Season 10 · Article 08

Generative 3D Meshes

Text/image → 3D model. 2025's breakthroughs: Trellis, Rodin, Meshy, Stable Fast 3D. The pipeline + clean-up steps + how to actually ship the output.

1. Tool comparison

Tool	Input	Output	Speed	Best for
Rodin Gen-1.5	Text / image	.glb with PBR	~60s	Game-ready props
Meshy	Text / image	.glb, .fbx, .usdz	~60s	General use
Trellis	Image	Mesh + radiance field	~30s	High quality research
Stable Fast 3D	Image	.glb	~2s	Speed-sensitive
Scenario	Text + style	.glb	~60s	Consistent packs

2. Typical quality

• Shape: 90% accurate to prompt. Complex details (fingers, small features) wobble.
• Topology: usually isotropic mess. Bad for rigging, OK for static props.
• Textures: PBR with color + normal, sometimes roughness. Quality varies.
• UVs: automatic, often non-ideal. Seams visible at steep viewing angles.

3. API example (Meshy)

const r = await fetch('https://api.meshy.ai/v2/text-to-3d', {
  method: 'POST',
  headers: { Authorization: 'Bearer ' + API_KEY, 'Content-Type': 'application/json' },
  body: JSON.stringify({
    prompt: 'a wooden treasure chest, medieval fantasy',
    art_style: 'realistic',
    negative_prompt: 'low quality, blurry'
  })
});
const { result } = await r.json();  // job ID
// Poll until done:
const model = await poll(result);
// Download model.model_urls.glb

4. Clean-up pipeline

# 1. Load in Blender
# 2. Retopologize if rigging needed (QuadRemesher, RetopoFlow, Instant Meshes)
# 3. UV unwrap (Smart UV Project → Pack Islands)
# 4. Bake textures onto new UVs
# 5. Fix PBR (roughness often too smooth)
# 6. Export clean glTF

# Then for web:
gltf-transform optimize input.glb output.glb
gltf-transform uastc output.glb --filter "*normal*"
gltf-transform etc1s output.glb --filter "*"

5. When it's "good enough" as-is

• Background props (chairs, crates, rocks).
• Concept-art stand-ins.
• Inventory icons (render to 2D).
• Prototypes before commissioning real art.

6. When it needs polish

• Hero characters: always.
• Animated objects: topology cleanup required.
• Tight geometry (doors, mechanisms): Blender modeling.
• Brand-specific style: fine-tune a LoRA on your IP.

7. Image-to-3D workflow

1. Artist sketches concept in 2D (Midjourney / Photoshop).
2. Upload sketch to Rodin / Meshy.
3. Receive 3D model matching the art direction.
4. Blender polish if needed.
5. gltf-transform compress.
6. Ship.

8. Cost

• Meshy / Rodin: $0.10-$0.50 per generation.
• Cheap enough to iterate 10× per asset.
• Save only the best output.

9. Takeaways

• Text/image → 3D is real in 2026. Output usable for backgrounds/props.
• Rodin/Meshy: 60s per model, $0.10-0.50.
• Topology usually bad — fix if rigging.
• Image-to-3D faster workflow than text.
• Always pipe through gltf-transform before shipping.