diff --git a/README.md b/README.md
index 8a773e5..db83cf8 100755
--- a/README.md
+++ b/README.md
@@ -1,22 +1,38 @@
-# ApDepth: Aiming for Precise Monocular Depth Estimation Based on Diffusion Models
+
+
ApDepth
+
Aiming for Precise Monocular Depth Estimation Based on Diffusion Models
+
-This repository is based on [Marigold](https://marigoldmonodepth.github.io), CVPR 2024 Best Paper: [**Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation**](https://arxiv.org/abs/2312.02145)
+
+
+
+
+
+
+
+
-[](https://haruko386.github.io/research)
-[](https://www.apache.org/licenses/LICENSE-2.0)
-[](https://huggingface.co/developy/ApDepth)
-[](https://huggingface.co/spaces/developy/ApDepth)
+---
-
+
-
+## Overview
-> We present **ApDepth**, a deterministic single-step diffusion framework for monocular depth estimation. Built upon Marigold (derived from Stable Diffusion) and fine-tuned on synthetic datasets (Hypersim and Virtual KITTI), ApDepth overcomes the feature representation bottlenecks of standard diffusion models. Ultimately, our framework achieves **highly competitive geometric accuracy** and **exceptional object edge refinement**, all while delivering significantly accelerated inference speeds.
+**ApDepth** turns Stable Diffusion 2.1 into a deterministic, single-step
+monocular depth estimator. It combines a frozen Depth Anything V2 prior with a
+fine-tuned diffusion U-Net to recover accurate scene geometry and crisp object
+boundaries without iterative denoising.
-## 📢 News
+> [!IMPORTANT]
+>
+> ApDepth builds on [**Marigold**](https://marigoldmonodepth.github.io), the
+> CVPR 2024 Best Paper [*Repurposing Diffusion-Based Image Generators for
+> Monocular Depth Estimation*](https://arxiv.org/abs/2312.02145).
+
+## News
+
+- **2026-09-22:** Preparing **ApDepth V2-1** with direct SD2.1 fine-tuning,
+ SDWT and optional FFT refinement.
- **2026-04-06:** `ApDepth V2-0` is released!
- **2026-04-03:** We officially release the code for **ApDepth**! Stage 1 feature-alignment training is maintained in [ApDepth_Stage1](https://github.com/Haruko386/ApDepth_Stage1); this repository provides Stage 2 training, inference and evaluation.
- **2026-01-15:** We successfully introduce a spatial-preserving **Conv Adapter** and a **Cosine Similarity Loss** to enhance feature alignment, alongside a **Pixel-level $L_1$ Loss** to establish an accurate global metric scale.
@@ -26,26 +42,23 @@ This repository is based on [Marigold](https://marigoldmonodepth.github.io), CVP
- **2025-08-10:** Trying to make some optimizations in Feature Expression.
- **2025-05-08:** Clone `Marigold` to local.
-## 🚀 Usage
-
-**We offer several ways to interact with ApDepth**:
-
-1. A free online interactive demo is available here:
+## Quick start
-2. If you just want to see the examples, visit our gallery:
+Try ApDepth in the [live demo](https://huggingface.co/spaces/developy/ApDepth),
+browse the [project gallery](https://haruko386.github.io/research), or run it
+locally using the instructions below.
-3. Local development instructions with this codebase are given below.
+### Requirements
-## 🛠️ Setup
-The Model was trained on:
+The model was trained on:
- Ubuntu 22.04 LTS, Python 3.12.9, CUDA 11.8, `NVIDIA RTX 6000 Ada Generation`
-The inference code was tested on:
+Inference was tested on:
- Ubuntu 22.04 LTS, Python 3.12.9, CUDA 11.8, `NVIDIA GeForce RTX 4090`
-### 🪧 A Note for Windows users
+#### Windows
We recommend running the code in WSL2:
@@ -53,7 +66,7 @@ We recommend running the code in WSL2:
1. Install CUDA support for WSL following [installation guide](https://docs.nvidia.com/cuda/wsl-user-guide/index.html#cuda-support-for-wsl-2).
