comp-visual-planning/toy-example

Compositional Visual Planning: toy example

Pretrained weights and training data of the toy example (Figure 2) in Compositional Visual Planning via Inference-Time Diffusion Scaling (project page). A short-horizon diffusion model is trained on circular arc clips. At test time, three 120° arc generators are composed into a three-petal "flower" by enforcing agreement at the chunk boundaries at every denoising step.

Files

File	Description	Size
diffusion_420000.ckpt	{"model": ..., "model_ema": ...} state dicts of the DiT1d denoiser (in_dim 60, emb_dim 256, d_model 384, 12 heads, 6 blocks; x0 prediction, 500 cosine diffusion steps) after 420k gradient steps on 8-token clips	132 MB
toy.npy	training data, float64 array of shape (64, 600, 2): 64 rotated flowers, each a closed curve of 600 (x, y) points made of three noisy 120° arcs	0.6 MB

Usage

The toy example code downloads both files when they are missing locally. To download them by hand:

pip install -U huggingface_hub
hf download comp-visual-planning/toy-example diffusion_420000.ckpt --local-dir ckpts
hf download comp-visual-planning/toy-example toy.npy --local-dir data

or from python:

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("comp-visual-planning/toy-example", "diffusion_420000.ckpt", local_dir="ckpts")
data = hf_hub_download("comp-visual-planning/toy-example", "toy.npy", local_dir="data")

Then, from the root of the toy example repository, run python eval.py method=cvp (ours) or python eval.py method=gsc (DiffCollage / Generative Skill Chaining baseline).

Citation

@misc{zhang2026compositionalvisualplanninginferencetime,
      title={Compositional Visual Planning via Inference-Time Diffusion Scaling},
      author={Yixin Zhang and Yunhao Luo and Utkarsh Aashu Mishra and Woo Chul Shin and Yongxin Chen and Danfei Xu},
      year={2026},
      eprint={2603.02646},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2603.02646},
}