Tencent Hunyuan3d

Tencent Hunyuan3D

Hunyuan3D is Tencent's family of generative 3D models. Its defining trait is a two-stage, shape-then-texture pipeline: one model generates untextured geometry from an image (or text), a second model paints materials onto that geometry. The family spans open-weight releases you can self-host and closed tiers available only through a hosted API. Choosing correctly between those two tracks — and reading the license before you ship anything commercial — is the core production judgment this skill supports.

All version, license, hardware, and access facts below were verified 2026-07-10. These are volatile; re-verify against first-party repos before relying on them. Facts are labeled [fact] (documented in a primary source), [first-party claim] (Tencent's own benchmark/marketing, not independently reproduced), or [heuristic] (production judgment, not a documented guarantee).

When this skill applies

Use it when the request involves:

• generating a 3D object/asset from an image or text prompt;
• generating a 3D scene/world (panorama, explorable environment) — the HunyuanWorld / HY-World branch;
• deciding self-host vs hosted API, or which Hunyuan3D tier fits;
• license/compliance questions (commercial use, Territory, MAU, attribution);
• GPU sizing for local inference or ComfyUI setup;
• reviewing or post-processing Hunyuan3D output (topology, UVs, PBR maps).

Do not use it for 2D image generation (that is HunyuanImage), video (HunyuanVideo), or the Hunyuan text LLMs (Hunyuan/Hy3) — those are separate model families that happen to share the "Hunyuan" brand.

The version landscape (verified 2026-07-10)

Two tracks. Getting the track right is the first decision, because it determines cost model, control, and legal footing.

Track A — Open-weight, self-hostable (download from GitHub + Hugging Face)

[fact] Hunyuan3D-2.1 is the reference open baseline for object generation. It is the version to self-host when you need PBR and want everything (weights and training code) under an open license. Sources: GitHub Tencent-Hunyuan/Hunyuan3D-2.1, arXiv 2506.15442.

[fact] FlashVDM accelerates the shape decoder: Tencent reports a Lightning Vectset Decoder with >45x speedup and Progressive Flow Distillation sampling in as few as 5 steps ([first-party claim] on the speedup magnitude). Enable via the turbo checkpoints / --enable_flashvdm.

Track B — Closed, hosted API only (no downloadable weights, verified 2026-07-10)

[fact] The 2.5 / PolyGen / 3.x tiers are closed-source: accessible only through hosted APIs, not as weights. The Tencent Cloud International Station and Hunyuan3D API for overseas users launched Nov 26 2025. If a task says "use the open weights" and also "use 3.0/PolyGen," that is contradictory — surface it. Sources: Tencent Cloud doc 1284/75539; hunyuan3d.cc version notes; Tencent press.

[heuristic] PolyGen exists specifically to solve the biggest weakness of the open 2.x meshes: topology. If your blocker is that self-hosted meshes are dense triangle soup, PolyGen (hosted) or a manual retopology pass is the answer — not a different 2.x checkpoint.

The license — read this before any commercial deployment

This is the single most consequential production fact in the family. The open weights are not MIT/Apache. They ship under the Tencent Hunyuan 3D Community License Agreement (versioned per model, e.g. "TENCENT HUNYUAN 3D 2.1 COMMUNITY LICENSE AGREEMENT"). [fact] Verified 2026-07-10 from the LICENSE files in Tencent-Hunyuan/Hunyuan3D-2 and Hunyuan3D-2.1.

Three restrictions that routinely surprise people:

1. Territory exclusion [fact]. The license grant is worldwide excluding the European Union, the United Kingdom, and South Korea. Use "outside the Territory" is prohibited. A team in Berlin, London, or Seoul is not licensed to use the open weights under this agreement. This is a hard legal blocker, not a preference — flag it whenever a user's location or target market is in those regions.

2. 1 million MAU commercial gate [fact]. If, on the model's release date, the products/services made available by or for the licensee had > 1,000,000 monthly active users in the preceding calendar month, the licensee must request a separate license from Tencent (email hunyuan3d@tencent.com with company name, sector, and use case). Below 1M MAU, commercial use is permitted within the Territory at no charge.

3. Prohibited uses [fact]. Among ~20 clauses: no use to improve competing AI models; no military use; no generating content that harms minors; no election disinformation; no malware; no undisclosed AI-generated content; no "high-stakes automated decisions affecting an individual's safety, rights, or wellbeing."

