Hunyuan3D 2 vs 2.1: Which One to Run for Your Asset
Same request, very different assets. Here's what changed from Hunyuan3D 2 to 2.1, from the bigger shape model to PBR materials, and which version to run for four common 3D jobs.
By Eftal Yurtseven · Nov 20, 2026 · 9 min read

You have one photo of a lamp, a sneaker or a ceramic dragon, and you need it as a 3D asset by this afternoon. Tencent gives you two versions of the same idea, and the version number makes 2.1 sound like a patch. It isn't. The hunyuan3d 2 vs 2.1 choice decides whether your asset comes back with flat painted color or with real material maps, and whether you wait seconds or minutes for it.
Both run on each::labs today: Hunyuan3D 2 and Hunyuan3D 2.1, side by side in the Hunyuan 3D family. They take the exact same request. So the question isn't which one is "better" in the abstract. It's which one fits the job on your desk.
What Changed From Hunyuan3D 2 to 2.1
The architecture didn't flip. Both versions split the work in two: a shape model builds a bare mesh from your image, then a paint model textures that mesh. Tencent released Hunyuan3D 2.0 on January 21, 2025 and Hunyuan3D 2.1 on June 13, 2025, and the 2.1 README names two headline changes: a fully open-source framework and physically based rendering (PBR) texture synthesis.
"Open" means something specific here. Hunyuan3D 2.0 shipped inference code and weights. With 2.1, Tencent says it released the full model weights and the training code for the first time, so teams can fine-tune the model on their own data. If you only call the model through an API, that changes nothing in your output. If your studio wants a version tuned on its own art style, it's the whole story.
The other two changes land directly in your asset: a bigger shape model, and a paint model that outputs materials instead of color. Both deserve a closer look.

Shape: A Bigger Model Reading the Same Photo
In Tencent's model tables, the 2.0 shape model (Hunyuan3D-DiT-v2-0) is listed at 1.1B parameters. The 2.1 shape model (Hunyuan3D-Shape-v2-1) is listed at 3.3B. Same job, roughly three times the capacity. Tencent's 2.1 shape benchmark table puts Shape 2.1 at the top of its comparison on the ULIP and Uni3D alignment scores, though 2.0 itself isn't in that table, so read it as "strong," not as a direct head-to-head.
The 2.1 technical report also spells out what the shape model does to your photo before it reads it: it removes the background, resizes the object to a standard size, centers it, and fills the background with white. That's a useful hint for preparation, and we'll come back to it.
In practice, a bigger shape model matters most on the parts that small models fumble: thin handles, gaps between fingers, the underside of a chair. On a simple blob of a pebble, the difference is smaller than the version number suggests.

Paint: From Flat Color to PBR Materials
This is the change that matters. The 2.1 README says the new PBR pipeline replaces "the prior RGB-based texture model." In plain terms, Hunyuan3D 2.0 paints color. Hunyuan3D 2.1 paints materials.
The 2.1 technical report is specific: the paint model generates albedo, metallic and roughness maps for the mesh. Albedo is the surface color with no lighting baked in. Metallic and roughness tell a renderer how the surface reflects light. Put those three into a game engine, a web viewer or a renderer and the object reacts to your lights, rather than carrying the photographer's lights painted onto it forever. Tencent also describes an illumination-invariant training strategy designed to keep the albedo free of baked lighting, a job the 2.0 lineup handed to a separate model Tencent calls an image delight model.
Tencent's numbers back the upgrade. On its texture benchmarks, Paint 2.1 beats Hunyuan3D 2.0 on every metric it reports: CLIP-FiD drops from 26.44 to 24.78, CLIP-I rises from 0.8893 to 0.9207, and CMMD and LPIPS both improve. Those are Tencent's figures, not ours, but the direction is unambiguous.
The paint model grew too: 1.3B parameters in 2.0, 2B in 2.1.
Hunyuan3D 2 vs 2.1 Side by Side
Here's the comparison that matters when you're actually choosing, read straight from Tencent's READMEs and the two each::labs model pages.
Request. Identical. On each::labs, both models take one required field, input_image_url, plus the same optional settings: seed, num_inference_steps (default 50), guidance_scale (default 7.5, range 0 to 20), octree_resolution (default 256) and textured_mesh (off by default). There's no text prompt field on either, so the photo is the whole brief.
Output. The sample output on the Hunyuan3D 2 page is a single GLB file. The sample on the Hunyuan3D 2.1 page is a ZIP archive, so plan one unzip step into your pipeline before the asset reaches your engine or viewer.
Speed. The model pages list an average run of about 20 seconds for Hunyuan3D 2 and about five minutes for Hunyuan3D 2.1. That gap shapes everything below.
Texture. RGB color for 2. PBR albedo, metallic and roughness for 2.1.
Model sizes. Shape 1.1B and paint 1.3B for 2. Shape 3.3B and paint 2B for 2.1.
Running locally. Tencent lists 6 GB of VRAM for 2.0 shape generation and 16 GB for shape plus texture. For 2.1 it lists 10 GB for shape, 21 GB for texture and 29 GB for both. On each::labs, none of that is your problem.
License. Each version has its own Tencent Hunyuan 3D Community License with the same structure: the territory excludes the European Union, the United Kingdom and South Korea, and products with more than one million monthly active users must request a license from Tencent. Read the license file for your version before you ship.
Because the request is the same, you can send one photo to both models and compare the results without rewriting anything. It's the fastest way to settle the question for your own object.

