SupaVoxel
3D generation model · Tencent

Hunyuan 3D 2.5, Explained: What Changed From 2.0

Hunyuan 3D 2.5 is Tencent's April 2025 upgrade to its image-to-3D system, with a 10B shape model and PBR materials; this page covers what is new and how to use it. Want a textured model today? Try SupaVoxel's generator: typically 3 to 5 minutes, nothing to install.

  • SupaVoxel: 3 free credits
  • No GPU or install
  • STL, OBJ, USDZ, 3MF on paid plans

By ·Updated

Steampunk robot knight, weathered teal metal, shield and gears
ExampleSteampunk robot knight, weathered teal metal, shield and gears
Press Generate to see it in 3D
Image → 3D

Drop in a picture, get a 3D model

Upload an image
Drag & drop or click · PNG, JPG, WebP, HEIC · up to 10 MB
Or try an example

3 free credits · Google sign-in · no card

Apr 23, 2025
Hunyuan 3D 2.5 launched on Tencent's 3D platform
10B
Parameters in the largest 2.5 shape model
768 × 768
Maximum multi-view texture resolution at inference
3–5 min
Typical textured model on SupaVoxel's generator
01

What is Hunyuan 3D 2.5?

Hunyuan 3D 2.5 is a version of Tencent's Hunyuan3D image-to-3D system that turns a picture into a detailed, textured 3D model. Tencent launched it on April 23, 2025 as an upgrade to its online 3D creation engine and opened an API on Tencent Cloud the same day. A technical report followed on June 19, 2025.

It keeps the two-stage design of Hunyuan 3D 2.0: first a shape model builds an untextured mesh from your image, then a texture model paints it. What changed is the scale and the materials. The shape stage uses a new foundation model called LATTICE, whose largest version has 10 billion parameters, and the texture stage now outputs physically based (PBR) materials instead of a plain color map.

Tencent offered Hunyuan3D 2.5 on its own platform and API. Unlike 2.0 and 2.1, it lists no downloadable 2.5 checkpoint. Our Hunyuan3D overview covers the whole family; this page sticks to 2.5.

02

Hunyuan 3D 2.5 vs 2.0: what changed

The short answer: 2.5 has a much larger shape model, accepts four views, adds PBR materials and rigging, and is offered hosted, with no open checkpoint listed. Here is the comparison side by side.

Released

Hunyuan3D 2.0
Jan 21, 2025
Hunyuan 3D 2.5
Apr 23, 2025 (report Jun 19, 2025)

Shape model

Hunyuan3D 2.0
Hunyuan3D-DiT, 1.1B
Hunyuan 3D 2.5
LATTICE, largest version 10B

Shape input

Hunyuan3D 2.0
One image (multi-view via the separate 2mv model)
Hunyuan 3D 2.5
One image or four multi-view images

Materials

Hunyuan3D 2.0
Color texture
Hunyuan 3D 2.5
PBR: albedo plus metallic and roughness

Extras

Hunyuan3D 2.0
Local Gradio app; Blender add-on (Jan 27, 2025)
Hunyuan 3D 2.5
Automatic rigging and skin weights, mesh reduction templates on Tencent's platform

Weights

Hunyuan3D 2.0
Open, run it yourself
Hunyuan 3D 2.5
Hosted by Tencent; no 2.5 checkpoint listed

VRAM to run locally

Hunyuan3D 2.0
16 GB for shape plus texture
Hunyuan 3D 2.5
Not applicable
Hunyuan3D 2.0Hunyuan 3D 2.5
ReleasedJan 21, 2025Apr 23, 2025 (report Jun 19, 2025)
Shape modelHunyuan3D-DiT, 1.1BLATTICE, largest version 10B
Shape inputOne image (multi-view via the separate 2mv model)One image or four multi-view images
MaterialsColor texturePBR: albedo plus metallic and roughness
ExtrasLocal Gradio app; Blender add-on (Jan 27, 2025)Automatic rigging and skin weights, mesh reduction templates on Tencent's platform
WeightsOpen, run it yourselfHosted by Tencent; no 2.5 checkpoint listed
VRAM to run locally16 GB for shape plus textureNot applicable

The full 2.0 model zoo, VRAM figures and license are on our Hunyuan 3D 2.0 page.

03

Hunyuan 3D 2.5 specs

What Tencent published about Hunyuan 3D 2.5 in its launch announcement and technical report. The last row shows SupaVoxel's generator for comparison.

Maker
Tencent Hunyuan team
Input
One image or four multi-view images; text prompts on Tencent's platform
Shape model
LATTICE, a large diffusion model with guidance and step distillation for speed; largest version 10B
Geometry
Tencent advertises 1024-resolution geometry and 10 times the mesh faces of the previous version
Texture
Multi-view PBR maps (albedo, metallic-roughness) at up to 768 × 768 per view
Access
Tencent's online 3D creation engine and Tencent Cloud API; no open weights listed
Alternative: SupaVoxel's generator
Works in the browser, textured model in typically 3 to 5 minutes, downloads and STL / OBJ / USDZ / 3MF exports on paid plans
04

How the Hunyuan 3D 2.5 pipeline works

Tencent's technical report describes four steps from picture to textured asset. The structure is the same as 2.0; the models inside are new.

