Xinhao Lou1,∗ · Dongye Xie2,∗ · Peng Sun1,4,∗ · Chenyi Zhuang3 · Kaicheng Yu1,† · Tao Lin1,†
1 Westlake University · 2 University of Cambridge
3 Independent Researcher · 4 Zhejiang University
∗ Equal contribution. † Corresponding authors.
Paper · Results · ImageNet guide · TODO
NextGAN learns how to update a generator by supervising the input gradients of a scalar critic.
NextGAN training pipeline (Figure 2 in the paper).
Paper results from Table 1 and our GitHub reproduction results are shown together below. All use one network evaluation and no classifier-free guidance.
| Source | Method | Generator | FID ↓ | IS ↑ |
|---|---|---|---|---|
| Paper | NextGAN | JiT-B | 2.87 | 207.1 |
| Paper | NextGAN | DiT-XL/1 + VA-VAE | 1.54 | 243.5 |
| Paper | NextGAN, adversarial form | DiT-XL/1 + VA-VAE | 1.36 | 241.7 |
| Github | NextGAN | JiT-B | 2.75 | 213.0 |
| Github | NextGAN | DiT-XL/1 + VA-VAE | 1.27 | 258.9 |
The GitHub results use EMA weights, 50K images, seed 1, eight GPU ranks, batch size 256 per GPU and BF16, with one training seed per model.
Code and recipe status. The current ImageNet implementation provides the field-regression objective with unit weight. The paper's experimental recipe uses a different regression scale in Appendix A. The included DiT configuration is described in the ImageNet guide. Exact Appendix A configuration alignment and the adversarial-form implementation remain release follow-ups.
NextGAN compares reference and generated images through frozen representation models. Each real–generated feature pair defines a displacement target. A scalar critic learns a potential whose input gradient matches that displacement at randomly interpolated features. The generator then increases the potential at its generated endpoints.
For images, the training representations are MAE, SigLIP2 and ResNet18. Their weights stay frozen while gradients pass through them to the generator. Inference uses only the generator and, for latent models, the decoder.
Use Linux, CUDA, Bash and Python 3.10–3.12. The included DiT training recipe uses eight GPUs; its evaluation protocol also uses eight GPU ranks.
git clone https://github.com/LINs-lab/NextGAN.git
cd NextGAN/imgnetPrepare the assets listed in ASSETS.md, then edit the paths in
configs/reproduced/nextgan_dit_vavae_256.yaml:
- ImageNet VA-VAE latent shards and pretrained DiT-XL/1 initialization.
- The matching converted VA-VAE model and channel statistics.
- MAE, SigLIP2 and ResNet18 weights for training.
- Inception weights and ImageNet reference statistics for FID evaluation.
Public NextGAN checkpoint downloads, a versioned asset manifest and a standalone sampling example are pending. The commands below use assets supplied locally.
export DM_MODEL_ZOO=/path/to/model_zoo
export DM_DATA_ZOO=/path/to/data_zoo
python3 reproduce.py \
--config configs/reproduced/nextgan_dit_vavae_256.yaml \
--stage check \
--runs-dir ./reproductions/nextgan-dit-256The reproduction entry creates a virtual environment, installs the pinned
dependencies and checks the configured assets. Use an explicit --config to
select the NextGAN recipe.
After the asset check succeeds:
EVAL_NPROC_PER_NODE=8 python3 reproduce.py \
--config configs/reproduced/nextgan_dit_vavae_256.yaml \
--skip-install \
--runs-dir ./reproductions/nextgan-dit-256This trains for 100K generator updates and evaluates the final EMA checkpoint using 50K images. It saves resolved configurations, source and asset hashes, dependency versions, logs and metrics in the run directory. The ImageNet guide covers individual checkpoints, checkpoint sweeps, resuming training and alternative NextGAN configurations.
- Release the Wan2.1 code.
- Release NextGAN2.0.
- Release the MiniMax-H3 code.
- Release the model checkpoints.
@article{lou2026nextgan,
title={NextGAN: Towards Simple and Scalable GAN Training for Image and Video Generation},
author={Lou, Xinhao and Xie, Dongye and Sun, Peng and Zhuang, Chenyi and Yu, Kaicheng and Lin, Tao},
journal={arXiv preprint arXiv:2611.06270},
year={2026},
url={https://arxiv.org/abs/2611.06270}
}This implementation builds on work including UCGM, JiT and FD-Loss, alongside the pretrained representation models used for training and evaluation. Original project code is licensed under Apache 2.0. Adapted third-party components retain their upstream terms; see THIRD_PARTY_NOTICES.md for attribution and component-specific licenses. Model weights and datasets have their own terms.
