infinigen2 examples

infinigen2 examples#

Each example is a single standalone script that renders a multi-pass video dataset with infinigen2. Each example script is extremely customizeable. You can call additional object / scene generators from the docs, change what ExportTypes are rendered, or add additional rendering steps with objects or settings changed.

flying_indoor [source]#

A stereo video of a furnished room with animated floating objects and lights, with ground truth for the left camera. Renders a fixed 24-frame 1280x720 clip.

Infinigen2 Flying Indoors Datarelease

▶ Watch the 2.0.0a2 flying-indoor video  |  📦 Download pregenerated data for this setting here!

If you wish to generate data similar to the video or datarelease, you must install infinigen==2.0.0a2. Otherwise, you will produce data from a newer version, possibly with different objects and materials, but still matching the overall design of flying objects in an indoor room.

This example is grouped under examples/flying_indoor/:

  • render.py — the render script.

  • cvdpack.json — the cvdpack config that packs a rendered clip into per-camera videos.

  • sbatch.sh — a SLURM array script that renders, packs, and offloads a dataset.

To render a single camera locally, please run:

uv run python examples/flying_indoor/render.py --seed 0 --camera_idx 0 --trajectory_seed 0 --output outputs/flying_indoor

  • --seed is the random seed used to determine the base 3D scene (including objects, materials, lights, layout).

  • --trajectory_seed is the random seed used to determine only the camera trajectory. Changing this seed re-randomizes the trajectory without affecting the base 3D scene.

  • --camera_idx {0,1} chooses which stereo camera to render (0 = left, 1 = right). You must run render.py twice for each seed pair (seed and trajectory_seed) to get stereo data.

By running both camera_idx for a single seed pair, you will receive the following data:

  • left — rgb + camera.npz, plus material-index, diffuse-color, environment passes.

  • right — rgb + camera.npz.

  • left gt — depth/surface-normal/object/optical-flow .npy.

  • right gt — depth .npy.

  • metadata.json (seed, per-pass runtimes, exports), written by the left-camera shard.

To render a dataset on SLURM, sbatch.sh runs one array task per (scene, trajectory, camera), with NUM_TRAJECTORIES unique trajectories per scene (default 4). Each task renders raw frames to fast local SCRATCH_DIR, packs them with cvdpack, and offloads only the compressed videos to FINAL_DIR, then deletes the raw frames:

sbatch examples/flying_indoor/sbatch.sh

💡 Edit the hardcoded SCRATCH_DIR (node-local scratch) and FINAL_DIR (permanent storage) at the top of sbatch.sh to suit your cluster, size --array as N_SCENES * NUM_TRAJECTORIES * 2, and adjust --partition/--account/other SLURM configs.

panning video of clay-material scenes [source]#

A linear camera pan of a livingroom, rendered in several passes: clay, ambient occlusion, rgb, and ground truth.

wget https://raw.githubusercontent.com/princeton-vl/infinigen/main/examples/render_clay_pan_video.py
uv run python render_clay_pan_video.py --seed 0 --output outputs/clay_pan_video

Each pass writes %c/<name>-%f.png under --output (%c = camera, %f = frame):

  • clay-flat + ao-flat — undisplaced mesh (DisplacementMode.NONE), with a camera-parented fill light.

  • clay + ao-disp — displaced mesh (DISPLACEMENT_AND_BUMP); fine surface detail shows up in the AO pass.

  • rgb — full materials, plus camera.npz.

  • gt — depth/surface-normal/object/optical-flow .npy; skipped with --skip_gt.