1. Find your drives in `/mnt//`; check [WSL FAQ](https://learn.microsoft.com/en-us/windows/wsl/faq#how-do-i-access-my-c--drive-) for more details. Navigate to the working directory of choice.
-### 📦 Repository
+### Installation
Clone the repository (requires git):
@@ -62,10 +75,9 @@ git clone https://github.com/Haruko386/ApDepth.git
cd ApDepth
```
-### 💻 Dependencies
+#### Conda
- **Using Conda:**
- Alternatively, create a Python native virtual environment and install dependencies into it:
+Create and activate the project environment:
```bash
conda create -n apdepth python==3.12.9
@@ -78,7 +90,7 @@ pip install -r requirements.txt
> Keep the environment activated before running the inference script.
> Activate the environment again after restarting the terminal session.
-### 🐳 Docker Setup
+### Docker
For a streamlined setup, we provide a Docker environment that pre-installs all necessary dependencies, including PyTorch, CUDA, and evaluation tools.
@@ -94,22 +106,24 @@ docker build -t apdepth:latest .
To utilize GPU acceleration, ensure the NVIDIA Container Toolkit is installed. We recommend mounting your local input and output directories to easily access your inference results:
-```Bash
+```bash
docker run --gpus all -it --rm \
-v $(pwd)/input:/workspace/ApDepth/input \
-v $(pwd)/output:/workspace/ApDepth/output \
apdepth:latest
-Once inside the container, the apdepth conda environment is activated by default, and you can directly execute the inference or training scripts.
```
-## 🏃 Testing on your images
-### 📷 Prepare images
+Inside the container, the `apdepth` Conda environment is activated automatically.
+
+## Inference
+
+### Prepare images
1. Use selected images under `input`
1. Or place your images in a directory, for example, under `input/test-image`, and run the following inference command.
-### 🎮 Run inference with paper setting
+### Paper setting
This setting corresponds to our paper. For academic comparison, please run with this setting.
@@ -124,7 +138,7 @@ python run.py \
You can find all results in `output/example-1`. Enjoy!
-### ⚙️ Inference settings
+### Inference options
The default settings are optimized for the best result. However, the behavior of the code can be customized:
@@ -140,7 +154,7 @@ The default settings are optimized for the best result. However, the behavior of
- `--color_map`: [Colormap](https://matplotlib.org/stable/users/explain/colors/colormaps.html) used to colorize the depth prediction. Default: Spectral. Set to `None` to skip colored depth map generation.
- `--apple_silicon`: Use Apple Silicon MPS acceleration.
-## 🦿 Evaluation on test datasets
+## Evaluation
Install additional dependencies:
@@ -180,11 +194,21 @@ You can get the result under `output/eval`
-## 🏋️ Training
+## Training
-The workflow is **Stage 1 feature alignment → Stage 2 main training → optional
-[VKITTI far-depth post-training](#post-training)**. Complete Stage 2, including
-its FFT refinement, before starting the optional post-training run.
+Three training methods are retained. The mixed SDWT + FFT method is the
+recommended recipe; the SDWT-only method is kept for controlled comparison, and the
+original Stage 1 + legacy FFT workflow remains reproducible.
+
+| Method | Initialization | Objective | Config |
+| --- | --- | --- | --- |
+| 1. Mixed SDWT + FFT (recommended) | Original SD2.1 | MSE + pixel L1 + gradient, then gradual SDWT + low-weight FFT | `config/train_sd2_sdwt_fft.yaml` |
+| 2. SDWT only | Original SD2.1 | MSE + pixel L1 + gradient, then gradual SDWT | `config/train_sd2_sdwt.yaml` |
+| 3. Legacy ApDepth/FFT | Stage 1 U-Net checkpoint | Reconstruction, then the original latent FFT transition | `config/train_apdepth.yaml` |
+
+Methods 1 and 2 are single-run Stage 2 training methods and do not execute or
+load Stage 1. Their implementation details and ablations are documented
+[here](doc/sd2_sdwt_training.md).
