Neural networks for 3D: models from text and photos with no Blender experience

⚡ The gist in 5 seconds - From a text prompt or a single photo, a neural network builds a 3D model in a minute that opens in Unity — no serious Blender experience required. - The tools are available via services like Meshy and Tripo, and through BotHub you can work with AI models without a VPN and pay with Russian cards. - The main limitation: the result often needs manual refinement, and a photo doesn't reveal the hidden sides of an object. ### 🔍 What was found A review on the BotHub blog on Habr breaks down the 3D generation pipeline. The tools fall into two types: generative diffusion models (Meshy, Tripo, and others) and reconstruction via NeRF / 3D Gaussian Splatting. Tripo P2.0 at Gamescom 2026 demonstrated generation of native quad meshes, immediately suitable for topology editing. Meshy's sixth version bet on geometry: smooth organic shapes, clean edges on hard surfaces, and a Low Poly mode for games without manual retopology. There is also automatic rigging: a skeleton in half a minute, a library of 600+ animations, and export to FBX or GLB. Export formats are strictly divided by task: GLB is the universal standard for the web, Three.js, and Android AR; FBX is for exchange between Maya, Blender, Unity, and Unreal and is the only one that properly carries skeletal animation; OBJ is geometry and UVs only, opens in almost everything; USDZ is for AR Quick Look in Safari, Messages, and Vision Pro. A believable render requires PBR materials: without roughness and metallic, the model looks like a plasticine lump no matter how many polygons you pour into it. A photo gives better geometry with the right angle, but only shows what's visible; text wins on control over materials and proportions but tends toward an averaged image from the training data. ### 💡 Why it matters 3D models are a bottleneck in game development, AR, e-commerce, and 3D printing: an artist spends from several hours to several days creating and rendering a single model. Neural networks promise to compress this into minutes, though some of the promises remain marketing. For beginners, the main barrier is removed: no need to spend weeks learning 3D editors — just write "a wooden chest with metal corners" and get a finished file. Practical uses include quick prototypes, turnkey characters for indie projects via automatic rigging, and product scanning for e-commerce. ### 🧩 Context Until recently, such generation seemed almost impossible; now it works just like with images and video. The neural network doesn't draw the model whole — it reconstructs the 3D structure from interpretable data, which is why proper instructions matter. If you don't want an averaged "chest," you'll have to describe textures and materials in detail. With photos it's the opposite: the network gets the real geometry of a specific object, but the back wall, cavities, and details at other angles are either inferred from symmetry or left as holes. Multi-angle capture partially covers the problem but doesn't give the full picture.

🔗 Read on habr.com