Image to 3D for games: from a picture to a GLB

A capybara, a treasure chest or a knight needs a mesh with a shape you’d recognize, and an image-to-3D tool can make that mesh from a single picture. You still have to clean it up for the game, though.

The raw result often has too many triangles, lighting baked into the colors and no skeleton. So we’ll go through the whole path, from the picture you feed in to the GLB your game loads.

Prices, limits and licenses are as of September 2026 and linked where they come up, so recheck them before you pay for anything.

1. Make a concept image the model can read

The output can only be as clean as the input, and Meshy’s guidance works for pretty much every tool (Meshy help).

• One object, filling the frame, on a plain white or removed background.

• Front view or a slight angle, no extreme top-down or bottom-up shots, and little perspective distortion.

• Even, diffused light. Shadows in the picture get baked into the texture, so a dramatic key light turns into a permanently dark cheek.

• Characters in an A-pose or T-pose, full body, limbs apart and uncropped. Overlapping arms break auto-rigging later.

Make the picture with any image model, and ask for exactly that. Something like “a single cartoon treasure chest, three-quarter front view, plain white background, soft even studio light, no shadows” does the job. For props with a different back or sides, use multi-view input.

Meshy’s multi-image endpoint accepts several views of one object (Meshy API).

Multi-view setting	Requirement
Input views	1 to 4
Meshy 7.1 order	Front view first

Tripo has its own multiview-to-model endpoint too (Tripo docs).

2. Pick the tool, and read its license first

Tool	Low-poly option	Rigging	License notes
Meshy 7.1	target_polycount from 100, Smart Topology	Humanoids, up to 300,000 faces	Free output is CC BY 4.0 with credit; paid users own it
Tripo P1 and H3.1	P1 is built for low-poly, 48 to 20,000 faces	Biped, quadruped, avian and more	Check the plan terms on Tripo’s site
Hyper3D Rodin Gen-2.5	Smart Low-Poly topology	Not covered here	Paid Creator and Business plans list “unlimited export and any use”
Hunyuan3D 2.1	Open weights	Not covered here	License excludes the EU, UK and South Korea
TRELLIS.2	Open weights, needs a 24 GB NVIDIA GPU	None	MIT
Stability SF3D, SPAR3D	Open weights	None	Free commercial use under US$1M annual revenue

The table comes from Meshy API, Meshy plans, Meshy license, Tripo models, Tripo P1, Rodin, Rodin pricing, Hunyuan3D license, TRELLIS.2 and the Stability license.

If you want to try them, new users who join through our Meshy invite link get 30 days of Meshy Pro for $1, according to Meshy’s referral terms. Tripo has a web studio too.

Both are referral links, so we may earn credits when you sign up, but a referral never changes what a guide says or how a game ranks.

Read the free tier’s terms before you generate final assets, because Meshy’s free allowance is limited to its Lite model (Meshy free plan).

Free-plan allowance	Limit
Credits	100 a month
Downloads	10 a month
Downloadable models	Meshy 6 Lite only

That output is also licensed CC BY 4.0 and needs a “Model created with Meshy” credit line (Meshy license). If you’d rather not credit a tool, or its free tier limits commercial use, generate final assets on a paid plan and note which plan made which file.

3. Ask for fewer triangles at the start

Every tool can produce millions of triangles, while a browser game really only wants thousands, so set the budget when you generate instead of fixing it later. For Meshy, choose topology: "triangle". These are the polygon controls Meshy and Tripo give you (Meshy API, Tripo P1).

Control	Range or default
Meshy target_polycount	100 to 300,000; default 30,000
Meshy Smart Topology	Default 4,000
Tripo P1 face_limit	48 to 20,000

Two more Meshy settings are really worth knowing. On the Meshy 6 model, remove_lighting is on by default and strips highlights and shadows from the base color, and setting pose_mode to a-pose or t-pose straightens a character for rigging (Meshy API).

In a browser, aim for the low end of a prop budget and only spend more on the hero object the camera lingers on. The starting ranges below come from a polycount guide (polycount guide).

Prop size	Triangle range
Tiny	100 to 2,000
Standard	1,000 to 10,000

4. Clean up what the tool got wrong

Open the GLB in Blender and actually look at it before you use it. These are the usual fixes.

• Too dense. The Decimate modifier’s Collapse mode cuts triangles by a ratio (Blender manual).

• Messy topology on something that bends. Automatic remeshers don’t produce topology that deforms well, so retopologize characters by hand or with the retopology tools (Blender manual). Static props don’t need it.

• Lost detail after decimating. Bake the dense mesh’s detail into a normal map on the light one, with “Selected to Active” baking (Blender manual).

If you work with a coding agent, Blender’s MCP server lets it run these steps for you, so you can just ask it to “decimate to 3,000 triangles, bake normals from the original, export a GLB”.