Based on the previously created environment, install extended requirements:
@@ -207,19 +231,62 @@ Prepare [Hypersim](https://github.com/apple/ml-hypersim) and [Virtual KITTI 2](h
------------
-**Stage 1: feature alignment pre-training**
+### Method 1: direct SD2 training with SDWT + FFT (recommended)
-Stage 1 is maintained separately in
-[Haruko386/ApDepth_Stage1](https://github.com/Haruko386/ApDepth_Stage1).
-Follow that repository's setup and training instructions. This repository contains
-only the main Stage 2 trainer; it does not include Stage 1 code or dependencies.
+This method starts from the original SD2.1 weights. During iterations 0-8,000,
+it learns the basic depth mapping with latent MSE, pixel L1 and gradient loss.
+From iteration 8,000 to 12,000, SDWT increases from 0 to 1.0 while latent
+FFT increases from 0 to 0.2. The reconstruction losses remain active for the
+rest of training, and the VAE decoder is fine-tuned from iteration 8,000.
-------------
+SDWT matches local depth distributions with an explicit pixel-displacement
+cost. It subtracts entropic self-costs, evaluates both regular and half-window
+shifted grids, and weights each window by its valid support. These changes keep
+limited tolerance to noisy boundary labels without making the loss invariant to
+arbitrary permutations inside a window.
+
+```bash
+python train.py --config config/train_sd2_sdwt_fft.yaml --no_wandb
+```
+
+Do not pass `--init_checkpoint`. Resume an interrupted run with:
+
+```bash
+python train.py \
+ --resume_run output/train_sd2_sdwt_fft/checkpoint/latest --no_wandb
+```
+
+### Method 2: direct SD2 training with SDWT only
+
+This method uses the same original SD2.1 initialization, reconstruction-loss
+schedule and VAE decoder adaptation as Method 1, but leaves latent FFT disabled.
+It is the direct comparison for measuring the contribution of FFT.
+
+```bash
+python train.py --config config/train_sd2_sdwt.yaml --no_wandb
+```
+
+Do not pass `--init_checkpoint`. Resume an interrupted run with:
+
+```bash
+python train.py \
+ --resume_run output/train_sd2_sdwt/checkpoint/latest --no_wandb
+```
-**Stage 2: main ApDepth training**
+For both direct-SD2 methods, `BASE_CKPT_DIR/stable-diffusion-2-1` must contain
+the original SD2.1 Diffusers pipeline. Their checkpoints save both the U-Net and
+the fine-tuned VAE.
-To initialize Stage 2 from an existing Stage 1 checkpoint, start a new run with
-`--init_checkpoint`. This loads only the U-Net; optimizer and iteration counters
+### Method 3: legacy Stage 1 + ApDepth/FFT training (ApDepth v2-0)
+
+First complete the feature-alignment Stage 1 training. Stage 1 is maintained
+separately in
+[Haruko386/ApDepth_Stage1](https://github.com/Haruko386/ApDepth_Stage1).
+Follow that repository's setup and training instructions. This repository contains
+the legacy Stage 2 trainer, but does not include Stage 1 code or dependencies.
+
+Then initialize the legacy Stage 2 from the Stage 1 checkpoint with
+`--init_checkpoint`. This loads the U-Net and a saved VAE if present; optimizer and iteration counters
start fresh using the Stage 2 config. `config/train_apdepth.yaml` keeps far-depth
supervision disabled throughout the main training stage:
@@ -232,12 +299,6 @@ The initialization directory must contain
`unet/diffusion_pytorch_model.safetensors`. `--init_checkpoint` rejects `.bin`
checkpoints.
