---
license: cc-by-nc-sa-4.0
tags:
  - panoptic-segmentation
  - unsupervised
  - cityscapes
  - mirror
---

# mbps-cups-e1

Public mirror of the two CUPS-released checkpoints needed to reproduce **CUPS Stage-1 (pseudo-label generation)** as part of E1 -- the "same-backbone CUPS rerun" control experiment for the MBPS NeurIPS 2026 paper.

## Files

| File | Size | SHA256 | Origin |
|---|---|---|---|
| `depthg.ckpt`     | 352 639 220 bytes | `09556e83b791a8177a369c218800d54bc8bfa77a61759840d9758f6211ecab39` | TUdatalib `4532/depthg.ckpt` (CUPS Eccv-W release) |
| `raft_smurf.pt`   |  21 049 890 bytes | `692f1deece783ae69d058608b18ddeb97096c2505bd4529a1a08b8f5cab20d47` | TUdatalib `4532/raft_smurf.pt` (SMURF official, converted by Christoph Reich) |

Original checkpoints: <https://tudatalib.ulb.tu-darmstadt.de/handle/tudatalib/4532>
Original CUPS repo: <https://github.com/robot-learning-freiburg/CUPS>

## Why this mirror exists

E1 reruns CUPS Stage-1 on an RTX A6000 (48 GB) -- the original GTX 1080 Ti
(11 GB) OOMs in SF2SE3 instance proposal generation. The mirror lets the
A6000 launcher pull both inputs with `wget` (no auth, no `huggingface-cli`).

## Usage

```bash
wget https://huggingface.co/qbit-glitch/mbps-cups-e1/resolve/main/depthg.ckpt
wget https://huggingface.co/qbit-glitch/mbps-cups-e1/resolve/main/raft_smurf.pt
```

The full launcher (`scripts/e1_stage1_pseudolabel_gen_a6000.sh` in the
MBPS repo) does this for you.

## NOT mirrored

- `cups.ckpt` -- this is the AFTER-Stage-3 final panoptic checkpoint.
  Using it as a Stage-1 input is circular. If you need it for inference
  comparison, fetch it directly from TUdatalib.
- `dino_RN50_pretrain_d2_format.pkl` -- only needed for Stage-2 backbone
  init in CUPS-original. E1 swaps in DINOv3 ViT-B/16 instead.

## License

These are mirrored CUPS / SMURF artifacts. Their licenses (CC BY-NC-SA 4.0,
research use) come from the original TUdatalib / SMURF releases and apply
unchanged to this mirror.