BrickFlow is the model from a NeurIPS 2026 paper by Peter Kulits and Cordelia Schmid: it takes part numbers and colors (in LDraw format) as input, uses a flow-matching denoiser to sample the position and orientation for each brick, and then snaps them into alignment based on connection relationships; the image-conditioned version (IC) can also arrange bricks according to an image. Weights are available in three pre-trained sizes—S, B, and L—with corresponding image-conditioned versions; the largest, IC-L, is 2.48 GiB. It includes scoring tools that check whether the result forms a connected structure and whether any parts intersect.
Why it matters Most image-to-3D outputs produce meshes or point clouds; this model directly generates LDraw brick assemblies, guided by connection relationships, and can verify whether the structure is connected and free of collisions.
Hugs: 2 · Hugs rising score: 2 · Downloads: 0
License: Commercial use allowed · ComfyUI: Not available · VRAM: Not specified
Reasons
- MIT license; pip package, dataset, training, and evaluation code are all public
Usage
- Run
pip install brick-flow, thenpython -m brickflow fetchto download default files, followed bypython -m brickflow fetch weights/ic-l; place your image and a corresponding.ldrparts list with the same name in the same directory (e.g.,photos/duck.pngandphotos/duck.ldr), then runpython -m brickflow sample weights/ic-l.pt photos --pt weights/pt-l.pt --samples 4 --guidance --out duck
Risks
- You must prepare the parts list yourself; the model only handles placement. The image-conditioned version depends on DINOv3, which requires login and agreement to terms.
- The model card does not specify VRAM requirements; default downloads are approximately 3.81 GiB; there is no online demo.