-To train the main stage directly from the base model without Stage 1 initialization:
-
-```bash
-python train.py --config config/train_apdepth.yaml --no_wandb
-```
-
The transition to latent frequency loss is controlled by
`latent_freq_loss.gradual_transition` in `config/train_apdepth.yaml`. When `true`,
the reconstruction loss weight decreases linearly from 1 to 0 between iteration
@@ -252,13 +313,30 @@ Resume from a checkpoint, e.g.
python train.py --resume_run output/train_apdepth/checkpoint/latest --no_wandb
```
+### Inference after Method 1 or Method 2
+
+Direct-SD2 checkpoints are training-component checkpoints, so inference combines
+the original SD2.1 pipeline with the saved U-Net and VAE. For Method 1:
+
+```bash
+python run.py --checkpoint "${BASE_CKPT_DIR}/stable-diffusion-2-1" \
+ --training_checkpoint output/train_sd2_sdwt_fft/checkpoint/iter_021000 \
+ --input_rgb_dir input/example-1 --output_dir output/sd2_sdwt_fft \
+ --ensemble_size 1 --processing_res 0
+```
+
+For Method 2, replace `train_sd2_sdwt_fft` with `train_sd2_sdwt` in the checkpoint
+path. `infer.py` also accepts `--training_checkpoint` for dataset evaluation.
+
------------
-## 🌤️ Optional post-training after Stage 2
+## Optional post-training after Stage 2
**Start this step only after Stage 2 training has finished.** The new VKITTI
far-depth supervision feature is enabled by `config/train_sky_finetune.yaml` for
-a separate post-training run initialized from the completed Stage 2 U-Net.
+a separate post-training run initialized from a completed checkpoint from any
+of the three methods. For Methods 1 and 2, the saved VAE is loaded and retained
+together with the U-Net.
Measured depths in [80, 655.35] m supply an additional clipped far-depth target
before VAE encoding; missing depth remains excluded. Evaluation masks and
@@ -279,12 +357,19 @@ If preprocessing replaced the far-plane values with zero, restore the original
depth files. Zero depth is not treated as sky. Adjust each dataset config's `dir`
to match the actual layout under `BASE_DATA_DIR`.
-Start a new post-training run from the final Stage 2 checkpoint (adjust the path
-if Stage 2 was saved elsewhere):
+Select the corresponding completed checkpoint:
+
+| Method | Example checkpoint |
+| --- | --- |
+| 1. SDWT + FFT | `output/train_sd2_sdwt_fft/checkpoint/iter_021000` |
+| 2. SDWT only | `output/train_sd2_sdwt/checkpoint/iter_021000` |
+| 3. Legacy ApDepth/FFT | `output/train_apdepth/checkpoint/iter_021000` |
+
+Then start a new post-training run, adjusting the path if needed:
```bash
python train.py --config config/train_sky_finetune.yaml \
- --init_checkpoint output/train_apdepth/checkpoint/iter_021000 --no_wandb
+ --init_checkpoint /path/to/completed/checkpoint/iter_021000 --no_wandb
```
This resets the optimizer, learning-rate schedule and iteration counter, then
@@ -292,7 +377,7 @@ runs 6000 new optimization steps at LR 5e-6 with 100 warmup steps and saves
backups every 1000 steps in `output/train_sky_finetune/checkpoint`. It retains the
90% Hypersim / 10% VKITTI sampling, original batch settings and periodic validation.
These 6000 steps use latent MSE + pixel L1 reconstruction plus the new far-depth
-loss; the Stage 2 FFT refinement has already finished. Initialization requires
+loss; the selected Stage 2 objective has already finished. Initialization requires
`unet/diffusion_pytorch_model.safetensors` in the supplied checkpoint directory.
If this post-training run is interrupted, resume its own checkpoint:
@@ -312,7 +397,11 @@ alone does not establish a metric improvement.
## Evaluating trained checkpoints
-Stage 2 and optional post-training update and save the U-Net. For inference, copy a complete
+The new direct-SD2 recipe saves both U-Net and VAE; use `--training_checkpoint`
+as shown above to load them together. The following instructions apply to legacy
+U-Net-only checkpoints.
