Preparing Custom Data

This guide describes how to prepare a custom scene for CHORD. It covers the mesh layout expected by the optimization code and the object/background 3D Gaussian fitting step.

Mesh Layout

CHORD expects a mesh directory with one scene mesh and one mesh per object:

data/<scene_name>/
├── scene.glb
├── obj_0.glb
├── obj_1.glb
├── obj_2.glb
└── ...

Important conventions:

• scene.glb contains the objects whose motion will be generated. It is used to compute scene scale and center for camera sampling.
• obj_1.glb is the default static background convention.
• Each obj_i.glb must preserve the object's world-space transform.

The image below illustrates the expected object/background split. obj_1.glb follows the default static background convention, while the other obj_i.glb files can be optimized as dynamic objects.

Train Scene Gaussians

Train all object and background Gaussians for a scene:

python scripts/preprocess/train_gaussian_for_scene.py -m trained/<scene_name> --mesh_source_path data/<scene_name>

The scene-level script discovers obj_<id>.glb files, trains dynamic objects first, and trains the default static background obj_1 last.

For manual debugging of a single object, use the lower-level scripts/preprocess/train_gaussian_per_obj.py script with --cur_train.

Command Line Arguments for scripts/preprocess/train_gaussian_for_scene.py

-m

Path where the fitted object and background Gaussians are saved, such as trained/<scene_name>.

--mesh_source_path

Path to the scene mesh directory, such as data/<scene_name>.

--objects

Optional subset of objects to train, such as obj_0 obj_2. If omitted, all top-level obj_<id>.glb files are trained.

--static_id

Object id used as the static background convention. The release setup uses 1, corresponding to obj_1.

--skip_static

Train only dynamic objects and skip the static background.

--dry_run

Print the per-object training commands without running them. This is useful for checking object discovery and training order.

--dynamic_elev_r, --static_elev_r

Upper elevation bound used for dynamic-object and static-background Gaussian fitting.

--dynamic_near_cam_radius, --static_near_cam_radius

Near-surface camera radius used for dynamic-object and static-background fitting.

--no_light

Disable the default simple lighting used during Gaussian fitting.

--no_zoom_in_train

Do not use zoomed-in training views for dynamic objects.

--no_light_bg

Do not apply additional lighting to dynamic-object background renders.

Preview Static Renders

After fitting the static Gaussians, use scripts/evaluation/render_scene_static.py to check the scene composition and prompt framing:

python scripts/evaluation/render_scene_static.py -m trained/<scene_name> --mesh_source_path data/<scene_name> --checkpoint -1 --image_width 832 --image_height 480 --cam_radius 1.8 --azim_l 0.0 --azim_r 360.0 --elev_l 10.0

Command Line Arguments for scripts/evaluation/render_scene_static.py

-m

Path to the fitted 3D Gaussian directory for the scene, such as trained/<scene_name>.

--mesh_source_path

Path to the scene mesh directory, such as data/<scene_name>.

--obj_num

Number of object meshes in the scene. The default -1 auto-detects this from the top-level obj_i.glb files in the mesh directory.

--checkpoint

Gaussian checkpoint iteration to render. Use -1 to load the latest checkpoint.

--image_width, --image_height

Output preview resolution in pixels.

--cam_radius

Camera distance from the scene center.

--azim_l, --azim_r

Azimuth range, in degrees, for rendering preview views around the scene.

--elev_l

Camera elevation, in degrees, for the preview orbit. Positive values place the camera above the scene.