HierAmp: Coarse-to-Fine Autoregressive Amplification for Generative Dataset Distillation

August 5, 2026 · View on GitHub

Official implementation of HierAmp (CVPR 2026). Object semantics are inherently hierarchical; HierAmp matches this hierarchy with the next-scale prediction of Visual Autoregressive Models (VAR). At each generation scale, a learnable class token identifies salient regions, and its induced activation map guides an additive attention-logit amplification that steers synthesis toward discriminative structures — with minimal inference overhead and no external segmenter.

Method

At each scale $n$ , a class token attends to the tokens of its own scale (scale-restricted attention). Its multi-head-averaged attention gives a semantic saliency map

$m_n = \frac{1}{H}\sum_h \alpha^{(h)}_{n,\mathrm{cls}},\qquad a_n = \mathrm{Top}\text{-}\rho_n(m_n)\in\{0,1\}^{L_n},$

and generation is steered by an additive logit bias applied before softmax:

$\tilde{L}^{(h)}_n = L^{(h)}_n + \beta_n\,\mathbf{1}\,a_n^{\top},\qquad \tilde{\alpha}^{(h)}_n=\mathrm{softmax}(\tilde{L}^{(h)}_n).$

At coarse scales this diversifies object layouts; at fine scales it concentrates token usage on object details. The final configuration uses $\beta_n = 5$ and $\rho_n = 50\%$ at all scales $n \in \{1,\dots,9\}$ (scale 0 contains a single patch and is excluded), applied at transformer block 14 of the depth-16 VAR.

Setup

pip install -r requirements.txt

Verified with PyTorch 2.5.1 + CUDA 12.4 (any torch ≥ 2.1 should work). flash-attn and xformers are optional accelerators, not requirements.

Download the pre-trained VAR-d16 and its VQVAE (both from FoundationVision/var):

wget https://huggingface.co/FoundationVision/var/resolve/main/vae_ch160v4096z32.pth
wget https://huggingface.co/FoundationVision/var/resolve/main/var_d16.pth

1. Fine-tune with class tokens (5 epochs)

The pre-trained VAR is fine-tuned for 5 epochs with a class token injected at every scale (0–9). Each class token is supervised by a lightweight shared classifier: the total loss is the VAR next-scale token cross-entropy plus $L_{cls} = \frac{1}{N}\sum_n -\log p_n(c_n^e)$ (Eq. 6, weight --cls_lw, default 1.0).

bash scripts/finetune.sh /path/to/imagenet 8
# → local_output/ar-ckpt-last.pth

The class token attends only to tokens of its own scale; its keys/values are never exposed to later scales (nor cached at inference), so the backbone's autoregressive structure is unchanged. $c_n^e$ is the last block's attention output at the class-token position of scale $n$ ; class-token positions are dropped before the token logits, so the token loss itself is untouched.

Notes:

Interrupted runs auto-resume: rerunning the same command picks up local_output/ar-ckpt-last.pth. Clear local_output/ (or change --local_out_dir_path) before starting a run with a different configuration.
--cls_scales= (empty) trains a plain VAR without class tokens, e.g. for baseline comparisons.
Progressive training (--pg > 0) is not supported together with class tokens.

2. Generate the distilled dataset

Paper configuration ( $\beta=5$ , $\rho=50\%$ , scales 1–9, block 14):

python sample.py --var_ckpt local_output/ar-ckpt-last.pth --save_dir output/in1k_ipc10 --ipc 10

Useful flags:

flag	default	meaning
`--ipc`	10	images per class
`--beta`	`5`	amplification strength; 1 value or 3 values for the coarse/mid/fine stage-aware schedule $\beta_{1:3},\beta_{4:6},\beta_{7:9}$
`--top_ratio`	`0.5`	$\rho$ ; 1 or 3 values, same convention
`--scales`	`1 … 9`	scales to amplify
`--layers`	`14`	transformer block(s) receiving the bias
`--no_amplify`	—	plain VAR baseline
`--classes` / `--start --end`	all 1000	class subset (e.g. ImageNet-100 / ImageNet-Woof index lists)
`--cfg / --top_k / --top_p / --seed`	4.0 / 900 / 0.95 / 0	sampling hyperparameters

Images are generated at 256×256 and saved at 224×224 for downstream training. See scripts/sample.sh for the full IPC sweep.

3. Evaluate

Train a student network (e.g. ResNet-18) on the generated images following standard dataset-distillation evaluation protocols with soft labels at 224×224. Please using the evaluation code in D3HR.

Notes for reproduction

Amplification requires the explicit-attention path; it is automatically used whenever class tokens are active, even if flash-attn/xformers are installed (they still accelerate all other layers/scales at training time).
--cls_scales at sampling must match the fine-tuned checkpoint (default: 0–9).
Generation is deterministic given --seed; each additional batch chunk of a class uses seed + chunk.

TODO

We will release the distilled dataset soon.

Citation

@inproceedings{zhao2026hieramp,
  title={HierAmp: Coarse-to-Fine Autoregressive Amplification for Generative Dataset Distillation},
  author={Zhao, Lin and Jiang, Xinru and Xiao, Xi and Fan, Qihui and Lu, Lei and Wang, Yanzhi and Lin, Xue and Camps, Octavia and Zhao, Pu and Gu, Jianyang},
  booktitle={CVPR},
  year={2026}
}

Acknowledgments

This codebase is built on VAR. We thank the authors for open-sourcing it.