3D asset generation for games, AR & e-commerce

Producing a production-ready 3D model by hand takes hours to days of specialist work, which is why only a fraction of product catalogues or game prop libraries exist in 3D. Generative 3D models turn a text prompt or one to a few photos into a textured mesh in seconds to minutes. The output is usually a good starting point (prototypes, background props, long-tail products) rather than a finished hero asset, so cleanup and human review remain part of the workflow.

Problem & value

• Today: 3D assets come from modelling in Blender/Maya/ZBrush, CAD conversion, photogrammetry or outsourcing. Cost and lead time limit how many SKUs get an AR/3D view and how many props a small game team can afford.
• AI value: fast concept-to-mesh iteration, rapid greyboxing and prop libraries, 3D/AR views for long-tail catalogue items, and variants (colour, material) without new shoots.
• Benefit evidence: platform vendors report that interactive 3D/AR views on product pages can lift engagement and conversion (Shopify documents 3D models and AR for products [4]; Apple and Google provide AR Quick Look and Scene Viewer [5]). Such figures are vendor-reported and vary by category; measure with your own A/B test. No independent, general uplift number is claimed here.

Variants

• Game and indie prop generation: text/image-to-mesh for background props, kitbashing, concept blockouts; game engines (Unity, Unreal, Godot) import GLB/FBX. Dialogue, levels and NPCs → GEN-013.
• E-commerce / AR product viewers: photo-to-3D for products, delivered as GLB (web, <model-viewer>) and USDZ (iOS AR Quick Look) [5]. For accurate real-object capture use scanning instead → GEN-025.
• Character and avatar assets: auto-rigging and animation of generated characters → GEN-022.
• Product ideation: early-stage form studies before CAD → GEN-014. Generated meshes are not CAD (no parametric, tolerance-accurate geometry) and must not be used for manufacturing.
• Synthetic training data: generated 3D assets rendered into scenes for vision models → GEN-020, ROBO-010.
• Private / hobby: 3D printing of figurines, custom game mods, AR décor previews.
• Explorable 3D world and environment generation (emerging variant): generative world models create navigable scenes from text, images or video. World Labs launched its Marble product in late 2025 (persistent, exportable 3D worlds, incl. Gaussian splats and meshes) [7], and Google DeepMind's Genie 3 is a real-time interactive world model that is a research preview [8]. Production evidence is still thin: use for previs, level blockouts and mood boards, not shipped content. Related: GEN-013 (game content), GEN-025 (capture-based scenes), ROBO-010 (simulation for robot training). It can be promoted to its own use case once production deployments exist.

How it works

1. prompt / product photos / concept art
2. (optional) image generation or clean-up: isolate object, neutral background, multi-view synthesis
3. 3D generation model (image/text-to-3D: latent 3D diffusion/flow or multi-view reconstruction)
4. mesh extraction + PBR texture synthesis (albedo, normal, roughness/metallic)
5. automatic post-processing: retopology / decimation, UV unwrap, LODs, texture compression
6. QA: automated checks (poly budget, watertightness, scale, material sanity) + human review in DCC tool
7. export (GLB/USDZ/FBX)
8. engine, CDN, PIM/DAM or e-commerce platform

• Core models: most current systems generate a 3D latent (structured voxel/latent representation or multi-view images lifted to 3D) from one or more images; text-to-3D typically goes via a generated image first. Texture/PBR generation is a separate stage in many pipelines.
• Human-in-the-loop: artists review geometry, topology and material accuracy; for e-commerce, product owners must confirm that colour, proportions, logos and features match the real product (misleading representation risk).

Data preparation

3D generation needs input material of good quality: product photos from several angles, concept art, CAD or existing meshes, brand style guides and technical specs (polygon budget, texture size, format such as glTF/USDZ). In an SMB these live in the PIM or ERP (article master data), on file shares, with suppliers or on photographers' drives. Typical gaps are inconsistent photo backgrounds, missing views, no SKU-to-image mapping, missing dimensions, unclear image rights and no defined target specs.