When to Pick Which: Four Jobs
Our position: default to 2.1 for anything a customer or player will see under changing light, and keep 2 for speed work where materials don't matter yet. Here's how that plays out by job.
Game prop
Pick Hunyuan3D 2.1. Game engines light everything dynamically, and a prop with lighting baked into its color looks wrong the moment a torch moves past it. Albedo, metallic and roughness maps are the inputs a standard metallic roughness material expects, so the asset drops into a familiar setup. The five minute run is a small wait for a prop that ships.
Product turntable
Pick 2.1 again, and for the same reason with higher stakes. A turntable rotates the product through the light, which is exactly when baked highlights betray themselves: the shine stays glued to one side of the bottle as it spins. Metallic and roughness maps let a chrome cap read as chrome and a matte label read as paper from every angle.
3D print
Here the paint stage is irrelevant, because a printer doesn't care about roughness maps. Run either model with textured_mesh left off and judge the geometry. We'd start with 2.1 for its larger shape model, especially on parts with thin features, and raise octree_resolution if fine edges come back soft. If you're printing a quick test piece rather than a finished part, 2 gets you a mesh to slice in a fraction of the time.
Quick concept blockout
Pick Hunyuan3D 2. When you're testing whether a creature design reads in 3D at all, or blocking out a scene with ten placeholder objects, a 20 second average run beats a five minute one every time. Leave texture off, check silhouettes, and promote the keepers to 2.1 once the design is locked. Fast drafts, careful finals.
Prepare the Input Photo Before Either Model Sees It
Both versions read a single image and nothing else, so the photo carries all the information you can give. Tencent's own preprocessing tells you what the model wants: one object, no background, centered, at a standard size. Hand it something close to that and you remove a whole class of failures.
Start with one object, fully in frame, nothing cropped at the edges. A three quarter angle shows the model front and side at once, which gives it more to work with than a dead flat front view. Soft, even light helps both versions, and it helps 2 most, since its paint model works in color and will happily keep whatever hard shadow you give it.
Then clean it. Run the photo through Bria Remove Background to isolate the subject, so stray props and table edges don't become geometry. If the source shot is cluttered or the object is partly hidden, an edit model such as Nano Banana 2 Edit can rebuild it as a clean studio shot first. And if you're starting from an idea rather than a photo, generate the reference with Nano Banana 2 using a plain prompt like this:
A single ceramic teapot shaped like a sleeping fox, three quarter view, centered, full object in frame, soft even studio lighting, plain white background, no props, no shadows on the backdrop.
That prompt asks for exactly what the 3D models prefer: one subject, a readable angle and nothing competing for attention. For the deeper version of this checklist, our Hunyuan3D 2.1 photo to textured 3D model guide walks through each step.

Run Hunyuan3D on each::labs: A Two-Model Workflow
The smartest way to use both versions is together, since one request works for each. Here's the loop we'd build.
First, prepare the photo as above and host it at a public URL. Second, send it to Hunyuan3D 2 with textured_mesh off. In about 20 seconds you'll see whether the silhouette is right. If the back is mush or a handle fused to the body, fix the photo, not the settings. Try two or three seeds while you're here; each one is a new interpretation of the same image.
Third, once the shape reads correctly, send the same photo to Hunyuan3D 2.1 with textured_mesh on. That's the run that gives you PBR materials, so it's worth the longer wait. Fourth, open the result and look at it under a light you control: rotate it, move the light, check the bottom. If the reflections behave as you move the light, the materials did their job.
If you do this more than a few times a week, chain it. Background removal, the quick 2 preview and the 2.1 final make a natural three step pipeline on each::labs flows, and the same pattern works with any image model from the model catalog feeding the first step. For more on settings and outputs, see 9 things to know about Hunyuan3D, or read Tencent's own Hunyuan3D 2.1 repository for the full model details.
Frequently Asked Questions
Hunyuan3D 2 vs 2.1: which one should I use by default?
For finished assets, 2.1. Its paint model outputs albedo, metallic and roughness maps instead of flat color, and its shape model is about three times larger. Keep 2 for fast previews and blockouts, where its 20 second average run lets you try many photos and seeds quickly.
Is Hunyuan3D-2.1 open source?
The weights and, for the first time in the series, the training code are public on Tencent's GitHub and Hugging Face. The license is the Tencent Hunyuan 3D 2.1 Community License rather than a standard open-source license, and its territory excludes the European Union, the United Kingdom and South Korea.
Can Hunyuan3D generate a 3D model from text?
Not through the each::labs request, which takes an image URL and no prompt field. Generate a clean studio image of your object with a text-to-image model first, then pass that image to either Hunyuan3D version.
Does Hunyuan3D 2.1 make textures by default?
No. On each::labs, textured_mesh is off by default on both models, which gives you geometry only. Turn it on for the 2.1 run when you want the PBR materials, and leave it off for 3D printing or shape checks.



