Code map
The public API / CLI / Space separation follows the organization of TencentARC/GAE-GeometricAutoEncoder (reference inspected 2026-10-15). PURE remains a reconstruction/post-training project; no GAE generation model or sampling code is included.
| Path | Responsibility |
|---|---|
pure/hub.py |
PURE, PUREConfig, strict HF loading, lossless saving |
pure/processing.py |
RGB preprocessing and consecutive video frames |
pure/pipeline.py |
Shared inference → optional sky masks → exports |
pure/sky.py |
Pinned SegFormer RGB sky segmentation |
pure/export.py |
Full filtered PLY, preview GLB, videos and camera files |
pure/models/ |
Pi3 v2 / DA3 Giant / VGGT adapters |
pure/_vendor/ |
Upstream network components, with original notices |
pure/loss.py, pure/utils/ |
Original public PURE objective and geometry |
pure/training/ |
Teacher caches, optimization, DDP and checkpoint resume |
scripts/demo/reconstruct.py |
Model-folder or Hub inference CLI |
scripts/train/train.py |
Training CLI; forwards to the existing tested engine |
scripts/data/prepare_cache.py |
Training-cache preparation CLI |
scripts/hub/convert_weights.py |
Three lossless model-folder conversions |
scripts/hub/upload.py |
Explicit upload of converted model folders |
configs/ |
Training configurations |
app.py |
Gradio GPU Space; uses the same pipeline as the CLI |
tests/ |
Training, adapter, serialization, preprocessing and export checks |
recovery/ |
Original recovery evidence; not imported at runtime |
Existing scripts/train.py, scripts/infer.py, scripts/prepare_cache.py and
Pi3 aliases remain available. Training checkpoint keys are unchanged. The new
PURE API wraps a backbone, while HF weight files retain the backbone's original
keys. To resume the training CLI from a single converted Safetensors file, use
--checkpoint hf_models/pi3/model.safetensors with the matching training config.
Training optimizer state remains in training checkpoints, not model folders.