Prompt Hunyuan 3D for image-to-3D

Last updated: October 12, 2026

Hunyuan 3D builds a textured mesh from one image of an approved design, and stays close to it: the shapes, the trim and the painted colours come back on the model. Pick it when the prop or character is already drawn and the job is getting it into 3D.

Let the agent write it

You don't have to set any of this up yourself. The Layer agent knows how to brief each model it runs, from the same guidance this page is built on, including how to make a clean source image first when all you have is key art. Attach the art in chat, name the model and say what the mesh is for, for example: Turn this chest into a game prop with Hunyuan 3D v3.1 Pro, about 40,000 faces, textured. The agent prepares the source, shows you its plan, and runs it when you approve. Use the rest of this guide when you want to do it yourself, and see Write a brief for the agent for how to brief it.

When to use it

Use Hunyuan 3D for props, pickups, furniture, weapons and characters whose look is already signed off as a 2D image. It reproduces hard-surface trim and hand-painted colour closely, and its texture follows the image rather than reinterpreting it.

It reads one picture, so it does not take direction in words. If you need to change the design, change the image first (see Edit an image with a prompt), then build the mesh. To describe a mesh from nothing, pick a model in the picker that takes a text prompt for 3D (see Choose a model).

Versions on Layer

Hunyuan 3D v3.1 Pro

The version to deliver with. It takes a Face limit and returns a full colour texture. The chest in this guide was made with it.

Hunyuan 3D v3.1 Fast

Quicker and cheaper, with no face limit. In our test it took less than half Pro's time and matched the chest's shape, but came back as an untextured grey model even with Generate textures on. Use it to check that a source image reads as the right shape, then build the keeper on Pro.

How to write the prompt

With a source image, Hunyuan 3D takes no prompt. Tencent's API does not accept a prompt and an image together, so on Layer anything typed in the Prompt box is not sent when a Source image is attached. Leave it empty. Everything you would say in words goes into the image instead:

Product render of a single stylised fantasy treasure chest game prop, straight-on front view at eye level, centred, filling about two thirds of the frame. Rounded wooden lid with iron bands, a gold lock plate with a keyhole on the front, studded corners. Hand-painted stylised game art look. Even, soft studio lighting with no hard shadows, plain light grey seamless background, no floor shadow, no text, no other objects.

That is the image prompt behind the chest in the clip above, run on GPT Image 2. It names one object, the view, the material and the style, then pins down everything the 3D model would otherwise copy by mistake: the background, the light and anything else in the frame.

What works

One object on a plain background

Give it the object alone, on a plain solid colour, filling more than half the frame. We built the same chest twice on the same settings: once from the clean render above, once from dark key art of it sitting on a dungeon floor of coins and bones. The clean render came back as the chest alone, on-model and fully textured. The key art came back fused to a slab of floor, with the coins and bones modelled in and part of the face budget spent on them.

Even, soft light

The texture is painted from the image, so the light in the image becomes part of the asset. In the same pair of runs, the clean render gave neutral grey iron, mid-brown wood and gold. The torchlit key art gave a dark texture with blue-black iron and orange light painted onto the wood, which would look lit twice in an engine. Ask for "even, soft studio lighting with no hard shadows" in the source.

A front view is enough for most props

A straight-on front view works as the only input; the model invents the sides and back. Our chest came back with a plausible back and lid from a single front render. Objects whose back cannot be guessed from the front (an asymmetric backpack, a creature with a tail) are where a single view falls short.

A render, not a flat drawing

Ask for the source as a stylised product render, with shaded form and depth, rather than a flat illustration. Images that already look three-dimensional give the model more to build from.

Make a clean source from key art

When the approved art is a scene, pull the object out of it with an image edit before you build: the same object, alone, front view, plain background, even light. The second prompt under Prompts to start from does this for a character.

What to avoid

Writing a prompt

Text in the Prompt box does nothing once a source image is attached. If the mesh needs to differ from the image, edit the image.

Busy scenes and dramatic light

Floors, props around the object and other characters are built into the mesh, and cast shadows and coloured light are baked into its texture. Key art makes a poor source; a plain render of the same design makes a good one.

Text and logos in the image

Tencent's input rules exclude text. Leave labels and logos out of the source and add them to the texture afterwards.

Settings that matter

Source image is the image the mesh is built from: JPG, PNG or WebP. Face limit (Pro only) caps the mesh, and the mesh meets it: our 40,000 run came back at exactly 40,000 triangles. Stay around 40,000 to 50,000 for game props and previews, and go higher only for close-up renders. Generate textures paints colour onto the mesh; off returns bare geometry. PBR materials adds metalness, roughness and normal maps so the model reacts to light in a game engine; leave it off for a hand-painted, unlit look.

Prompts to start from

These are image prompts for the source, the one place words steer the result.

A prop render to build from

Product render of a single stylised fantasy treasure chest game prop, straight-on front view at eye level, centred, filling about two thirds of the frame. Rounded wooden lid with iron bands, a gold lock plate with a keyhole on the front, studded corners. Hand-painted stylised game art look. Even, soft studio lighting with no hard shadows, plain light grey seamless background, no floor shadow, no text, no other objects.

A character pulled out of key art

Attach the key art as a reference and run:

The same fire mage character from the reference, alone, full body from head to boots, standing straight in a neutral pose with arms slightly away from the body, staff held upright at his side, seen straight-on from the front at eye level, centred and filling most of the frame. Same costume, colours and painterly game style. Even, soft studio lighting with no hard shadows, plain light grey background, no floor, no scenery, no text.

Use it in Layer

  1. Add the approved render to a session's board: one object, front view, plain background.

  2. Click View on the image, then Turn into 3D model. Create mesh opens in the side panel with the image as its Source image.

  3. Click the model name at the top of the panel and choose Hunyuan 3D v3.1 Pro. The source image stays in place.

  4. Leave the Prompt box empty. With a source image, Hunyuan 3D builds from the image alone.

  5. Keep Generate textures on, and leave Face limit at 40,000 for a game prop. Turn on PBR materials if the asset goes into a lit engine scene.

  6. Click Create 3D model. The cost shows on the button before you click.

  7. When the mesh lands on the board, click View to open it and turn it around. If something is off, fix the source image and run it again.

Learn more

Related