DeXplicit

DeXplicit 3D model predicts metric object motion from an initial scene and supplied hand actions. DeXplicit renderer turns those forecasts into future RGB using initial-image appearance and calibrated camera controls.

Code, installation, inference and training · Project page

File	Model
dexplicit-3d-model.safetensors	DeXplicit 3D model; full state, 425M learned parameters
dexplicit-renderer-high.safetensors	DeXplicit renderer; high-noise adapter
dexplicit-renderer-low.safetensors	DeXplicit renderer; low-noise adapter
dexplicit-renderer-compact.safetensors	DeXplicit renderer; optional compact variant

The default inference command predicts the 3D conditions itself and generates 33 frames at 832 × 480. It requires an initial metric scene and features, supplied hand actions, an initial RGB image and calibrated camera controls. Future object observations do not enter that pipeline.

The standard renderer-training recipe also generates its conditions with the frozen DeXplicit 3D model. It uses recorded future RGB only as supervision and rejects measured or unlabelled conditioning bundles.

python -m dexplicit.download --output-dir weights
python -m dexplicit.infer --help
python -m dexplicit.prepare_renderer --help
python -m dexplicit.train_renderer --help

checkpoints.json and SHA256SUMS bind the canonical filenames to their unchanged tensor hashes. dexplicit-3d-config.json contains the DeXplicit 3D model configuration. Upstream Wan backbones, VAE, text encoder and tokenizer are obtained separately at pinned revisions. Optimizer states, raw datasets, MANO assets and DINOv3 weights are not redistributed.

Use and licensing

The DeXplicit 3D model and DeXplicit renderer weights, including the compact renderer, are licensed under Apache-2.0. Upstream models and datasets retain their own terms; see licensing and data sources.