Trellis 2 Turn a Reference Image into a 3D Asset

Start with an Image That Explains the Shape

Show the complete object and the features you want to carry into 3D.

Choose a reference that shows the whole subject and the connections between its parts. A handle seen edge-on, a leg hidden behind another leg, or a dark surface against a dark background gives the model less information to work with. Use a view that reveals the features you care about, rather than choosing a dramatic camera angle for its own sake.

The upload accepts JPG, PNG, or WebP up to 10MB. Keep space around the object so the silhouette is not cropped. Reflections, cast shadows, and background objects can make the intended boundary ambiguous; simplify the reference when they compete with the subject.

If the idea exists only in words, develop a reference with the AI image generator. For an existing photograph, background removal can help isolate the object. Choose the reference that best communicates the object’s outline and depth.

Balance Shape, Mesh Size, and Texture Detail

These controls solve different problems. Change the one related to what you need to improve.

Begin with the defaults when assessing a new reference. Review the proportions before spending effort on finer detail. A higher resolution is worth evaluating when small geometric features matter, but it cannot make an obscured feature known or guarantee that a thin connection will survive.

Use a smaller simplification target when the object needs to be easier to load or edit. Inspect curved edges and narrow parts afterward: reduction can remove detail you wanted to keep. This control is a mesh simplification target, so do not treat its number as directly equivalent to the face-count settings of another model.

Increase texture size when surface markings matter at the intended viewing distance. If the silhouette is wrong, a larger texture will not fix it. For a small background prop, extra texture detail may contribute little to the final scene. Compare the downloaded asset at the size it will actually appear.

What Trellis 2 Adds to Surface and Material Generation

Microsoft’s O-Voxel representation is designed for varied geometry and material properties.

The published approach represents open surfaces and complex shapes alongside material information such as base color, roughness, metallic properties, and opacity. Use that capability to explore layered forms and mixed finishes, such as painted panels with exposed metal or a decorative object with translucent sections.

For a mechanical object, inspect the gaps between plates and the joints between small components. For a translucent object, inspect both its shape and how its material renders under different lighting. A surface can look plausible while hiding overlapping geometry; material detail and mesh structure deserve separate checks.

This workflow returns a textured model. It does not offer a dedicated geometry-only switch or an existing-mesh retexturing input. If you need untextured output from the start, Hunyuan 3D 3.1 offers an explicit Geometry option.

Trellis 2 vs the Original TRELLIS

TRELLIS.2 emphasizes geometry and material generation through its O-Voxel representation.

The original TRELLIS project uses structured latents that can be decoded into representations including meshes, 3D Gaussians, and radiance fields. Its public project also includes text generation and multi-image examples.

TRELLIS.2 introduces the O-Voxel approach for geometry and materials. Here, its practical interface is single-image input, separate resolution and texture controls, and GLB output. If an older tutorial asks you to select a Gaussian export or enter a text prompt directly into TRELLIS, it describes a different workflow.

If you are moving from an older TRELLIS workflow, plan around GLB mesh editing and material review rather than Gaussian or radiance-field output.

Trellis 2, Meshy 7, or Tripo 3.1?

Start with the references you have and the decisions you want to make before generation.

Choose Trellis 2 to explore a single image with separate shape-resolution and texture-size controls. Its focused input is useful when one reference already communicates the object and you want to adjust those dimensions of detail independently.

Consider Meshy 7 when you want texture direction or a topology choice. The Meshy workflow here accepts text or images and includes a texture prompt, a texture reference, and triangle or quad topology options. Choose those controls when directing a finish or requesting a topology treatment is central to your task.

Consider Tripo 3.1 when you have consistent views of the same subject. Its multi-view input provides named front, left, back, and right slots. That can be a more appropriate starting point than asking a single-image workflow to infer an important hidden feature.

Use the 3D generator to select the workflow that fits your input. When comparing single-image results, keep the reference consistent and check the same outline, hidden surfaces, material details, and usable mesh structure.

Check the Download Beyond the Front View

Import the GLB into your scene to assess its shape, materials, and loading needs.

Rotate underneath and behind the model before downloading. In your editor, review untextured geometry for unwanted interior surfaces, merged gaps, and disconnected pieces. Then restore the materials and check seams, reflections, and transparent areas under the destination lighting.

For a web scene, evaluate file size and responsiveness at the planned viewing distance. For a close-up render, prioritize the visible outline and material behavior. For printing, establish a suitable solid volume and physical size before checking fragile features. Choose the level of detail around the destination and viewing distance.