README.md

---
license: other
license_name: originlab-noncommercial-research
license_link: LICENSE.md
extra_gated_heading: "Request access to OriginLab Lotus Game-Depth models"
extra_gated_prompt: "By requesting access you agree to CC-BY-NC-4.0 plus a model-release requirement: non-commercial research use only, and any model you derive must be publicly released with open weights and a model card. Commercial use requires a separate agreement with OriginLab."
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
  "I will use these models for non-commercial research only": checkbox
tags:
- depth-estimation
- monocular-depth
- lotus
- diffusion
pipeline_tag: depth-estimation
datasets:
- originlab/game-depth
---

<p align="center">
  <img src="https://huggingface.co/originlab/lotus-game-depth/resolve/main/assets/logo.png" alt="OriginLab" width="320">
</p>

# OriginLab Lotus Game-Depth (pretrained + NYU fine-tuned)

Website: [originlab.ai](https://originlab.ai)

Two Lotus-recipe latent-diffusion depth checkpoints (SD2-base UNet, 8-channel conv_in, single-step x0 at t=999, trunc_disparity), in one repo:

- **`pretrained/`** - trained from scratch on the **[OriginLab Game-Depth](https://huggingface.co/datasets/originlab/game-depth)** dataset (game-engine z-buffers), no real data. Zero-shot KITTI AbsRel **0.191** (Lotus 0.224, Marigold 0.244).
- **`nyu-ft/`** - the above fine-tuned on real NYU Depth V2. NYU AbsRel **0.116**, on par with a fairly-tuned Lotus baseline (0.115) using 0% indoor pretraining data and ~4x fewer frames.

**Dataset:** https://huggingface.co/datasets/originlab/game-depth

## Load

```python
from diffusers import UNet2DConditionModel
# game-pretrained
unet = UNet2DConditionModel.from_pretrained("originlab/lotus-game-depth", subfolder="pretrained/unet")
# NYU fine-tuned
unet = UNet2DConditionModel.from_pretrained("originlab/lotus-game-depth", subfolder="nyu-ft/unet")
```

Each subfolder contains the full pipeline (unet, vae, text_encoder, scheduler, ...); run with the Lotus single-step depth pipeline.

## License
Non-commercial research use with a model-release clause (see LICENSE).