snskysk/event2rgb-distillation

The Shape of Events — pretrained RGB students

Pretrained weights for "The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation" (NeurIPS 2026).

• 📄 Paper: https://arxiv.org/abs/2609.30478
• 💻 Code: https://github.com/snskysk/event2rgb-distillation

We distil an event-domain teacher (DiST, trained on N-ImageNet) into an RGB ImageNet student. The resulting student suppresses its dependence on unstructured high-frequency texture and instead relies on edge-based object shape, which shows up as color invariance, shape bias, robustness to high-frequency band noise, and better linear-probe transfer.

Available weights

All files are weights-only state_dicts in safetensors format that load directly into torchvision models with strict=True.

File	Arch	Paper name	ImageNet top-1
resnet34-baseline.safetensors	ResNet-34	baseline	73.90
resnet34-ed-student.safetensors	ResNet-34	ED_Student (α = 0.2, main result)	69.70
resnet34-ed-softlabelonly.safetensors	ResNet-34	ED_SoftLabelOnly (α = 1.0)	53.50
resnet34-self-distilled.safetensors	ResNet-34	Self Distilled (control)	74.72
resnet34-gray-distilled.safetensors	ResNet-34	Gray Distilled (control)	74.60
resnet50-baseline.safetensors	ResNet-50	baseline	78.16
resnet50-ed-student.safetensors	ResNet-50	ED_Student	67.47

Start with resnet34-ed-student — it is the model behind the main claims of the paper.

Usage

import torchvision
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

path = hf_hub_download("snskysk/event2rgb-distillation",
                       "resnet34-ed-student.safetensors")

model = torchvision.models.resnet34(weights=None)
model.load_state_dict(load_file(path), strict=True)
model.eval()

Inputs use the standard ImageNet preprocessing (resize 256 → center crop 224, mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]). Outputs are 1000-way ImageNet-1k logits.

What ED_Student buys you

Relative to the RGB baseline, on four axes (Table 2 of the paper; positive is better):

Axis	baseline	ED_Student
Clean ImageNet top-1	73.9	69.7
Color — mean top-1 over 10 hue rotations	61.8	+3.6
Shape — ImageNet-Sketch top-1	23.7	+4.4
High-frequency band noise — reduction in accuracy drop at 0.95π	29.6 pp drop	+17.4
Transfer — mean linear-probe accuracy over 19 tasks	73.5	+1.6

ED_Student is the only supervision among those evaluated that improves on all four axes at once. It trades roughly 4 points of clean ImageNet accuracy for that robustness profile.

Notes

• These are the student networks only. The frozen DiST event teacher used during distillation is not redistributed here — it is the public checkpoint released by the N-ImageNet authors (82magnolia/n_imagenet), and should be obtained from that repository under its own terms.
• The original training checkpoints were PyTorch Lightning files that also contained optimizer state and the teacher weights; these files contain only the student state_dict, which is why they are about a quarter of the size.
• Remaining checkpoints from the paper (B&W Distilled, Adversarial Training, Noise Training, SIN, and the additional DiST teachers) are currently distributed from the links in the GitHub README and may be added here later.

License

MIT, for research and reproduction purposes. The datasets used upstream (ImageNet-1k, N-ImageNet) carry their own terms, which users are responsible for complying with.

Citation

All authors contributed equally.

@inproceedings{yasuki2026event2rgb,
  title     = {The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation},
  author    = {Yasuki, Shunsuke and Kihara, Soshun and Taki, Masato},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
  eprint    = {2609.30478},
  archivePrefix = {arXiv}
}