# Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning
This repository contains four experiments on suppressing a specified class, style, mode, or spatial region in diffusion sampling.
Each experiment reports an avoidance metric and a distance or quality measure for the remaining target distribution.
The implementation and reproduction commands are documented in the README of each experiment.
## Experiments
### [`exp_checkerboard`](exp_checkerboard/)
Two-dimensional toy distributions provide the main visual demonstration.
The experiments remove modes from a ring mixture or a tile from a checkerboard while retaining the shape of the kept distribution.
The additional metrics are `leak` and the empirical 2-Wasserstein distance `W2`.
### [`exp_gmm`](exp_gmm/)
Three-dimensional twelve-component ring GMMs provide a controlled quantitative evaluation.
The score-based setting uses exact scores, and the reference-based setting uses samples from the avoided components.
The reported metrics are `post_mean`, `leak`, and the empirical 2-Wasserstein distance.
### [`exp_cifar10`](exp_cifar10/)
The CIFAR-10 experiment suppresses one class at a time or the vehicle group consisting of Plane, Car, Ship, and Truck.
It reports the avoided-class rate together with FID, KL, and mean pairwise LPIPS distance.
### [`exp_stable_diffusion`](exp_stable_diffusion/)
The Stable Diffusion v1.4 experiment generates the prompt "a cafe terrace at night under a starry sky" with less Van Gogh style.
It reports the style detection score, FID, KID, CSD-FD, and ESS/M for the weighted estimates.
## Citation
Hirohane Takagi and Masashi Sugiyama. Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning. The 40th Annual Conference on Neural Information Processing Systems (NeurIPS2026), In Advances in Neural Information Processing Systems, **:****--****, 2026.
```bibtex
to be added
```
