Skip to content

tools.md

Latest commit

 

History

History
219 lines (168 loc) · 7.91 KB

tools.md

File metadata and controls

219 lines (168 loc) · 7.91 KB
Raw

Demos and conversion tools

The repository includes optional command-line tools for visualization, retargeting, pose sampling, and DCC integration. Install the demo extras before using tools that render images or video:

uv pip install -e ".[demo]"

Full-body demo

Render one or more identity backends with the shared SOMA body rig:

python tools/demo_soma_vis.py \
  --data-root assets \
  --output-dir out/body-demo \
  --identity-model-type soma,mhr \
  --lod mid

Use --random-shape to animate identities and --motion-file to provide a custom motion. Both body and hand demos default to assets/example_animation.usd, the canonical example animation updated alongside the corrective model. The mid, low, and xlo LODs share the same public pose contract.

SOMA Hand demo

Render the left and right wrist-local layers from a body motion:

python tools/hand/demo_soma_hand_vis.py \
  --data-root assets \
  --hand-type left,right \
  --remove-wrist-translation \
  --output-dir out/hand-demo

Use --shape-only --random-shape to render native hand identity variation without a body motion. MANO and MHR identity backends are selected with --identity-model-type and require their corresponding local model assets. For MANO, pass the separately licensed model explicitly with --hand-type right --mano-model-path /path/to/MANO_RIGHT.pkl (or the left-hand equivalent). By default, motion renders apply the wrist's animated world transform. Use --remove-wrist-translation to retain wrist orientation while keeping the hand centered. The default camera framing leaves room around the motion-wide hand bounds; increase --camera-framing-scale to zoom out further.

Both demos leave pose correctives disabled by default. Use --apply-correctives to enable them.

Both demos accept --skeleton-overlay to draw the public skeleton as octahedral bones over the mesh, and --skeleton-style {light,skin} to pick a neutral light-gray or a darker skin-toned bone color.

README teaser

make_teaser.py composes the body and hand demo renders into the single README teaser GIF with one title and label specification, so both rows stay aligned and typographically consistent. The published assets/images/soma-in-action.gif uses correctives. To regenerate it, opt in when rendering and composing:

python tools/demo_soma_vis.py \
  --identity-model-type soma,mhr,smplx,anny,garment \
  --apply-correctives --procedural-transforms on --pose-batch-size 64 \
  --skeleton-overlay --skeleton-style skin \
  --image-size 1440 --max-frames 640 --output-dir out/teaser/correctives/body
for backend in soma mhr mano; do
  python tools/hand/demo_soma_hand_vis.py \
    --hand-type right --identity-model-type "$backend" \
    --apply-correctives \
    --remove-wrist-translation --camera-framing-scale 1.0 \
    --skeleton-overlay --skeleton-style skin \
    --image-size 1440 --max-frames 640 --output-dir out/teaser/correctives/hand
done
python tools/make_teaser.py --apply-correctives --fps 12

This writes assets/images/soma-in-action.gif, the single published teaser. Demos and the compositor still default to without correctives. To render that mode, omit --apply-correctives at both stages and use out/teaser/no_correctives/body and out/teaser/no_correctives/hand for the source directories. Both modes use the same default GIF output path.

The compositor's --apply-correctives flag selects pre-rendered inputs; it does not calculate corrective offsets. Use --render-dir to change the common render root, or --body-videos, --hand-videos, --body-labels, --hand-labels, and --output to supply custom inputs, labels, and output. Explicit video lists override the mode's default inputs.

The MANO backend additionally needs --mano-model-path, and the SMPL-X and GarmentMeasurements backends need their locally licensed model files. The compositor requires ffmpeg on the path.

The example USD runs at 24 fps; keeping every second frame at 12 fps preserves its playback speed. The 640-frame source range ends after the second rightward kick: the last GIF sample is source frame 638, back in neutral standing before the next action. This gives a 26.67-second teaser. The recipe enables procedural body transforms.

Sample the hand articulation prior

sample_soma_hand_pose_pca.py draws reproducible coefficients from the distributed articulation prior, converts the bind-relative exponential maps to the layer's absolute-local rotation convention, poses the skinned mesh, and writes MP4/GIF output.

python tools/hand/sample_soma_hand_pose_pca.py \
  --hand-asset assets/SOMAHand.npz \
  --hand-type right \
  --num-poses 60 \
  --seed 20260829 \
  --output-prefix out/soma-hand-samples

Use --n-components, --sample-scale, and --lod to control the sampled prior and rendered geometry.

SMPL-family to SOMA

Convert an SMPL animation to SOMA and optionally export the recovered pose:

python -m tools.smpl2soma --output-npz out/smpl-soma.npz

The converter uses PoseInversion: analytical fitting is the default, and --autograd-iters adds differentiable FK refinement. SMPL/SMPL-X model files must be supplied under their own license.

MHR to SOMA

Convert MHR-format parquet data, including SAM 3D Body outputs:

python -m tools.mhr2soma \
  --input /path/to/parquet-directory \
  --output-npz out/mhr-soma.npz

Use --max-samples for bounded local checks. The tool also exposes the reusable RTS smoothing presets through its --smooth options.

Convert identity backends

The identity conversion tools accept a canonical SOMA animation NPZ written by soma.io.save_soma_npz, preserve its animation, and optimize target identity parameters against the source geometry in the SOMA bind pose.

Convert a full-body SMPL-X fit to native SOMA identity parameters:

python -m tools.convert_identity_backend input_smplx.npz output_soma.npz \
  --target-backend soma \
  --source-model-path /path/to/SMPLX_NEUTRAL.pkl

The source backend and coefficients are read from the input NPZ. This example therefore converts its stored SMPL-X betas to native SOMA identity coefficients and bone scales. Other supported full-body targets include MHR, Anny, SMPL, SMPL-H, SMPL-X, and GarmentMeasurements.

Use the separate hand entry point for native SOMA Hand, MHR hand, and MANO:

python -m tools.hand.convert_identity_backend input_mhr_hand.npz output_soma_hand.npz \
  --target-backend soma

Native SOMA bone scales are optimized by default for body and hand targets. Global scale remains fixed to the input value unless --optimize-global-scale is explicitly supplied. The output is another SOMA NPZ containing the target parameters plus conversion_* source metadata, loss history, and per-identity bind-pose vertex error. Use --no-optimize-scale-params to keep target scale parameters neutral.

AMASS to SOMA

Convert one AMASS sequence or a directory tree of SMPL motion files:

python -m tools.convert_amass_to_soma \
  --input /path/to/sequence.npz \
  --output-npz out/soma.npz \
  --no-render

For batch conversion, use --input-dir and --output-dir. The exported NPZ contains SOMA poses, root translations, joint names, reconstruction errors, and identity parameters.

MANO interoperability

The hand tools convert pose and identity parameters using user-supplied MANO v1.2 files:

  • tools/hand/mano2soma.py
  • tools/hand/soma2mano.py

Both commands fail with an actionable model-path error when the licensed MANO file is unavailable. See SOMA Hand data assets for the topology correspondence and local asset setup.

DCC integration

Procedural transform references and setup instructions are maintained with their integrations:

  • Blender: tools/soma_procedural_blender/README.md
  • Maya: tools/soma_procedural_maya/README.md

The procedural control format documents the shared sidecar schema used by the Python runtime and DCC implementations.