Preparation steps: (1) select a pilot range of 20 to 50 articles; (2) standardise photography (neutral background, consistent lighting, multiple angles), possibly with a simple turntable; (3) map images to SKUs and add dimensions and materials from the ERP/PIM (PREP-005 for extraction, PREP-010 for tagging); (4) define target specs per channel (web viewer, AR, game engine); (5) clarify rights with suppliers and photographers; (6) collect gold-standard models for QA (PREP-016). Tools: background removal, Blender for clean-up, a DAM. Share of effort: about 25 to 35 percent, driven by photography and master data quality. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path A, because hosted image-to-3D services and e-commerce viewers deliver usable assets at per-asset prices and self-hosting needs HW3.
• Enterprise: Path B, because catalogues with thousands of SKUs need API integration into PIM/DAM pipelines and consistent technical specs.
• Developer: Path C, because open image-to-3D models (TripoSR, TRELLIS class) can be integrated in pipelines on own GPUs without per-asset fees.
• Private: Path A, because hosted tools suffice for hobby game and AR projects, as HW3 is rarely available.

Path A: Buy

• Generative 3D platforms: Meshy, Tripo (Tripo3D), Rodin (Hyper3D / Deemos), Luma (Genie), Sloyd (parametric game props), Scenario and Layer for game-art pipelines; Adobe Substance 3D tools for texturing and materials. Most vendors are US/Asia-based; check EU data processing terms. EU: Sloyd (Sweden).
• E-commerce 3D/AR delivery: Shopify 3D models and AR [4], Apple AR Quick Look and Google Scene Viewer [5], plus 3D platforms such as Sketchfab (hosting/viewers).
• Choose A when: low volume, no strict visual-fidelity needs, need results within days.

Path B: Build on cloud APIs

• Meshy, Tripo and Rodin offer REST APIs for text/image-to-3D with GLB/FBX/USDZ output; orchestrate with n8n, queues (async jobs) and webhooks. Use an image model (e.g. FLUX, Gemini image, GPT image) to produce clean multi-view or isolated product renders first, and a VLM to check the result against the source photos.
• Post-process with Blender in headless/Python mode (decimate, UV, LOD) and glTF tooling (gltf-transform, gltfpack, Draco/meshopt compression, KTX2 textures).
• Choose B when: you need a catalogue-scale batch pipeline with PIM/DAM integration.

Path C: Self-host (open source)

• Open-weight models: Tencent Hunyuan3D (2.x incl. texture generation) [1], Microsoft TRELLIS (structured 3D latents) and its successors [2], Stable Fast 3D / Stable Point Aware 3D (Stability AI) [3], TripoSR. Check each licence (several have territory, size or commercial-use restrictions; Hunyuan3D licence excludes the EU/UK/South Korea in some releases, so verify before EU use).
• Stack: PyTorch + CUDA, ComfyUI 3D nodes, Blender for post-processing, gltf-transform.
• Hardware: HW2 (24 GB) runs most image-to-3D models at reduced resolution; HW3 is comfortable for high-resolution geometry plus texture generation and batch throughput.
• Choose C when: IP or data must stay in-house (unreleased products), or volume makes API cost high.

Path D: Custom training

Fine-tune (LoRA or full) on a licensed, in-house library of assets in a consistent style (e.g. a game's art direction or a brand's product category). Needs hundreds to thousands of clean, correctly licensed meshes with textures. Worth it only for studios with a stable style and engineering capacity.

State of the art (2026)

• Image-to-3D models produce usable textured meshes in seconds to minutes, with PBR materials increasingly standard. Open-weight models (Hunyuan3D, TRELLIS family) are close to commercial services for single objects [1][2].
• Limits and failure modes: wrong or hallucinated back sides from a single view, thin or hollow structures, text/logo distortion, inconsistent scale, messy triangulated topology that is poor for animation/deformation, baked lighting in textures, and no engineering tolerances. Hard-surface precision parts and branded packaging still need CAD or scanning.
• World models: Marble [7] and Genie 3 [8] show scene-level generation; consistency over long trajectories, scene scale, collision/physics and editability are open problems.
• Standards: glTF 2.0 (Khronos) and OpenUSD (Alliance for OpenUSD) are the interchange formats [6].

Shared resources & combinations

• Same stack (3d-gen + image-gen + 3d-vision): GEN-001 (product images and renders), GEN-025 (capture-based 3D), GEN-014 (product ideation), GEN-009 (architecture visualisation), GEN-022 (rigging and animation).
• Hardware: one HW3 GPU server can host image generation and 3D generation jointly.
• Downstream: GEN-013 game content, ROBO-010 simulation and digital twins, GEN-020 synthetic data.

