---
license: cc-by-nc-4.0
library_name: sceneworks
base_model: facebook/sapiens-normal-0.3b
tags: [sapiens, surface-normals]
---

# Sapiens normal 0.3B (SceneWorks safetensors rehost)

Safetensors re-container of Meta's Sapiens surface-normal 0.3B checkpoint, used by SceneWorks' MLX and Candle trainers for the frozen, differentiable surface-normal loss in character LoRA training.

## Provenance
- Source: [facebook/sapiens-normal-0.3b](https://huggingface.co/facebook/sapiens-normal-0.3b) at revision `ada60522133d083464e5b70172adab7568ce6f49`, file `sapiens_0.3b_normal_render_people_epoch_66.pth`.
- Source sha256: `47792bbed2d3ea7537970230d5c6c7fb2b45c89bac384435bfbd6b2db03efb61` (1,358,042,350 bytes).
- License: CC-BY-NC-4.0, inherited from the source (non-commercial use only). Attribution: Meta Platforms, "Sapiens: Foundation for Human Vision Models".
- Conversion: `torch.load(weights_only=True)` → the `state_dict` tensors saved with `safetensors.torch.save_file`, keys UNCHANGED (`backbone.*`, `decode_head.*`); no tensor renamed, cast or dropped (the non-tensor `meta` entry is omitted). 304 tensors. Equivalent to `scripts/reference/body_losses_reference.py --convert` in the SceneWorks inference repo.

`model.safetensors` sha256: `b37670417447b6da441111e0c950bf9e694bf92777616d81ee72b82d2f9387fc`