Attribution [fact]: redistribution requires a NOTICE file ("Tencent Hunyuan 3D 2.1 is licensed under the Tencent Hunyuan 3D 2.1 Community License Agreement, Copyright © 2025 Tencent."), and products are expected to be marked "Powered by Tencent Hunyuan."

[heuristic] The hosted API does not erase these questions — it changes which document governs. Tencent Cloud and third-party hosts (fal, Replicate, etc.) each impose their own Terms of Service. Do not assume "hosted = unrestricted commercial"; check the specific host's ToS, and if a user is in EU/UK/KR, using a hosted API is often the cleaner route than self-hosting weights they aren't Territory-licensed for — but confirm the host permits it. When a commercial deployment is on the line and the facts are ambiguous, tell the user to get their own legal read; do not assert a clearance you cannot source.

How generation actually works (the two-stage pipeline)

[fact] (arXiv 2506.15442, Hunyuan3D-2.1) The pipeline is deliberately modular — you can run either stage alone.

Stage 1 — Shape (Hunyuan3D-DiT + ShapeVAE):

• Input: a single image (background removed), or multi-view images (2mv), or text (which is routed through text-to-image first).
• A flow-based diffusion model over a vectset latent produces a signed-distance / occupancy field; marching cubes extracts a watertight triangle mesh.
• octree_resolution is the primary geometry-density / quality knob.
• Output: an untextured mesh (GLB/OBJ, trimesh object).

Stage 2 — Texture / material (Hunyuan3D-Paint):

• Input: the mesh from Stage 1 (Hunyuan3D's own, or your own imported mesh).
• A mesh-conditioned multi-view diffusion model renders the object from several viewpoints (training used 512×512 views) and generates view-consistent material maps, then bakes them into UV textures.
• [fact] 2.1 produces PBR maps under the Disney Principled BRDF: albedo (light-free, via an illumination-invariant training strategy), metallic, and roughness. It uses a 3D-aware RoPE for cross-view consistency. 2.0's Paint output was closer to baked diffuse color; 2.1 is the first open PBR.

[heuristic] Because Stage 2 accepts arbitrary meshes, a strong workflow is: use Hunyuan3D only for texturing an existing (hand-modeled or retopologized) mesh, or only for shape and texture elsewhere. Treat the two stages as independent tools.

Input strategy: image beats text for control

[heuristic] Text-to-3D internally does text→image→3D, so you surrender control at the image step. Feed your own reference image whenever geometry matters. Good shape inputs:

• single, clearly separated subject, whole object in frame, not cropped;
• plain / removable background (the pipeline runs background removal; a busy background degrades it);
• front-facing, roughly orthographic framing, even lighting, minimal harsh shadow (shadows can be baked into geometry);
• for objects where the back or sides matter, use 2mv with multiple views — a single front image forces the model to hallucinate the unseen faces, and it often gets concavities, back detail, and thin structures wrong.

Self-hosting: hardware and setup (verified 2026-07-10)

[fact] VRAM footprints (from the model READMEs — note the two open lines report slightly different numbers, so plan for the larger):

[heuristic] Practical reading: a 24 GB card (RTX 3090/4090) runs 2.1 shape and texture if you offload between stages rather than holding both in memory at once; running both simultaneously wants ~29 GB, i.e. a 32–48 GB card or sequential execution with model offloading. 2mini is the choice for 8 GB consumer cards (shape only). The distilled Turbo + FlashVDM path is for fast iteration, not maximum fidelity.

[fact] Reference environment for 2.1: Python 3.10, PyTorch 2.5.1 + CUDA 12.4. Setup compiles a custom rasterizer and a DifferentiableRenderer / mesh painter (C++/CUDA extensions) — budget time for the build step and a matching CUDA toolchain. Real-ESRGAN weights are fetched separately for texture upscaling. A Gradio app (gradio_app.py) and a REST endpoint (/generate, base64 image in → GLB out) ship in the repo. Source: Hunyuan3D-2.1 README.