Effort, cost & KPIs

• Path A: days; subscription or credit-based pricing per generation.
• Path B: 1-3 months for a catalogue pipeline with QA and integration.
• Path C: 2-4 months incl. model evaluation, licence review and Blender automation; GPU server.
• KPIs: share of generated assets accepted without rework, artist hours per asset, cost per asset, time to publish a SKU in 3D, 3D/AR view engagement and conversion (A/B), return rate for 3D-enabled products, defect rate in product-fidelity QA.

People & change

Involve management for the business case, the e-commerce or product lead as process owner, and the 3D artists, photographers and marketing staff who will review results. IT is needed for the PIM/DAM integration and, in Path C, GPU hardware (HW3); legal for image rights; suppliers if their product data or CAD files are used.

Works council: no in most cases, as the output is product content and not employee monitoring; §87(1) no. 6 BetrVG applies only if the platform logs individual performance. §90(1) no. 3 BetrVG requires timely information if the roles of in-house 3D artists change through AI. Data protection: normally no personal data; involve the DPO only if images show people, and conclude an AVV (Art. 28 GDPR) with cloud vendors.

AI literacy: support staff competence under Art. 4 AI Act with a hands-on workshop on image-to-3D quality checks, mesh clean-up and licence terms (GUIDE-006). Start with a pilot group of two or three artists on one product range, measure time per asset and rework rate against manual modelling, then decide go or no-go (GUIDE-009). Communicate that AI produces first drafts and artists keep quality control; every asset is checked for shape accuracy, since a wrong 3D model can mislead customers about a product. Typical concerns are job loss for artists and artefacts in geometry; address them by retraining towards review and finishing work, and by publishing only approved models. See GUIDE-008.

Risks, regulation & compliance

• EU AI Act: typically limited risk. Art. 50(2) requires providers of generative systems to mark synthetic output machine-readably; deployers of deepfake-like content must disclose it (Art. 50(4)). Pure 3D props are rarely deepfakes, but photoreal avatars of real people could be. No Art. 5 prohibition applies for ordinary asset generation; generating likenesses of real persons without consent must be avoided [9].
• Copyright and licensing: training-data provenance and output ownership differ per vendor; prefer vendors with commercial-use terms and IP indemnities; keep records of prompts and source images. Use of protected designs as input (competitor products, brand characters) can infringe design/copyright and trademark rights. The EU text-and-data-mining opt-out (DSM Directive Art. 4) and AI Act GPAI copyright obligations affect model providers.
• GDPR: mostly no personal data; applies if photos contain people or faces, or if avatars resemble real persons. Check vendor processing and data-retention terms.
• Consumer law: 3D/AR product views must accurately represent the product (UCPD, misleading presentation); label generated visuals where they could mislead.
• Mitigations: human fidelity check against the real product, licence inventory per model, provenance metadata (C2PA where supported), no CAD/manufacturing use of generated meshes.

Getting started

1. Pick 20 products or props across easy and hard categories (reflective, transparent, thin, branded).
2. Run them through 2 commercial services and 1 open-weight model; compare geometry, texture, polygon cost.
3. Define acceptance criteria (poly budget, fidelity checklist) and an automated glTF validation step.
4. Pilot on a few product pages with an A/B test, or in a prototype level for games.
5. Review licences and AI Act labelling before scaling; decide between buy, API and self-host.

Sources

1. Tencent Hunyuan3D (GitHub) - https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1
2. Microsoft TRELLIS (GitHub) - https://github.com/microsoft/TRELLIS
3. Stability AI Stable Fast 3D - https://github.com/Stability-AI/stable-fast-3d
4. Shopify Help Center - 3D models and AR for products - https://help.shopify.com/en/manual/products/product-media/3d-models
5. Apple Developer - AR Quick Look - https://developer.apple.com/augmented-reality/quick-look/
6. Khronos Group - glTF 2.0 specification - https://www.khronos.org/gltf/
7. World Labs - Marble - https://www.worldlabs.ai/blog/marble-world-model
8. Google DeepMind - Genie 3 - https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/
9. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
10. Meshy API documentation - https://docs.meshy.ai/