---
license: mit
tags:
- image-generation
- one-step-generation
datasets:
- ILSVRC/imagenet-1k
---

# Generative Modeling via Drifting: Pretrained Weights

JAX weights for the ImageNet 256x256 models of [Generative Modeling via Drifting](https://arxiv.org/abs/2602.04770) (NeurIPS 2026, oral). Code, usage, and evaluation instructions: [lambertae/drifting](https://github.com/lambertae/drifting). Project page: [lambertae.github.io/projects/drifting](https://lambertae.github.io/projects/drifting/).

## Files

| Path | Description |
|---|---|
| models/gen/jax/latent_L_sota | Latent-space generator, L/2 (FID 1.54, 1-NFE) |
| models/gen/jax/latent_B_sota | Latent-space generator, B/2 (FID 1.75, 1-NFE) |
| models/gen/jax/pixel_L_sota | Pixel-space generator, L/16 (FID 1.61, 1-NFE) |
| models/gen/jax/pixel_B_sota | Pixel-space generator, B/16 (FID 1.76, 1-NFE) |
| models/gen/jax/ablation | Ablation-setting generator, B/2 |
| models/mae/jax/mae_latent_640 | Latent-space MAE-640 feature encoder |
| models/mae/jax/mae_pixel_640 | Pixel-space MAE-640 feature encoder |
| models/mae/jax/mae_latent_256 | Latent-space MAE-256 feature encoder (ablations) |

The code downloads these weights automatically.

## License

MIT, same as the [code repository](https://github.com/lambertae/drifting/blob/main/LICENSE).

## Citation

Please cite the paper; the BibTeX entry is in the [code repository](https://github.com/lambertae/drifting#citation).