photograph it alone.

The same watch photographed on the wrist, and what Meshy modeled.

Le Mans cathedral, from one postcard view

From the photographed side, the cathedral is impressive: the tower, the tiers of flying buttresses, the radiating chapels of the apse, the lancet windows, even the city wall in the foreground and a base plate under everything. Walk around the model, though, and the building is wrong. The nave, which in reality stretches away to the west, is missing. To fill in the side it could not see, Meshy copied the rounded east end. Seen from above, the building is round at both ends instead of forming the cross shape of a real cathedral.

Le Mans cathedral: the photo and the model from the photographed side.

The invented side, and the plan from above: round at both ends, not a cross.

The file is 36 MB and contains just under three million triangles. At the exported height of 50 mm, the buttresses and pinnacles are thinner than any FDM nozzle can produce. Scaled to 150 mm, the buttresses reach roughly a millimeter, and a resin printer could resolve them. On an FDM printer, only the base and the tower would survive; the rest is a cloud of stone lace too fine to print.

Three AI-generated inputs: a plaque and two figurines

We also fed Meshy three images we generated with Gemini, which turn out to be ideal source material: one object, evenly lit, on a plain background. The first, a bas-relief plaque of the 3Dnatives team, is the best printing candidate of the whole set. Meshy turned it into a relief 6.6 mm thick with a flat back, the five faces readable, the gears in the border and, for once, legible text: the large 3DNATIVES survived, the smaller tagline did not. Meshy’s own printability check gave it a green tick, though the file still holds eight bodies and 16 non-manifold junctions.

The 3Dnatives plaque: the Gemini image and the relief.

The edge view: 6.6 mm thick with a flat back.

Two poses of a mecha figurine came out as the kind of model Meshy was built for: a rifle held in two hands, antennae, backpack cylinders, a tablet, a base with cables and panels. At 50 mm the antennae are fragile, and both files carry dozens of detached fragments (19 and 36 bodies), but scaled to 100 mm or more they would print on resin without an argument.

Figurine 1: the Gemini image and the model.

Figurine 1 from behind.

Figurine 2, holding a tablet: the Gemini image and the model.

Figurine 2 from behind.

5. From model to print

Four points about the files themselves apply across our tests, and Meshy’s marketing does not highlight any of them.

First, every 3MF Meshy exported is exactly 50 mm tall. The mascot, the car, the cathedral, the plaque and a watch lying flat all arrived at the same height. The 3MF is a Bambu Studio project generated on Meshy’s server and centered on the build plate, and it carries no real-world dimensions. That isn’t a flaw in itself, but users need to know it: nothing about a Meshy model is a measurement, and you set the scale yourself in the slicer.

Second, “watertight” needs unpacking. None of our exported files had a single open edge. But only two of the nine models, Leo and the Ligier, were a single closed solid. The other seven are stacks of self-touching pieces: between 8 and 51 bodies per file, most of them tiny two-triangle slivers, joined at edges shared by more than two triangles (between 16 and 99 per file). Bambu Studio flagged nothing to repair, and in practice a slicer will accept these files. They are printable, but they are not clean.

Model Triangles Pieces in the file One clean solid
Leo 384,932 3 Yes, 51 cm³ at 50 mm
Ligier JS2 R 777,908 2 Yes, 396 cm³ at 182 mm
Jacob & Co. 1,779,838 35 No
Seiko, on desk 1,746,428 51 No
Seiko, on wrist 1,223,204 14 No
Le Mans cathedral 2,972,546 17 No
3Dnatives plaque 1,311,110 8 No
Figurine 1 2,638,506 19 No
Figurine 2 2,589,916 36 No

What the slicer receives, measured on the exported 3MF files. No file had a hole; “pieces” counts separate bodies, most of them slivers a few triangles wide.

Third, the files are heavy: between 385,000 and 2.97 million triangles per model, and up to 36 MB per 3MF. Slicers handle them, but slowly. Meshy’s Smart Topology mode can bring a model under 15,000 triangles in about 10 seconds, but it is a separate step that costs extra credits, and it would erase the fine relief that makes these models interesting.

Fourth, texturing a model does not give you a color print file. We compared the Ligier’s 3MF exports before and after texturing: the geometry is identical, and neither file contains any color information (no paint data, color groups or vertex colors). The 8K texture stays in the GLB. To print in color with an AMS, you have to run the separate multi-color tool, which costs another 10 credits and produces a different 3MF.

Conclusion

Pros:
  • The fastest and easiest route from a photo to a printable figurine we have used
  • No open edges in any file, and simple objects come out as one clean, closed solid
  • Three photos of a real object can be enough for a convincing model
  • Direct export to eight slicers, and a free printability check
Cons:
  • Exports carry no real-world scale: everything arrives 50 mm tall
  • Complex objects come out as many self-touching bodies, and small lettering is unreadable
  • The unseen side of every object is invented
  • Retries use up credits quickly, and texturing does not give you a color print file

Meshy was built for games and animation, and it shows: a full pipeline from prompt to rigged, animated, textured asset in a few minutes, in a browser. The printing suite added over the past year sits on top of that engine. It gained one-click slicer export and a printability check, but it inherited models designed to look right on screen rather than to match a real object.

For the objects its printing tools target, Meshy does what it says. A mascot, a toy, a figurine, a plaque, a pet: photograph it alone on a plain background, and you will have a printable solid in a minute, for a few tens of credits, with a back side that nobody will question. The Ligier shows that three photos from three angles can capture a far more complex shape, and the Seiko shows that a phone on a desk is enough to get there.

But Meshy is not a scanner, and it is not CAD. Every surface the camera did not see is a guess, and the guess is confident enough to fool anyone who has not seen the real object. Anything with small lettering, thin features or a tolerance belongs elsewhere. As for the 97% figure, Meshy’s benchmark measures whether a slicer accepts the file, and on that measure our results agree: Bambu Studio flagged nothing on any of our nine models. Being accepted is not the same as being clean, though. Only two of our models came out as a single closed solid.

At $20 a month, Meshy is worth it for anyone who prints for fun, sells small custom pieces or needs a display model of something they can photograph. To find out more about Meshy, visit the company’s website HERE.

Have you printed a model generated by Meshy? Let us know in a comment below or on our LinkedIn or Facebook pages! Plus, don’t forget to sign up for our free weekly Newsletter to get the latest 3D printing news straight to your inbox. You can also find all our videos on our YouTube channel.

*Photo credits: 3Dnatives, except the Ligier JS2 R (Ligier Automotive), the Astronomia (Jacob & Co.) and the cathedral (stock photo). Model renders: 3Dnatives, from Meshy exports.

Lily-Swann Frost: Writer and digital marketer at 3Dnatives, covering the latest developments in additive manufacturing, 3D printing, and advanced manufacturing technologies.
Related Post
Disqus Comments Loading...