ComfyUI

[fact] (docs.comfy.org, verified 2026-07-10) ComfyUI has native support for the shape/geometry stage only — workflows for Hunyuan3D-2, 2mv, and 2mv-turbo generate untextured GLB into ComfyUI/output/mesh. Texture/material generation is not in the native nodes. For PBR texturing inside ComfyUI, use a community wrapper (e.g. ComfyUI-Hunyuan3DWrapper). Plan the texture stage separately if you are ComfyUI-only.

Hosted access routes (verified 2026-07-10)

• Tencent Cloud "Hunyuan 3D APIs" (doc 1284/75539) — first-party; International Station live since Nov 26 2025. Backs the closed 2.5 / PolyGen / 3.x tiers.
• Hugging Face Space tencent/Hunyuan3D-2.1 — a free demo of the open model, good for a quick capability check, not for production volume.
• Third-party hosts — fal.ai, Replicate, 3D AI Studio, Atlas Cloud, and others wrap Hunyuan3D behind their own APIs and billing.

[fact] Tier/pricing signals seen across hosts (indicative, varies by provider): "Rapid" tiers finish in 2–3 min with fixed mid-range polygon budgets and ~1K textures; "Pro" tiers invoke the full model with configurable polygon counts (40K–1.5M) and up to 4K PBR. Per-model prices around $0.02–$0.38 were observed; one credit scheme charged Pro 60 credits (+20 PBR, +20 multi-view, cap 100) vs Rapid 35 (+20 PBR, cap 55). Treat all prices as volatile — quote the host's live pricing, never these numbers, to a user.

Mesh output characteristics and required post-processing

[fact] Open 2.x meshes come from marching cubes over a diffusion field. That means they are watertight but also dense, uniformly-triangulated, non-quad, and lacking artist topology (no edge loops). UVs are auto-generated for texturing, not laid out for hand-editing.

[heuristic] Consequences for a production pipeline — plan these passes:

• Retopology. Raw meshes are unsuitable for rigging/animation or clean deformation. Retopologize (manual, an auto-retopo tool, or route the asset through PolyGen, which was built for exactly this) before animating.
• Decimation. For real-time (game/AR/web) budgets, decimate the dense mesh to a target triangle/polygon count; expect to bake normal maps from the high-res mesh to preserve detail.
• UV re-layout if a human needs to paint or edit textures, rather than only consuming the auto-baked PBR set.
• Scale / orientation / origin normalization — generated assets rarely come in your engine's units, up-axis, or pivot convention.
• PBR map validation — confirm albedo, metallic, and roughness channels all exported and are separated correctly (some export paths flatten to diffuse).

Reviewing generated assets — quality checklist

[heuristic] Before accepting an asset, check:

1. Silhouette & proportions from all sides (orbit it) — front-only inputs frequently produce a plausible front and a mangled back.
2. Unseen-face invention — concavities, undercuts, and thin features (straps, handles, hair) are the common failure zone; verify they exist and are solid.
3. Floaters / disconnected shells / non-manifold artifacts from marching cubes.
4. Geometry density vs need — is it far denser than the use case warrants (decimate) or too coarse (raise octree_resolution / use 2.5/3.0)?
5. Topology — acceptable for a static prop as-is; not acceptable for animation without retopology.
6. PBR correctness — albedo should be lighting-free (no baked highlights or cast shadows); metallic/roughness should read physically (metal vs dielectric).
7. Texture seams and stretching at UV boundaries; resolution adequacy (1K vs 4K depending on tier and camera proximity).
8. Scale/orientation normalized to target engine.

Production decision: open self-host vs hosted commercial

[heuristic] Self-hosting an open 2.x model wins when:

• you generate high volume — marginal cost per asset approaches zero vs per-call API fees;
• data cannot leave your environment (unreleased IP, private client assets);
• you need PBR at no per-asset cost and full pipeline determinism/integration;
• you want the training code (2.1 ships it) to fine-tune on a domain;
• and — the gating condition — you are inside the Territory (not EU/UK/KR) and under 1M MAU, or have Tencent's commercial license.

[heuristic] A hosted API wins when:

• you need peak fidelity or resolution (2.5, 3.0) or clean topology (PolyGen) that the open weights don't provide;
• you have no suitable GPU or don't want to maintain the CUDA/build toolchain;
• volume is low or spiky (pay-per-use beats idle GPU);
• you are in EU/UK/KR and self-hosting the open weights is not Territory-licensed (use a host that permits your use — and still check its ToS).