+
+Legacy Stage 2 and optional post-training update and save the U-Net. For inference, copy a complete
ApDepth pipeline checkpoint to a new directory, then replace that copy's `unet/`
folder with the selected training checkpoint's `unet/`. The pipeline's VAE,
tokenizer, text encoder and scheduler are still required; a training checkpoint
@@ -324,11 +413,11 @@ alone is not a complete inference pipeline. Then refer to [evaluation](#evaluati
-## ✏️ Contributing
+## Contributing
Please refer to [this](CONTRIBUTING.md) instruction.
-## 🤔 Troubleshooting
+## Troubleshooting
| Problem | Solution |
| --------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------- |
@@ -341,7 +430,7 @@ Please refer to [this](CONTRIBUTING.md) instruction.
-## 🎓 Citation
+## Citation
Please cite our paper:
```bibtex
@@ -353,10 +442,8 @@ Please cite our paper:
}
```
-## 🎫 License
+## License
This work is licensed under the Apache License, Version 2.0 (as defined in the [LICENSE](LICENSE.txt)).
-By downloading and using the code and model you agree to the terms in the [LICENSE](LICENSE.txt).
-
-[](https://www.apache.org/licenses/LICENSE-2.0)
+By downloading and using the code and model you agree to the terms in the [LICENSE](LICENSE.txt).
diff --git a/doc/badges/v21-demo.svg b/doc/badges/v21-demo.svg
new file mode 100644
index 0000000..d821362
--- /dev/null
+++ b/doc/badges/v21-demo.svg
@@ -0,0 +1,7 @@
+
diff --git a/doc/badges/v21-paper.svg b/doc/badges/v21-paper.svg
new file mode 100644
index 0000000..bce8d0a
--- /dev/null
+++ b/doc/badges/v21-paper.svg
@@ -0,0 +1,7 @@
+
diff --git a/doc/badges/v21-training.svg b/doc/badges/v21-training.svg
new file mode 100644
index 0000000..524748a
--- /dev/null
+++ b/doc/badges/v21-training.svg
@@ -0,0 +1,7 @@
+
diff --git a/doc/badges/v21-version.svg b/doc/badges/v21-version.svg
new file mode 100644
index 0000000..c127127
--- /dev/null
+++ b/doc/badges/v21-version.svg
@@ -0,0 +1,7 @@
+
diff --git a/doc/badges/v21-website.svg b/doc/badges/v21-website.svg
new file mode 100644
index 0000000..8ba559a
--- /dev/null
+++ b/doc/badges/v21-website.svg
@@ -0,0 +1,7 @@
+
diff --git a/doc/badges/v21-weights.svg b/doc/badges/v21-weights.svg
new file mode 100644
index 0000000..7c7ccc0
--- /dev/null
+++ b/doc/badges/v21-weights.svg
@@ -0,0 +1,7 @@
+
diff --git a/doc/cover.jpg b/doc/cover.jpg
deleted file mode 100644
index a445bff..0000000
Binary files a/doc/cover.jpg and /dev/null differ
diff --git a/doc/sd2_sdwt_training.md b/doc/sd2_sdwt_training.md
new file mode 100644
index 0000000..21ed4e5
--- /dev/null
+++ b/doc/sd2_sdwt_training.md
@@ -0,0 +1,140 @@
+# 从原版 SD2 直接训练：SDWT 二阶段方案
+
