2026-09-06

Turn a Character Illustration into a Game-Ready 3D Model — T-Pose First

Single-image 3D generators got good fast. Upload a picture of a treasure chest and about two minutes later you have a textured GLB, and it genuinely looks like that chest. Then you try it on your game's main character and the result is unusable — not because the model is bad, but because of two things that have nothing to do with generation quality: the pose the art was drawn in, and the mesh that comes back.

This post is about the three tools that fix both, and the order to run them in.

Why the character comes back broken

Character art is almost always an action pose. Sword raised, one leg forward, cloak flaring, weapon gripped in the hand.

A 3D generator sees one image, and what it does with that pose is fuse it into the geometry. The sword becomes part of the hand — one continuous surface, with no seam to separate them. The torso behind the raised arm was never visible in the reference, so it is invented. The bent elbow is baked into the mesh, which means a skeleton can never straighten it: rig that model and you get an arm that bends from an already-bent starting point.

None of this is fixable downstream. You cannot un-fuse a weapon in a retopology step, and you cannot un-bend an elbow that is geometry rather than pose. The fix has to happen while the character is still a 2D image.

Step 1 — T-pose & parts

The T-pose splitter takes your reference and redraws the character standing in a T-pose, with the props lifted off and drawn separately — one image per asset.

It starts with an analysis pass, and the analysis is free. A vision model reads the image and reports what it found: whether the pose is already a rest pose, what body plan the character has (humanoid, quadruped, avian, serpentine, mechanical), which props it is carrying or wearing, how each one is attached, and how much of the body is hidden behind them. Nothing is generated and nothing is charged until you have seen that plan and agreed with it.

Then you choose how finely to split.

Coarse takes off only what is held in the hands. Armour, cloak and hair stay on the body.

Standard also takes off gear that is clearly strapped or hung on — pauldrons, backpacks, cloaks, quivers.

Modular takes off hair, helmet and armour pieces too, leaving the body as a bare base you can dress. Pick this one if you plan to swap or sell equipment, with the caveat that every piece you remove is more hidden surface the model has to invent underneath.

Don't split keeps the character whole, weapon and all, and only fixes the pose.

Two things worth knowing. If the reference is already a rest pose, the body is not redrawn and not charged. And you can drop in up to ten characters at once — a whole cast analysed in one pass, and you pay only for the pieces you actually generate.

The output is a set of transparent PNGs. They are already useful as 2D assets, and each one has a button that sends it straight to Image → 3D.

Step 2 — Image to 3D

Four models, and the choice matters more than it looks: the same job costs 6 credits or 38 depending on which one you pick.

Tripo3D H3.1 multiview (6 credits) needs two to four views of the same subject. That requirement is really a feature — the back and sides come from your images instead of a guess. It outputs quads directly.

Seed3D 2.0 (10 credits) is the single-image default. A solid all-rounder, but with one image the back is invented rather than observed.

Hunyuan 3D V3 (11 credits) costs almost the same as Seed3D and does considerably more: up to four named views (back, left, right), PBR textures, quad topology and a face-count target.

Meshy V7 (38 credits) is by far the most expensive. Try Tripo first.

If the model offers a face limit, set it. Left on auto, Hunyuan V3 returned a 500,000-triangle mesh in a 44 MB file — a fine high-poly to bake from, and unusable in an engine as it stands. Which is what the next step is for.

Step 3 — Mesh cleanup

Generated meshes come back as a uniform soup of tiny triangles: the same density on a flat cape as on a face, no edge loops anywhere, nothing that deforms sensibly. Mesh cleanup rebuilds them.

It is free, and it runs on your machine — a Web Worker in your browser, not our servers. That is why we can give it away.

Two modes. Triangle decimate is instant, keeps your UVs and textures, and leaves the topology messy — right for props and set dressing. Quad retopology rebuilds the surface as an all-quad mesh with real edge flow. Slower, and worth it for anything that has to deform: quads subdivide cleanly and fold at joints the way an artist would have built them.

A quad rebuild carries no UVs of its own, so unwrapping runs by default and the base colour is carried across automatically — the model never turns grey on you. Normal-map baking is a separate button, and it is what makes 4,000 triangles read like the 500,000 you started with.

One number to remember: 5,000 quads for a character. Below that, a limb gets one or two edge rings at each joint, which is not enough to skin — the elbow collapses when it bends. The tool warns you when your target falls under that line. Props and background pieces are fine much lower.

What a character actually costs

One character split into a body and two props: 3 credits for the body (drawn at 2k, because a hand is only a few dozen pixels in a full-body frame) plus 2 credits per prop — 7 credits. Three Seed3D conversions on top, 30. Mesh cleanup, nothing. Call it 37 credits for a fully separated, modular character, or 13 if you keep it whole and skip the split.

The analysis is free, a part that fails is refunded automatically, and a reference that is already in T-pose is not charged for the body.

Where the pipeline stops today

At the end of step 3 you have a clean, textured, correctly unwrapped mesh at a face count an engine will accept. You do not yet have a rigged character — skinning it and putting animation on it are the next two tools, and they are still behind a coming-soon gate.

If you want something moving today, one route already runs end to end: rig the mesh in Mixamo, which is free and takes a couple of minutes, then bring the rigged GLB into 3D → Sprite and render it out as a 4, 8 or 16-direction sprite sheet. That path is finished — and for a 2D game, it was the destination anyway.