File size: 3,571 Bytes
4db71f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a2ff861
4db71f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
---
license: mit
pretty_name: BrickFlow
task_categories:
- image-to-3d
tags:
- lego
- ldraw
- brick-assembly
- 3d
- procedural
size_categories:
- 100M<n<1B
---

<div align="center">

# BrickFlow: Connectivity-Guided Brick Reconstruction

[Peter Kulits](https://kulits.github.io/) and [Cordelia Schmid](https://cordeliaschmid.github.io/)

NeurIPS 2026

</div>

The data of BrickFlow, generated with [ProcBrick](https://github.com/kulits/ProcBrick). This repo contains the
connection atlas and BrickFlow's splits. See [kulits/BrickFlow](https://huggingface.co/kulits/BrickFlow) for the model.
The code of the [`brickflow`](https://github.com/kulits/BrickFlow) and [`procbrick`](https://github.com/kulits/ProcBrick) packages is on GitHub.

## Contents

| folder | files | size | contents |
| --- | --- | --- | --- |
| `assets/` | 10 | 1.11 GiB | ProcBrick's assets: `charts`, `atlas.index`, `atlas.tiles`, `pose_dictionary`, `parts.json`, `mates`, `sampling_policy.json`, `assembly_model`, `appearance_policy.npz`, `appearance_model` |
| `train/pt/` | 3 | 19.4 GiB | pretraining: a directory of 100,000,000 trees of 100 parts (`data`, `index`, `manifest.json`) |
| `train/ic/` | 15,651 | 100.5 GiB | image-conditioned finetuning: 1,000,000 renders with their images |
| `val/procbrick/` | 207 | 1015 MiB | evaluation: 9,900 renders with their images |

A tree is built one part at a time through a compatible connector of a part already placed, so each prefix is a valid
structure. Trees are stored compressed; any tree, or its first k parts, decodes alone with `assets/` to a BrickNet
`Tree`.

Renders show a prefix of each of their trees (`assemblies/`), with print variants and materials (`appearance/`; the
appearance policy gives each material's LDraw color), part counts (`sizes.npy`), a camera (`frames.npy`), and an image:
the images are AVIF (RGBA) members of uncompressed tar shards, `shards/NNN/SSSSS.tar` holding renders of NNN parts.
Renders 2j and 2j + 1 are partners whose part counts sum to 102.

The BrickNet evaluation set (512 models from BrickNet's validation split) comes with the BrickNet dataset: request
access to BrickNet, then download `brickflow_val.tar.xz` (46 MiB) from the BrickNet dataset folder and unpack it into
`procbrick/` in BrickNet's data directory to get `val/bricknet`.

The formats are specified in the DATA.md files of the `procbrick` and `brickflow` packages.

## Download

The packages fetch folders into `procbrick/` in BrickNet's data directory (`$BRICKNET_DATA`, else the platform
user-data directory), at the same paths as here:

```bash
pip install procbrick
python -m procbrick fetch                 # assets and BrickNet's colliders
python -m procbrick fetch train/pt        # train/pt, train/ic, val/procbrick
```

```bash
pip install brick-flow
python -m brickflow fetch                 # assets, val/procbrick, and the PT-L weights
python -m brickflow fetch train/ic
```

```python
import procbrick

trees = procbrick.load_trees("train/pt", range(1000), parts=10)   # the first 10 parts of the first 1,000 trees

renders = procbrick.load_renders("val/procbrick")                 # the evaluation set
open("render.avif", "wb").write(renders.image_bytes(0))           # RGBA image of render 0
```

## License

The data are under the MIT license.

## Citation

```bibtex
@InProceedings{kulits2026brickflow,
  author    = {Kulits, Peter and Schmid, Cordelia},
  title     = {BrickFlow: Connectivity-Guided Brick Reconstruction},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}
```