We Tested Meshy 6: Can AI Finally Generate Good 3D Printable Models From Photos?

Meshy frames its new Creative Lab as a tool for democratizing manufacturing. That may be a bit of a stretch. While you can absolutely manufacture parts created in Meshy, the platform doesn’t offer the level of control typically associated with the word “manufacturing.”

According to the company, the tool is designed for a broad audience: 3D printing hobbyists, makers, tabletop gamers creating custom miniatures, and everyday consumers looking to produce personalized gifts. Not mentioned are engineers or artists looking for accurate 3D models of their concepts (more on accuracy later).

Meshy AI told All3DP that the 3D printing community is quickly becoming one of the platform’s largest user bases. Still, the majority of subscribers are individuals and companies that need 3D assets for digital environments in marketing, animation, game development, or even manufacturers looking to generate 3D models at scale, like model resellers.

The Creative Lab also reflects a broader industry trend toward “phygital” experiences—blending digital design with physical production—and the growing role of generative AI in automating end-to-end workflows, the company says.

If the platform delivers on its promise, it could make high-quality, full-color 3D printed products far more accessible.

Here’s how it went for us.

Turning an Image Into a 3D Model

We uploaded four images of a well-known statue here in Munich called Bavaria. You may recall we used the same images a few weeks ago while testing another image-to-3D generator, Hitem3D.

Meshy operates on a credit-based system, with a highly limited free plan and then subscriptions that top up your credit account: $16 per month for 1,000 credits or $48 for 4,000, plus an undefined “enterprise” plan.

In the latest version of Meshy, you can upload multiple reference images (up to 10), which the platform then interprets as a 3D model. Since each reference you upload costs 20 credits, we went with four for the Bavaria at first. Within minutes, Meshy generated four 3D model options.

Meshy, like all AI, can astound you one minute and confound you the next. The AI, which is custom built by Meshy, draws on additional visual references beyond the images you upload. The result in our case looked like a mash-up of our photos and the AI’s general idea of what a “statue” should look like.

Still… wow. The result was surprisingly close.

Of the four generated options, we chose the one that looked most accurate even though it gave the lion at Bavaria’s side two tails in the back.

The fun or frustrating part about Meshy, depending on how you look at it, is that you never get the same model twice. We uploaded the same four images a few times, and, with each round, we got slightly different versions and were never able to reproduce the two-tailed lion again.

Each new generation costs additional credits, of course. However, Meshy also offers a free “Retry” option for generated models (with the paid plans). In our testing, retries only made subtle changes—perhaps improving or reducing accuracy by around 10%.

To test for maximum accuracy, we uploaded 10 images of the Bavaria from a wide range of angles expecting that more references would create a better result. Surprisingly, more photos didn’t help. From 10 images costing 200 credits we received 10 3D models where each seemed to be generated from one reference photo instead of the AI stitching together all 10 reference images.

By contrast, in another tool, Hitem3D, we could label images as the front, back, left side, and right side. In Meshy, the AI didn’t seem to interpret the photos as being of the same object, it was more like batch processing models instead of compiling them to create one accurate model.

There’s a chat helper you can use to refine your processes if you aren’t getting the results you need. When we told the AI chat helper that we wanted the Bavaria statue it said: “You can upload multiple images, and I will do my best to combine them into a 360-degree capture. Please ensure that the images cover all angles of the statue for the most accurate representation.” So, clearly, the aim is to use all the images as references for one model, but our results didn’t reflect this.

If you want to 3D print a bust of a person in your life or maybe your pet by uploading several photos of them to Meshy, the result will be an AI-interpretation of that person. It’s not a stand-in for 3D scanning or true photogrammetry. We put it to the test with 10 photos of the actor Pedro Pascal and got 10 3D models.

Again, Meshy delivered 10 3D models where each was based on one of the reference photos. It did not use all the reference photos to create a more accurate 3D model.

Toggling on “Image Enhancement” in the upload window does create better results. Without it, we got arms floating in space and other odd and un-printable anomalies.

Printing Your Models, Importing to CAD

Once you create a model from an image, you can’t refine it with text or by uploading new images. There’s only that “retry” option mentioned earlier. So there’s no way to directly manipulate the model, as in, removing the lion’s second tail.

However, for more detailed model manipulation, you can export the model file directly to Blender or download the file (as 3MF, STL, OBJ, or usdz, glb, and fbx for applications like game design and animation) for upload to any type of CAD software, if you have those skills. One giant plus for Meshy AI is that the model files are large enough (hundreds of thousands of mesh faces) to be easily edited in CAD to produce high-quality models.

The process above is not the only way to get from 2D to 3D using Meshy. There’s the new Nano-Banana-powered text-to-image option which uses the Google’s Nano Banana image generator. Here you can couple a text prompt with reference photos. Using the same four images plus a description of the famous statue, it created one near-perfect image and three “interpretations.”

There’s a disconnect in how these separate tools interact, though. Any image generated in the Nano Banana space can only be used as the sole reference image for a 3D model, rather than as a part of the 10-image option. Yet — and this is a hint at how fast platforms like this are evolving — there’s a beta-test option that lets you upload three more reference images to refine your Nano Banana image, resulting in maybe a 5% improvement. So in the image above using the Nano Banana option, to get a 3D model, we could only take the top right image and put that in the 3D model generator which was essentially like only using one source photo.

Going back to our preferred whimsical two-tailed interpretation of the Bavaria Statue, sending it to our Bambu Lab P2S was far easier than anticipated. Simply hit the export button and it opened up in the slicer. We downsized it to 50% since the file was enormous, but didn’t adjust any other settings. The file was not printable, though until we cleaned up the non-manifold edges Formware’s free online STL repair tool, which handled the issue in under a minute. A few hours of printing later, and we had a nicely detailed statue (printed in Cairo Gold PLA Extrafill from Filamentum) that we were really happy with.

But what if you have a drawing or just one image to work with? We uploaded a colored cartoon-style unicorn drawing and had Meshy render it in 3D (see below).

In all, we were happy with the results even though there’s a extra leg on the unicorn, which is a flaw that could be a deal breaker for many applications.

This colorful model gave us the opportunity to try Meshy’s color 3D print options and the new Texture feature that lets you color your models. The Texture is far more than texture, it’s coloring and refining. In the texture window you can prompt the AI to adjust the colors of your model or you can take a paintbrush tool and color them yourself. You can even upload images as references.

Once you’re happy with your model painting, send it to your own multi-color printer by clicking on the Multi-Color Printing button. There you can define the color count from 1 to 16, the level of detail from 3 to 6 (use 6), and match the models colors to standard PLA colors in regular, matt, silk, wood, and other types.