+本方案从原始 Stable Diffusion 2.1 权重直接开始一次连续训练，不依赖一阶段
+checkpoint。前 8,000 步使用 latent MSE、pixel L1 和梯度损失建立基本深度映射；
+8,000 至 12,000 步逐渐加入本文的 **Spatially Debiased Window Transport
+(SDWT) Loss** 并解冻 VAE decoder。
+
+SDWT 受到 Marigold V2 局部最优传输思路的启发，但不是 SinkLoss 的改名版本。
+它针对原方法的四个限制重新定义了代价、目标函数、窗口划分和聚合方式。
+
+## 原局部匹配的限制
+
+1. **忽略空间位置。** 仅使用深度差构造代价时，同一窗口内任意像素排列都可能
+ 获得近似相同的损失。它能容忍边缘标注错位，也可能接受错误的局部几何。
+2. **熵偏差。** 带熵正则的交叉传输成本在预测等于真值时通常仍不为零，损失的
+ 数值和梯度会受温度影响。
+3. **固定窗口接缝。** 非重叠窗口无法匹配落在相邻窗口两侧的同一条边缘。
+4. **稀疏窗口权重过大。** 按非空窗口平均时，只有一个有效像素的窗口与完整窗口
+ 权重相同。
+
+## SDWT
+
+对窗口中的预测深度 `p_i`、目标深度 `g_j` 和归一化坐标 `x_i`、`x_j`，
+SDWT 使用联合代价
+
+```text
+C_ij = rho(p_i - g_j) + lambda_s * ||x_i - x_j||_1
+rho(e) = sqrt(e^2 + epsilon^2) - epsilon
+```
+
+其中空间项限制远距离交换，Charbonnier 项在零点平滑，同时保留对异常深度误差
+的鲁棒性。无效像素在构造传输矩阵之前被精确移除，不再用极大哨兵值参与
+Sinkhorn 归一化。
+
+令带熵正则的局部传输目标为 `OT_tau(A, B)`，最终目标使用去偏形式
+
+```text
+SDWT(A, B) = OT_tau(A, B)
+ - 0.5 * OT_tau(A, A)
+ - 0.5 * OT_tau(B, B)
+```
+
+因此相同预测和目标的损失为零。每张图同时计算原始窗口网格和偏移半个窗口的
+网格，以覆盖原网格边界；最后按每个窗口的有效像素数聚合。默认参数为 5×5
+窗口、10 次传输归一化、温度 0.1、空间权重 0.15。
+
+去偏项和双网格需要额外的传输计算，因此 SDWT 会比单次局部匹配更慢。实现按
+有效像素数分组、分块计算并使用激活重算，目标自传输项不建立梯度图，以限制显存
+增长。论文实验应同时报告训练吞吐与峰值显存。
+
+## 总训练目标
+
+令已完成更新数为 `s`：
+
+```text
+r = clip((s - 8000) / 4000, 0, 1)
+
+L = (1 - 0.5r) L_latent_MSE
+ + (1 - 0.75r) L_pixel_L1
+ + (0.5 - 0.4r) L_gradient
+ + r L_SDWT
+```
+
+- `s < 8000`：训练 U-Net，VAE 保持冻结。
+- `8000 <= s < 12000`：逐渐加入 SDWT，并开始微调 decoder 和
+ `post_quant_conv`。
+- `s >= 12000`：保留全部基础重建项，SDWT 权重保持 1.0。
+
+推荐配置额外将 latent FFT 从 0 渐增到 0.2。FFT 约束全局频谱，SDWT 处理带
+有限位移容忍度的局部几何，两者作用不同。上述超参数是 SD2 的实验起点，需要
+通过验证集和消融实验确定最终论文设置。
+
+## 训练
+
+推荐的 SDWT + FFT：
+
+```bash
+python train.py --config config/train_sd2_sdwt_fft.yaml --no_wandb
+```
+
+不使用 FFT 的对照：
+
+```bash
+python train.py --config config/train_sd2_sdwt.yaml --no_wandb
+```
+
+这两个配置都不能传入 `--init_checkpoint`。恢复训练：
+
+```bash
+python train.py \
+ --resume_run output/train_sd2_sdwt_fft/checkpoint/latest --no_wandb
+```
+
+推理时组合原始 SD2.1、训练后的 U-Net 和 VAE：
+
+```bash
+python run.py --checkpoint /path/to/stable-diffusion-2-1 \
+ --training_checkpoint output/train_sd2_sdwt_fft/checkpoint/iter_021000 \
+ --input_rgb_dir input/example-1 --output_dir output/sd2_sdwt_fft \
+ --ensemble_size 1 --processing_res 0
+```
+
+## 论文消融
+
+所有实验应保持初始化、数据、随机种子、训练步数与评估协议一致。
+
+| 配置 | 验证内容 |
+| --- | --- |
+| `config/train_sd2_sdwt_fft.yaml` | 完整 SDWT + FFT |
+| `config/train_sd2_sdwt.yaml` | 去除 FFT |
+| `config/ablation/sd2_sdwt_no_transport.yaml` | 去除整个 SDWT 项 |
+| `config/ablation/sd2_sdwt_no_spatial.yaml` | 去除空间位移代价 |
+| `config/ablation/sd2_sdwt_no_debias.yaml` | 去除自成本校正 |
+| `config/ablation/sd2_sdwt_single_grid.yaml` | 去除半窗口偏移网格 |