  1. 01

    Clean up the input image

    An image processor removes the background and resizes the picture so the object sits centered for the shape model.
  2. 02

    Generate the shape with LATTICE

    The LATTICE shape model, conditioned on one image or four views, produces an untextured mesh with sharp edges and smooth surfaces.
  3. 03

    Prepare the mesh for painting

    The mesh is processed to extract normals, a UV map and other geometry signals that the texture model uses as conditions.
  4. 04

    Paint PBR materials

    The texture model generates multi-view albedo and metallic-roughness maps, using normal and position maps rendered from the mesh plus your reference image.
05

Hunyuan3D 2.5 shape benchmarks

Tencent measured how closely each model's shapes match the input images (-I) and text prompts generated from those images (-T). Higher is better in every column.

Hunyuan 3D 2.5

ULIP-T
0.07853
ULIP-I
0.1306
Uni3D-T
0.2542
Uni3D-I
0.3151

Hunyuan3D 2.0

ULIP-T
0.0771
ULIP-I
0.1303
Uni3D-T
0.2519
Uni3D-I
0.3151

TRELLIS

ULIP-T
0.0769
ULIP-I
0.1267
Uni3D-T
0.2496
Uni3D-I
0.3116

Commercial model 1

ULIP-T
0.0741
ULIP-I
0.1308
Uni3D-T
0.2464
Uni3D-I
0.3106

Commercial model 2

ULIP-T
0.0746
ULIP-I
0.1284
Uni3D-T
0.2516
Uni3D-I
0.3131
ModelULIP-TULIP-IUni3D-TUni3D-I
Hunyuan 3D 2.50.078530.13060.25420.3151
Hunyuan3D 2.00.07710.13030.25190.3151
TRELLIS0.07690.12670.24960.3116
Commercial model 10.07410.13080.24640.3106
Commercial model 20.07460.12840.25160.3131

Tencent's own numbers; the commercial models are not named and we have not reproduced them. Version 2.5 leads on ULIP-T and Uni3D-T, ties 2.0 on Uni3D-I and trails Commercial model 1 on ULIP-I. Two older open baselines, Michelangelo and Craftsman 1.5, are left out here. The report itself says these metrics do not fully show the jump in detail.

06

Hunyuan3D 2.5 texture benchmarks

For textures Tencent compared against published texturing methods. Lower is better for CLIP-FID, FID, CMMD and LPIPS; higher is better for CLIP-I. Three text-only methods (Text2Tex, SyncMVD, Paint-it) are left out; 2.5 also beats them on every metric they report.

Hunyuan 3D 2.5

CLIP-FID
23.97
FID
165.8
CMMD
2.064
CLIP-I
0.9281
LPIPS
0.1231

Paint3D

CLIP-FID
26.86
FID
176.9
CMMD
2.400
CLIP-I
0.8871
LPIPS
0.1261

TexGen

CLIP-FID
28.23
FID
178.6
CMMD
2.447
CLIP-I
0.8818
LPIPS
0.1331
MethodCLIP-FIDFIDCMMDCLIP-ILPIPS
Hunyuan 3D 2.523.97165.82.0640.92810.1231
Paint3D26.86176.92.4000.88710.1261
TexGen28.23178.62.4470.88180.1331

In a user study ranking textured results against three unnamed commercial models, Tencent reports a 72% win rate for image to 3D, which it says is nine times the win rate of its Commercial Model 1.

07

What is new in Hunyuan 3D 2.5

These are the upgrades Tencent highlights over the 2.0 release, in its launch post and report.

A much bigger shape model

LATTICE scales data, model size and compute. Tencent says quality improves steadily as the model is scaled up; its largest version has 10B parameters.

Finer detail, cleaner surfaces

The report shows correct finger counts and bicycle-wheel patterns, with sharp edges and smooth surfaces, a balance the report says earlier models, 2.0 included, struggled with.

Four-view input

The shape model accepts a single image or four multi-view images, so you can show the model the back of an object instead of leaving it to guess.

PBR materials

The paint stage now outputs albedo and metallic-roughness maps. A shared attention mask keeps the channels aligned, and a separate training loss separates lighting from color.

Sharper texture alignment

A two-phase training scheme adds zoom-in crops on top of 6-view 512 × 512 training, and inference runs at up to 768 × 768 per view.

Rigging on the platform

Tencent's launch added automatic bone binding and skin weights, plus mesh reduction templates for text and image to 3D.
08

Hunyuan 3D 2.5 limits to know

Hunyuan 3D 2.5 was a big step up by Tencent's own measures, but Tencent lists no downloadable 2.5 checkpoint, and its API now lists newer 3.x versions.

