---
license: cc-by-nc-4.0
library_name: pytorch
pipeline_tag: depth-estimation
tags:
  - bokeh
  - depth-estimation
  - monocular-depth
  - flux
  - lora
  - unidepth
  - icml-2026
---

# BokehDepth — Released Checkpoints

This repository hosts the released checkpoints for **BokehDepth: Boosting Monocular Metric Depth Estimation via Bokeh Rendering (ICML 2026)**.

- 📄 Paper: <https://arxiv.org/abs/2512.12425>
- 🌐 Project page: <https://fogradio.github.io/BokehDepth_Project/>
- 💻 Code: <https://github.com/fogradio/BokehDepth>  

BokehDepth is a **two-stage** framework. Stage-1 turns a single sharp image into a calibrated multi-strength bokeh stack (no depth map needed); Stage-2 fuses the resulting defocus cues to produce sharper, more reliable **metric** depth. The three files in this repository correspond exactly to the two-stage inference pipeline.

| File | Stage | Role | Size |
|---|---|---|---|
| [`bokeh_lora.bin`](./bokeh_lora.bin) | Stage-1 | Bokeh generation LoRA adapter on top of FLUX.1-Kontext | ≈ 556 MB |
| [`bokeh_lora_ft.bin`](./bokeh_lora_ft.bin) | Stage-1 | Robustness-finetuned variant of `bokeh_lora.bin` | ≈ 556 MB |
| [`UDv2_dsfa_release.pth`](./UDv2_dsfa_release.pth) | Stage-2 | UniDepthV2 + DSFA depth estimator | ≈ 5.19 GB |

---

## Stage-1 — Bokeh Generation LoRA

Stage-1 uses **FLUX.1-Kontext** (rectified-flow MMDiT) plus a lightweight bokeh cross-attention adapter. Heterogeneous optical settings (focal length, aperture, focus distance) collapse into a single calibrated scalar **K** from the thin-lens circle-of-confusion model, which captures the near-linear relation `r ≈ K · Δdisp` between blur radius and disparity offset. Conditioned on K, Stage-1 turns one sharp image into a multi-strength bokeh stack with no depth map at any point.

### `bokeh_lora.bin` — base LoRA

The base Stage-1 checkpoint. Trained on the unified Stage-1 data pipeline that aligns real defocused photos, synthetic renderings, and paired datasets onto the shared **K** axis. 

### `bokeh_lora_ft.bin` — robustness fine-tune

A continued fine-tune of `bokeh_lora.bin` that additionally mixes in **synthetic bokeh renderings produced by [BokehMe](https://github.com/JuewenPeng/BokehMe)** from subsets of the standard monocular-depth datasets **KITTI / Hypersim / NYU-v2 / vKITTI 2**. Since these datasets cover many scenes where the foreground is ambiguous, low-contrast, or simply absent, the resulting checkpoint is noticeably more robust at generating clean bokeh on such "no-clear-subject" inputs (driving scenes, dense indoor clutter, distant cityscapes, etc.) while preserving the calibrated K-control of the base LoRA.

Both LoRAs are wrapped at inference time by `BokehFluxControlAdapter` (see [`bokeh-generation/model/bokeh_adapter_flux.py`](https://github.com/...) in the code repository) and are loaded with `lora_rank=128`, `lora_alpha=128` over FLUX transformer blocks `0–56`.

---

## Stage-2 — UniDepthV2-DSFA

### `UDv2_dsfa_release.pth`

The Stage-2 metric depth model: **UniDepthV2** (ViT-L/14 DINOv2 backbone) with our **Divided Space Focus Attention (DSFA)** module inserted into the depth encoder. DSFA first runs spatial attention inside each frame conditioned on that frame's blur strength `K_f`, then runs focus attention across frames at matching spatial locations, modulated by FiLM. Each location can therefore read how its blur grows with K, which is the physical depth-from-defocus cue. Only reference-frame tokens are passed downstream, so the original DPT decoder and metric head stay untouched.

This checkpoint is the **plug-and-play DSFA build** dropped onto UniDepthV2 and trained jointly with the Stage-1 bokeh stack as input. Use it together with the config `UniDepth/configs/config_v2_vitl14_DSFA_inference.json` in the code repository.

---

## How to use

```bash
# from the project root
bash run_inference.sh
```

`run_inference.sh` expects all three files to live exactly here, i.e. under `weights/`:

```
weights/
├── bokeh_lora.bin          # or bokeh_lora_ft.bin (see ADAPTER_CKPT env var)
├── bokeh_lora_ft.bin
└── UDv2_dsfa_release.pth
```

Override which Stage-1 LoRA is used with:

```bash
ADAPTER_CKPT=weights/bokeh_lora_ft.bin bash run_inference.sh   # robust default
ADAPTER_CKPT=weights/bokeh_lora.bin    bash run_inference.sh   # base LoRA
```

The Stage-2 weights path is fixed via `WEIGHTS_PATH=weights/UDv2_dsfa_release.pth` (default).

---

## Citation

If you use these checkpoints, please cite:

```bibtex
@inproceedings{zhang2026bokehdepth,
  title     = {Boosting Monocular Metric Depth Estimation via Bokeh Rendering},
  author    = {Zhang, Hangwei and Fortes, Armando and Wei, Tianyi and Pan, Xingang},
  booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
  year      = {2026}
}
```

## License & acknowledgements

Released under **CC BY-NC 4.0** for research use only. Stage-1 builds on **FLUX.1-Kontext** (Black Forest Labs) and Stage-2 builds on **UniDepthV2**; both upstream licenses apply to their respective base weights. The robustness fine-tune additionally relies on synthetic bokeh produced by **BokehMe** on standard monocular-depth datasets (KITTI / Hypersim / NYU-v2 / vKITTI 2) — please respect each dataset's individual license when redistributing derived data.