+| `config/ablation/sd2_sdwt_equal_windows.yaml` | 改回非空窗口等权 |
+| `config/ablation/sd2_sdwt_no_decoder.yaml` | 保持 VAE decoder 冻结 |
+
+除常规 AbsRel 和 delta1 外，应报告边缘指标、细结构区域指标、不同有效像素率区间
+的误差，以及训练显存和耗时。只有在完整方法相对各项消融取得可重复改进后，才能
+把这些设计写成实证贡献。
+
+TensorBoard 记录 `train/sdwt_loss`、`train/sdwt_weight`、
+`train/gradient_loss`、`train/latent_mse`、`train/pixel_l1` 和
+`train/freq_loss_weight`。
+
+CPU 验证：
+
+```bash
+python -m pytest tests/test_sdwt_training.py -q
+```
+
+## 方法来源
+
+局部最优传输的研究方向来自：
+
+> Pavlovic et al., “Marigold V2: Revisiting Diffusion Transformers for
+> Monocular Depth Estimation,” 2026, arXiv:2609.08084.
+
+论文写作中应明确引用该工作，并将 SDWT 表述为空间约束、去偏、双网格和
+有效支持聚合方面的扩展。
diff --git a/input/example-0/01.jpg b/input/example-0/01.jpg
new file mode 100644
index 0000000..9721ec4
Binary files /dev/null and b/input/example-0/01.jpg differ
diff --git a/input/example-0/02.png b/input/example-0/02.png
new file mode 100644
index 0000000..b1098e5
Binary files /dev/null and b/input/example-0/02.png differ
diff --git a/input/example-0/03.jpg b/input/example-0/03.jpg
new file mode 100644
index 0000000..253d9da
Binary files /dev/null and b/input/example-0/03.jpg differ
diff --git a/input/example-0/04.jpg b/input/example-0/04.jpg
new file mode 100644
index 0000000..10d877d
Binary files /dev/null and b/input/example-0/04.jpg differ
diff --git a/input/example-0/05.jpg b/input/example-0/05.jpg
new file mode 100644
index 0000000..fab3862
Binary files /dev/null and b/input/example-0/05.jpg differ
diff --git a/input/example-0/06.jpg b/input/example-0/06.jpg
new file mode 100644
index 0000000..a1c5482
Binary files /dev/null and b/input/example-0/06.jpg differ
+ 
+ 
+ 
+ 
+ 
+
+## Quick start
-2. If you just want to see the examples, visit our gallery: 
+Try ApDepth in the [live demo](https://huggingface.co/spaces/developy/ApDepth),
+browse the [project gallery](https://haruko386.github.io/research), or run it
+locally using the instructions below.
-3. Local development instructions with this codebase are given below.
+### Requirements
-## 🛠️ Setup
-The Model was trained on:
+The model was trained on:
- Ubuntu 22.04 LTS, Python 3.12.9, CUDA 11.8, `NVIDIA RTX 6000 Ada Generation`
-The inference code was tested on:
+Inference was tested on:
- Ubuntu 22.04 LTS, Python 3.12.9, CUDA 11.8, `NVIDIA GeForce RTX 4090`
-### 🪧 A Note for Windows users
+#### Windows
We recommend running the code in WSL2:
@@ -53,7 +66,7 @@ We recommend running the code in WSL2:
1. Install CUDA support for WSL following [installation guide](https://docs.nvidia.com/cuda/wsl-user-guide/index.html#cuda-support-for-wsl-2).
1. Find your drives in `/mnt/