Do
  • Use 2.0 or 2.1 if you need weights you can run yourself
  • Treat the benchmark tables as Tencent's own figures
  • Plan time for retopology before animation or games
  • Check a mesh in a slicer before you 3D print it
Avoid
  • Search Hugging Face for 2.5 weights; Tencent lists none
  • Expect to pick 2.5 in Tencent's international API today; its Pro endpoint lists only versions 3.0 and 3.1
  • Expect clean quad topology; Tencent said in April 2025 it was still improving mesh decimation
  • Confuse the 10B figure with VRAM; Tencent lists no local hardware requirements for 2.5
09

Can you still use Hunyuan 3D 2.5? An alternative

Tencent lists no open Hunyuan 3D 2.5 weights, so there is nothing to install. At launch you used it on Tencent's 3D creation platform, with 20 free generations a day, or through the Tencent Cloud API. Tencent has since moved on: its international Pro 3D API lists only model versions 3.0 and 3.1, and the nearest open releases you can run yourself are Hunyuan 3D 2.0 and 2.1.

If you just want a textured model from a picture, SupaVoxel's image to 3D generator runs online with nothing to install. Upload one image, add up to 4 extra angles if the back matters, and get a textured model in typically 3 to 5 minutes. Most new accounts get 3 free credits with no card, enough for one textured model, and failed generations refund their credits.

View your model free in the browser; downloads, including STL, OBJ, USDZ and multi-color 3MF exports, come with a paid plan from $25 a year. Whatever made your mesh, the free STL repair tool fixes print errors, and the mesh simplifier slims a heavy mesh for a game engine.

10

What SupaVoxel's generator makes from one picture

These six models come from SupaVoxel's gallery, each made from one image with SupaVoxel's generator, so you can judge the result.

SupaVoxel gallery: steampunk robot knight with shield and gears
SupaVoxel gallery: bronze tiger sculpture with deep spirals
SupaVoxel gallery: chibi dragon warrior in a plush hood
SupaVoxel gallery: fantasy mushroom cottage with coral details
SupaVoxel gallery: ornate jewelry box with an enamel rose
SupaVoxel gallery: opera actress bust with dense ornament
FAQ

Frequently asked questions

What is Hunyuan 3D 2.5?

Hunyuan 3D 2.5 is a version of Tencent's image-to-3D system, launched on April 23, 2025. It keeps the two-stage design of 2.0, a shape model followed by a texture model, but uses a new LATTICE shape model of up to 10B parameters and paints PBR materials. Tencent offered it on its online 3D platform and through the Tencent Cloud API, not as open weights.

Is Hunyuan3D 2.5 open source?

Not as far as Tencent's listings show. Tencent published a technical report for version 2.5, but lists no downloadable 2.5 checkpoint on Hugging Face. The open image-to-3D releases nearest to it are 2.0, with its 2mini and 2mv variants, and 2.1, which added PBR materials and training code in June 2025. If you need weights to run on your own GPU, use 2.0 or 2.1.

What is the difference between Hunyuan3D 2.5 and 2.0?

Version 2.5 replaces the 1.1B Hunyuan3D-DiT shape model with LATTICE, up to 10B parameters, accepts one image or four views, and outputs PBR materials instead of a color texture. Tencent also added automatic rigging on its platform. Version 2.0 is open and runs locally with 16 GB of VRAM; Tencent lists no 2.5 weights. See our Hunyuan 3D 2.0 page for the open model.

Is Hunyuan 2.5 better than 2.1?

Tencent describes 2.5 as building on both 2.0 and 2.1: its texture model extends the 2.1 PBR architecture, and its shape model is far larger than the 3.3B shape model of 2.1. The practical difference is access. Version 2.1 has open weights and needs about 29 GB of VRAM for shape plus texture; Tencent offered 2.5 hosted and lists no 2.5 weights.

Can I still use Hunyuan3D-2.5?

Not by picking it by name in most places. Tencent Cloud's international Pro 3D API lists only model versions 3.0 and 3.1. For a self-hosted option use the open 2.0 or 2.1 releases. For a no-install option, SupaVoxel's generator returns a textured model from one picture in typically 3 to 5 minutes.

Was Hunyuan 3D 2.5 free?

At launch, Tencent gave every user 20 free generations a day on its 3D creation platform, double the earlier allowance, and opened an API on Tencent Cloud. Tencent lists no 2.5 weights to download. SupaVoxel gives most new accounts 3 free credits with no card, enough for one textured model you can view free in the browser.

Does Hunyuan 3D 2.5 support text to 3D?

The 2.5 report describes an image-to-3D model, conditioned on one image or four views. Tencent's platform also advertised text to 3D at launch, with new pipeline templates for it. If you only have an idea in words, SupaVoxel's text to 3D page starts from a prompt.

Is Hunyuan 3D 2.5 good for 3D printing?

Tencent's report shows much finer detail than 2.0, with smooth surfaces and sharp edges, though these are Tencent's own examples. You still need to check the mesh in a slicer, and in April 2025 Tencent said it was still improving mesh decimation. Run any mesh through the free STL repair tool if it reports holes. SupaVoxel's generator exports STL or 3MF directly on a paid plan.
Start free

No Hunyuan 3D 2.5 download? Make a 3D model online

Upload one picture and get a textured 3D model in typically 3 to 5 minutes, with nothing to install. Start with 3 free credits, no card needed.

No credit card required · 3 free credits to start

Hunyuan 3D 2.5: What Changed From 2.0 | SupaVoxel