[first-party claim] On Tencent's own benchmarks, Hunyuan3D-DiT (2.1) edges comparable open models on shape metrics (Uni3D-T 0.2556 vs TripoSG 0.2506, Step1X-3D 0.2554) and 2.1's Paint improves texture CLIP-FID to 24.78 from 2.0's 26.44. These are vendor-reported; treat "best" claims skeptically and, for a real selection, run your own assets through candidates rather than trusting a leaderboard.

Complete example (labeled example, not a required formula)

Intent: A studio outside the EU/UK/KR needs ~500 stylized prop meshes for a mobile game, textured, on an in-house pipeline, no per-asset cloud fee, budget for one 24 GB GPU.

Decision & reasoning:

• Track: self-host — high volume + fixed GPU makes marginal cost ~0, and 500 assets would be a recurring API bill.
• License: studio is in-Territory and well under 1M MAU → free commercial use; add the NOTICE file and "Powered by Tencent Hunyuan" mark. [fact]
• Model: Hunyuan3D-2.1 — need PBR for the game's lighting, want the open weights + code. On a 24 GB card, run shape then texture sequentially with offload (combined would want ~29 GB). [fact]/[heuristic]
• Input: provide a front reference image per prop (concept art), background removed; for props with important back detail, capture two views and use 2mv. [heuristic]
• Iterate with the Turbo/FlashVDM checkpoint to lock silhouette and framing cheaply, then re-run the full model for the final. [heuristic]
• Post-process each asset: decimate to the mobile poly budget; bake normals from the dense mesh; retopologize any prop that will deform/animate (or send those few through hosted PolyGen); validate the albedo/metallic/roughness set; normalize scale/up-axis to the engine. [heuristic]

Expected result: watertight, PBR-textured props at controlled poly counts, no per-asset fee, license-clean.

Likely failure modes: hallucinated back faces on single-view props (mitigate with 2mv); dense triangle topology unusable for the few animated props (retopo/PolyGen); albedo with baked shadows if lighting in the reference was harsh (use flat, even reference lighting).

Variation: the same studio in Germany cannot use the open weights under the Community License at all — it must switch to a hosted API whose ToS permits the use, changing the economics and the whole plan. [fact]

Common failure modes (summary)

• Back/side geometry wrong from single-view input → use 2mv or multiple views.
• Topology unusable for animation → retopologize or use PolyGen; the open meshes are marching-cubes triangle soup by design. [fact]
• VRAM OOM running shape+texture together → run stages sequentially with offload, or drop to 2mini for shape. [heuristic]
• Baked shadows/highlights in "albedo" → the input reference had directional lighting; re-shoot flat. [heuristic]
• Assuming open = unrestricted → it is the Community License: Territory-limited (no EU/UK/KR), 1M-MAU gate, prohibited-use clauses, attribution. [fact]
• Assuming ComfyUI gives you textures → native ComfyUI is geometry-only; texture needs a community wrapper. [fact]
• Confusing tracks → 2.5/PolyGen/3.x are API-only; you cannot download their weights. [fact]

Primary sources (all verified 2026-07-10)

• GitHub Tencent-Hunyuan/Hunyuan3D-2 — 2.0 model zoo, variants, FlashVDM, VRAM.
• GitHub Tencent-Hunyuan/Hunyuan3D-2.1 — 2.1 weights + training code, setup, VRAM.
• LICENSE files in both repos — Tencent Hunyuan 3D 2.0 / 2.1 Community License (Territory, 1M MAU, prohibited uses, attribution).
• arXiv 2506.15442 — Hunyuan3D 2.1 technical report (two-stage architecture, PBR / Disney BRDF, benchmark numbers).
• arXiv 2506.16504 — Hunyuan3D 2.5 technical report.
• GitHub Tencent-Hunyuan/HunyuanWorld-1.0 and Tencent-Hunyuan/HY-World-2.0; arXiv 2507.21809 — world/scene models.
• docs.comfy.org Hunyuan3D-2 tutorial — ComfyUI native geometry-only support.
• Tencent Cloud "Hunyuan 3D APIs" (doc 1284/75539) and Tencent press releases — hosted API, International Station launch (Nov 26 2025).
• hunyuan3d.cc version notes; Tencent Hunyuan announcements — PolyGen (Jul 8 2025), 2.5, 3.0/3.1 timeline and closed-source status.