TencentARC / Pixal3D

Overview

Pixal3D is an image-to-3D model from TencentARC. It builds a 3D mesh that matches the input photo’s viewpoint, not a generic pose. The pipeline lifts 2D image features into 3D with pixel back-projection, then generates structure, detailed geometry, and textures in stages. You get a ready-to-import GLB asset, which reduces manual cleanup for games, AR, and product renders.

What you can build

• Turn product photos into GLB assets for ecommerce viewers
• Create prop meshes for Unity or Unreal from concept art
• Make quick graybox assets from a single reference image
• Generate textured kitbash parts for Blender scenes
• Convert sprite-like artwork into simple 3D stand-ins

Inputs

• One reference image (PNG, JPG, or WebP). PNG with transparency works best for clean silhouettes.
• A generation resolution choice (for example 1024 or 1536). Higher values usually add detail.
• A texture map size for the exported GLB (from 1024 up to very large sizes, if supported). Larger textures increase file size.
• An optional random seed as a whole number. Use it to reproduce the same result.
• A mesh detail target for export, given as an approximate face budget. Higher budgets keep more geometry detail.
• Advanced quality controls for the 3-stage generation: - Structure stage step count and guidance strength. This controls the coarse 3D layout.
  • Shape stage step count and guidance strength. This controls fine geometry detail.
  • Texture stage step count and guidance strength. This controls how strongly textures follow the input colors.
  • Guidance rescale controls for each stage. These can reduce artifacts when guidance is high.
  • Timestep rescale controls for each stage. These shift where the sampler spends effort.
• An advanced detail cap for the high-resolution stage, expressed as a token budget. Higher budgets can capture finer structure.

Outputs

• A downloadable 3D GLB file.
  • It contains a triangle mesh with UVs.
  • It includes a material setup with textures baked at your chosen texture size.
  • The mesh is simplified toward your chosen face budget for easier downstream use.
• The seed actually used for the run, so you can re-run the same setup later.

Recommended settings

• For most assets: use 1024 resolution and a 2048 texture size.
• For close-up hero assets: use 1536 resolution and a 4096 texture size, if available.
• If geometry looks too soft: increase the shape stage steps before raising guidance.
• If you see noisy bumps or spikes: lower shape guidance slightly, or raise shape guidance rescale.
• If textures look oversaturated: keep texture guidance low and rely on more texture steps.
• For real-time engines: keep the face budget closer to a few hundred thousand, then add LODs in your engine.

Limitations

• Single-image 3D is ambiguous. Hidden surfaces may be guessed or incomplete.
• Occlusion hurts results. Hands over objects, hair, and clutter often cause holes.
• Busy backgrounds can confuse segmentation. A clean cutout or transparent PNG helps.
• Thin structures and reflective surfaces can break into disconnected parts.
• Low-resolution, blurry, or heavily compressed inputs reduce mesh detail and texture clarity.
• If the subject is cropped out of frame, the model may invent missing geometry.

Safety & compliance

• Follow the Pixal3D license terms when you use the weights and outputs.
• Don’t use it to create illegal content or to violate privacy rights.
• Don’t upload images you don’t have rights to use, especially brand assets or copyrighted designs.
• Access restrictions may apply in some regions, depending on the model host and license gating.