EntVista/metrixel-character-renders

Metrixel Animated Character Renders

Multi-view renders, signed-distance-field volumes and per-view mesh tensors produced by Metrixel from a small set of rigged, animated humanoid characters.

Each character is captured from four camera angles (0°, 90°, 180°, 270°) across ~30 sampled frames of its motion clip, at 512×512. Every frame/angle carries a matching 64³ signed-distance-field volume and a per-view mesh tensor, so the geometry, the implicit surface and the image are aligned sample-for-sample.

This is the output of a pipeline run, not the input to one: the source character models are not redistributed here. The point is to show what a Metrixel run produces — same command, different assets.

Want this on your own models? Sign up free — 1,000 export tokens, no credit card. Point the Metrixel CLI at your model library to reproduce the same multi-view renders, SDF volumes and mesh tensors.

This is a feature showcase / sample, not a full training corpus. It exists to demonstrate Metrixel's batch pipeline and on-disk schema.

Generated output only. The source models are not redistributed here — see the provenance table for where each one came from.

Preview

What's inside

data/<asset_id>/
├── images/      512x512 RGBA renders, one per camera angle per frame (transparent background)
├── sdf/         signed-distance volumes — raw little-endian float32, 64³ grid, row-major (z,y,x)
├── meshes/      per-view mesh tensors (.pt: vertices + normals + UVs + faces)
├── animations/  rig take in USD (.usda)
└── textures/    embedded material textures (.jpg)
metadata.jsonl   one row per asset: paths + license + geometry stats

Image filenames encode the camera + frame as AZ_AZ_AZ_FRAME (3-digit azimuth, 4-digit frame). SDF files are sdf_<az>_<el>_<roll>_<frame>.bin; load with numpy.fromfile(path, dtype='<f4').reshape(R, R, R) (negative = inside, positive = outside, zero on the surface).

Loading & visualizing

Open-source tooling for the .pt mesh tensors and the raw SDF volumes lives in metrixel-3d-dataset-tools (MIT):

git clone https://github.com/entvistastudio/metrixel-3d-dataset-tools
# mesh_visualization/ — load + view the per-view .pt tensors
# sdf_visualization/  — inspect the float32 SDF volumes

How it was generated

Metrixel 2026.3.0, CLI (preview — pin the version for reproducibility). SDF backend actually used: vulkan.

metrixel --mode cli --gen_mode batch \
  --input <animated-characters>/ --output <out>/ \
  --width 512 --height 512 \
  --single_frame=false --camera_step=90 \
  --frames_per_second=30 --frame_stride=50 \
  --animation_stay_in_place=true \
  --frame_format=png \
  --camera_distance_factor=1.4 \
  --gen_sdf=true --sdf_format=bin --sdf_backend=auto \
  --manifest=true

Provenance

asset	source	licence	tris	generated
AncientGeneral.glb	generated	evs-sample-assets-1.0	59999	2026-08-10
AncientKing.glb	generated	evs-sample-assets-1.0	60000	2026-08-10
AncientOfficer.glb	generated	evs-sample-assets-1.0	59999	2026-08-10
AncientRankedOfficer.glb	generated	evs-sample-assets-1.0	60000	2026-08-10
AncientSolider.glb	generated	evs-sample-assets-1.0	59999	2026-08-10

Licensing

Released under the EVS Sample Assets Licence 1.0 (evs-sample-assets-1.0). You may use, modify, redistribute and sell work made with these outputs, in commercial projects, with no attribution required.

One category of use is excluded: this material may not be used as training, fine-tuning or validation data for generative 3D models (text-to-3D, image-to-3D, mesh generation), nor included in datasets assembled for that purpose. Analytical and computer-vision research, benchmarking of existing tools, and reconstruction or segmentation work are all expressly permitted — see Clause 2.

The underlying characters were produced with generative tooling and the motion is retargeted from a separately-sourced capture. Everything here is provided as is, with no warranty of uniqueness, originality or non-infringement.

Honest limits

• Renders are consistent, not cinematic: fixed lighting, transparent background, four fixed angles at 512×512.
• Four camera angles only. This is enough to show view consistency; it is not dense multi-view coverage.
• The source geometry is clean and watertight, so the SDF path here never exercises the generalized-winding-number route that non-watertight real-world scans require. Do not read this set as evidence for that case.
• Every character performs the same motion clip, so the variation across the set is in the characters, not the motion.
• Small set, single class (humanoid characters). Not a benchmark and not a corpus.

Related

• Open-source readers for the .pt mesh tensors and raw SDF volumes: metrixel-3d-dataset-tools (MIT).
• The source characters these were rendered from are published separately as metrixel-rigged-characters — rigged, animated GLBs rather than pipeline output.