README.md

---
license: bsd
---
# NeuMERL

The Neural Augmented [MERL dataset](https://www.merl.com/research/downloads/BRDF) (NeuMERL, 2400 BRDFs) dataset adopted in the paper [NeuMaDiff: Neural Material Synthesis via Hyperdiffusion](https://arxiv.org/abs/2411.12015).

Please download it at [./NeuMERL-2400.npy](./NeuMERL-2400.npy), with Pytorch weights of shape (2400, 675).

Alternatively, please download them separately (for i from 1 to 24) at [./NeuMERL(24*100)/mlp_weights_all_{i}.npy](https://huggingface.co/datasets/Peter2023HuggingFace/NeuMERL/tree/main/NeuMERL(24*100)).

![img-visual](./img/visual.png)

## Usage

The dataset is stored in a numpy file, with all the MLP weights linearized in a single array. For NBRDF MLP architecture use, please refer to [./nbrdf-release.py](./nbrdf-release.py)

- Input: Cartesian coordinate for positional samples (hx, hy, hz, dx, dy, dz).
- Output: MERL reflectance value.
- To convert it back to model weights `.pth`, update the example [example-brdf.pth](./example-brdf.pth) (alum-bronze).
  - To then convert it back to MERL binary, load the model from `.pth` and record the inference results with all positional input.
  - For details, please refer to [Github](https://github.com/PeterHUistyping/M3ashy) repo.

```
fc1.weight       (21, 6)
fc1.bias         (21,)
fc2.weight       (21, 21)
fc2.bias         (21,)
fc3.weight       (3, 21)
fc3.bias         (3,)
```

## Dataset formation

To form the AugMERL dataset, we first augment the original [MERL dataset](https://www.merl.com/research/downloads/BRDF) with color channel permutation. The first group materials (1-600) are all without interpolation. Then, we augment the BRDFs via direct linear interpolation, forming three groups of materials (601-1200, 1201-1800, 1801-2400), where each group follows the same color channel permutation as the first group ([Sec 3.1](https://arxiv.org/abs/2411.12015)).

Next, we adopt neural fields as a low-dimensional, continuous representation for materials, fitting them to individual materials in Aug-MERL to create a new dataset of neural material representations, Neural MERL (NeuMERL). For weight initialization, we use the same seed for all materials, see [mlp_weights_ini.pth](./mlp_weights_ini.pth).

### Color channel permutation

Color channel permutation: RGB (1-100), RBG (101-200), GRB (201-300), GBR (301-400), BRG (401-500), GRB (501-600), similarly afterwards.

```python
file_index = mat_id // 100
rgb_type = file_index % 6
if rgb_type == 1:   brdf = brdf[..., [0, 1, 2]]  # merl_1
elif rgb_type == 2: brdf = brdf[..., [0, 2, 1]]  # merl_2
elif rgb_type == 3: brdf = brdf[..., [1, 0, 2]]  # merl_3
elif rgb_type == 4: brdf = brdf[..., [1, 2, 0]]  # merl_4
elif rgb_type == 5: brdf = brdf[..., [2, 0, 1]]  # merl_5
elif rgb_type == 0: brdf = brdf[..., [2, 1, 0]]  # merl_6
```

### BRDF interpolation

For material from 601 to 2400, we interpolate the BRDFs of two materials from the original MERL dataset, forming a new BRDF.

- For 601-1200, i=0;
- For 1201-1800, i=1;
- For 1801-2400, i=2.

Please refer to [./interpolate/interpolate_interpolation_{i}.txt](https://huggingface.co/datasets/Peter2023HuggingFace/NeuMERL/tree/main/interpolate), each line is in the format of

```
MERL_BRDF1_name MERL_BRDF2_name \alpha 
```

, where $\alpha \in [0.3, 0.7]$ is the linear interpolation coefficient, outputting the interpolated BRDF

$$
f^{*}_r = \alpha \times f^{1}_r + (1-\alpha) \times f^{2}_r
$$

## Citation

If you found the paper or base model useful, please consider citing,

```
@inproceedings{
    M3ashy2026, 
    author = {Chenliang Zhou and Zheyuan Hu and Alejandro Sztrajman and Yancheng Cai and Yaru Liu and Cengiz Oztireli}, 
    title = {M$^{3}$ashy: Multi-Modal Material Synthesis via Hyperdiffusion}, 
    year = {2026}, 
    booktitle = {Proceedings of the 40th AAAI Conference on Artificial Intelligence}, 
    location = {Singapore}, 
    series = {AAAI'26} 
}
```