5. Rig and animate

Humanoids are the easiest, of course. Meshy’s auto-rigging handles humanoid meshes with clear limbs up to 300,000 faces, and it outputs FBX and GLB with walking and running clips (Meshy rigging). Tripo rigs a lot more body types, including quadrupeds, birds, snakes and even fish, and it can name the bones for a Mixamo-compatible skeleton (Tripo rigging).

Mixamo, Adobe’s auto-rigger and clip library, is free with an Adobe ID and only handles bipedal humanoids (Adobe FAQ).

For moves no library has, AI 3D animation for games compares text-to-motion, video capture and physics, and the NVIDIA Kimodo guide generates clips from text on your own GPU.

6. Shrink the file with glTF Transform

Generated GLBs tend to arrive pretty heavy, mostly because of their textures, and Meshy’s go all the way up to 8K (Meshy API). glTF Transform’s CLI optimizes a file in one command (glTF Transform).

npm install --global @gltf-transform/cli
gltf-transform optimize chest.glb chest-web.glb --compress meshopt --texture-compress webp --texture-size 1024
gltf-transform inspect chest-web.glb

inspect reports whether a scene is geometry-heavy, texture-heavy or has too many draw calls, so you can check the result and don’t have to guess. There are three things to know here.

• Compression cuts download size, not frame time. glTF Transform’s docs say neither Draco nor Meshopt improves runtime performance directly (Meshopt docs). Fewer triangles and smaller textures do.

• Compressed textures cut GPU memory. WebP shrinks the download, but the GPU still unpacks it to full size, while KTX2 stays compressed on the GPU. Use --texture-compress ktx2 when memory is tight, which it is on phones.

• Decimate first. The docs warn that optimize defaults may not suit every scene, so bring a million-triangle mesh down in Blender or with a simplify option (gltf-transform optimize -h lists them) before you ship it.

In Khronos’ own example, KTX2 barely changed the file size but cut GPU memory by about 82% (KTX artist guide).

7. Load it in Three.js

Three.js loads a Meshopt- or Draco-compressed GLB with KTX2 textures like this (GLTFLoader docs, three.js example).

import { GLTFLoader } from 'three/addons/loaders/GLTFLoader.js';
import { DRACOLoader } from 'three/addons/loaders/DRACOLoader.js';
import { KTX2Loader } from 'three/addons/loaders/KTX2Loader.js';
import { MeshoptDecoder } from 'three/addons/libs/meshopt_decoder.module.js';

const draco = new DRACOLoader().setDecoderPath('/draco/');
const ktx2 = new KTX2Loader().setTranscoderPath('/basis/').detectSupport(renderer);
const loader = new GLTFLoader().setDRACOLoader(draco).setKTX2Loader(ktx2).setMeshoptDecoder(MeshoptDecoder);

const { scene: chest } = await loader.loadAsync('/models/chest-web.glb');
world.add(chest);

Copy the Draco decoder files from three/examples/jsm/libs/draco/gltf/ and the Basis transcoder from libs/basis/ into your public folder, and reuse one loader for every model.

If you’ve got twenty coins on screen, they should cost one draw call, not twenty. InstancedMesh draws one geometry and material many times (Three.js docs), and renderer.info.render.calls tells you how many draw calls a frame costs.

Measure the model on your cheapest target phone

No standard sets a triangle budget, so measure on the cheapest phone you want to support. Two numbers from people who build for the web make a decent first target, though. Don McCurdy, the author of glTF Transform, suggests “something like <100 draw calls and <100,000 vertices if you can” (three.js forum). PlayCanvas calls 100 to 200 draw calls “a rough target for low end mobile devices” (PlayCanvas).

Textures actually bite harder than triangles. One 4096 by 4096 RGBA texture takes about 64 MB of GPU memory before mipmaps.

So start with textures 1024 or 2048 pixels square, unless the camera sits right on the object.

Keep the licenses and disclose generated assets

• Steam has a content survey where you disclose pre-generated AI content that ships with the game (Steam).

• itch.io strictly enforces AI disclosure on asset pages and may delist untagged ones (itch.io).

• Copyright is its own question. The US Copyright Office says AI output is protected only where a human supplied enough of the expressive choices (USCO), and luckily your edits, cleanup and composition do count. Keep the concept images and notes on what you changed.

Put the model in a game

How to make a 3D game with AI builds the Three.js game around your models, and MCP servers for game development lets your agent drive Blender for the cleanup. If you’d rather end up with Gaussian splats than a plain mesh, our splat game write-up shows how we got there from two-view drawings in Blender 5.3.

For 2D sprites, textures and style, see how to make game assets with AI. Or browse the Three.js games in the catalog to see what other people built.