Fast and Useful: Adapting Distilled Generators for Visual Learning

Patryk Bartkowiak1,4, Dominik L. Michels2,4, Sören Pirk3,4, Wojtek Palubicki1,4
1Adam Mickiewicz University, 2KAUST, 3Kiel University, 4GreenMatterAI
arXiv Code
Two sampling paths from the same Gaussian noise for the class red fox. Top: the released FLUX.2-klein student takes four steps, and its four fox images look alike; classifiers trained on its images reach 17.1% real-image accuracy. Bottom: the first-step adapter makes one jump, the unchanged student finishes in one more step, and the foxes differ more in pose, framing and background; accuracy rises to 43.8%.

First-step adaptation restores variety and improves learning from synthetic images. The adapter learns to jump to an intermediate state of the unguided base, replacing the released FLUX.2-klein student's first three predictions. The unchanged student then finishes generation, requiring two denoiser evaluations instead of four. Fox images show one selected class with matched prompts and noise seeds. ResNet-50 top-1 accuracy on real ImageNet-100 rises from 17.1 ± 0.4% to 43.8 ± 0.4% after training on 500 generated images per class (mean ± SD; N = 3 classifier seeds for the released student, 3 adapter runs × 3 classifier seeds for ours).

Abstract

Distilled text-to-image models generate images quickly, but classifiers trained on those images can perform substantially worse than classifiers trained on images from their slower base models. We adapt released distilled generators to produce better training data while retaining fast generation. We fit a low-rank update to the student's first prediction so that it reaches an intermediate state produced by its unguided base, then disable the update and let the unchanged student finish generation. The base supplies training targets and is absent at inference. Guided targets and shorter jumps give classifiers lower accuracy in our ablations, showing why the state used for supervision matters. Across three generator families, the adapted images train better classifiers than images from the released students. With class-name templates at full ImageNet-1k scale, ResNet-50 top-1 rises from 8.4% to 34.2% for FLUX.2-klein and from 12.7% to 32.1% for SD-Turbo, using two and three denoiser evaluations per image, respectively. Fixed adapters also improve results on additional classes, fine-grained datasets, and other learner architectures and training recipes. Template-based adaptation remains useful against richer-prompt alternatives, while the benefit of combining adaptation with LLM generation prompts depends on the generator. Adapting a single generation step makes fast generators substantially more useful for training visual models.

One adapted prediction, then the released student

Guidance costs training value mostly in the first, high-noise steps, and handing a distilled model's early steps to its unguided base helps more than handing over late ones. We correct only the first prediction.

  1. Targets from the unguided base. For each prompt and starting noise, the base model runs without classifier-free guidance on a fine schedule, and we save its state at a chosen noise level σa.
  2. A low-rank update to the first prediction. With the released weights frozen, a rank-32 LoRA learns to jump from the noise straight to that saved state in one prediction.
  3. The released student finishes. At generation the adapter is active for the first prediction only; the unchanged student completes its own schedule. The base is not loaded.
Left: noise-level timelines for FLUX.2-klein. The unguided base takes 34 fine steps to reach sigma_a = 0.72 (used for training only); the adapted first step regresses onto that state in one jump and the student finishes in one step; the released student takes four steps. Right: images for the prompt hare from the same noise per column, completed from the teacher target, by the adapter with 2 NFE, and by the released student with 4 NFE.
One adapted prediction followed by unchanged student sampling. Left: klein's released four-step student (orange), its unguided base on a fine grid (blue), and the adapted sampler (violet). We train the first prediction to reach the base state at σa; the unchanged student then finishes in one step, giving two denoiser evaluations without the base at inference. Right: matching noise and hare prompts, completed from the teacher target, by the adapter, or by the released student. Images are selected to illustrate the sampling paths.
ℒ(ΔW) = 𝔼x1, c ‖ x1 + (σa − 1) vθ+ΔW(x1, 1, c) − x⋆σa(x1, c) ‖2

Flow-matching form; SD-Turbo uses the matching Euler update in its scheduler's timestep. One-time preparation (targets and fitting): 5.7 H100-hours for klein, 0.9 for SD-Turbo.

Released studentBase teacherJump endpointNFE, adaptedNFE, released
FLUX.2-kleinklein baseσ = 0.717524
SD-TurboSD 2.1t = 49934
Z-Image-TurboZ-Imageσ = 0.857179

Adapted images train better classifiers

A ResNet-50 trained from scratch on generated images only, tested on real images. Across three generator families the adapter beats the released student and the base-first hybrid, which needs the base at generation. The much slower unguided base is shown for reference.

Real-image top-1 accuracy by sampler
Pick a model and a dataset; hover a bar for the numbers.

Accuracy without the slow base

The adapted samplers are as fast as the released students or faster, because the one adapted jump replaces several student steps. The unguided bases buy their extra accuracy with more generation time per image.

ImageNet-100 top-1 against generation time
Seconds per image on one H100 (log scale), including prompt encoding and decoding.

Fixed adapters transfer to fine-grained datasets

The adapters trained on ImageNet-100 class names are applied unchanged to Food-101, Stanford Cars and Oxford-IIIT Pets. They beat SDA, a public diversity LoRA for Z-Image-Turbo, and improve DP-DMD's released student.

Fine-grained top-1 accuracy, 500 generated images per class
Pick a model and a dataset; hover a bar for the numbers.

The gains hold across learners

Learner, trained from scratchFLUX.2-kleinSD-Turbo
releasedadaptedreleasedadapted
ResNet-50 + RandAugment22.0 ± 0.346.7 ± 0.423.4 ± 0.748.0 ± 0.6
Recipe of Adamkiewicz et al.17.9 ± 1.039.0 ± 0.321.7 ± 1.238.6 ± 0.5
ConvNeXt-T19.9 ± 0.239.0 ± 0.220.7 ± 0.541.8 ± 1.4
ViT-S/1615.6 ± 0.126.9 ± 0.513.8 ± 0.426.8 ± 0.3

ImageNet-100 top-1 (%), mean ± SD over three classifier seeds, one adapter run, the same generated sets for every learner.

Acknowledgments

EOSC-ARENA Poznańskie Centrum Superkomputerowo-Sieciowe (PCSS)

This work was supported by the European Union through the EOSC-ARENA project (AI Research Enhancement through Networked Agents), Grant Agreement No. 101292597.

The authors thank Poznańskie Centrum Superkomputerowo-Sieciowe (PCSS) for providing the computational resources used in the experiments.

BibTeX

@misc{bartkowiak2026fastusefuladaptingdistilled,
      title={Fast and Useful: Adapting Distilled Generators for Visual Learning},
      author={Patryk Bartkowiak and Dominik Michels and Sören Pirk and Wojciech Pałubicki},
      year={2026},
      eprint={2610.32512},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.32512},
}