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.
Figures & tables
Figure 2: 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.
Student
Base teacher
Jump endpoint
NFE
FLUX.2-klein
klein base
σ=0.7175
2
SD-Turbo
SD 2.1
t=499
3
Z-Image-Turbo
Z-Image
σ=0.8571
7
Table 1: Teacher targets and adapted sampling schedules. The first prediction jumps to the target; the unchanged student completes generation. NFE counts all denoiser evaluations in the adapted sampler. Flow-model targets use noise level σ ; SD-Turbo uses its scheduler timestep t .
Figure 3: First-step adaptation improves the training value of released fast generators. ResNet-50 top-1 on real ImageNet-100 after training on 500 generated images/class. Adapters exceed released students and base-first sampling (a–c), the tested SDA adapter (c), and the released DP-DMD student when applied on top of it (d). Blue: unguided bases. Error bars: mean ± SD under the reporting convention in Sec. 4.1 ; DP-DMD plus ours uses one adapter run. Z-Image’s released reference uses SDA’s eight-step schedule; Tab. S6 also reports the nine-step student.
Sampler
200 classes
ImageNet-1k
FLUX.2-klein
Released student
19.1 ± 0.0
8.4 ± 0.1
Base-first
43.7 ± 0.1
–
Ours
48.6 ± 0.3
34.2 ± 0.1
Unguided base, 20 steps
–
37.2 ± 0.1
SD-Turbo
Table 2: The gains extend to additional classes and full ImageNet scale. ResNet-50 top-1 (%), mean ± SD over three classifier seeds, one adapter run. The 200-class protocol informed development (500 images/class); ImageNet-1k uses fixed samplers at 1,281 images/class. Bold: best fast sampler. Complete primary results, including ImageNet-100, COCO-target adapters, and costs, are in Tab. S6 .
Figure 4: Adaptation improves classifier accuracy while retaining fast generation. ImageNet-100 top-1 against measured H100 seconds/image (log scale); vertical bars show SD. Unguided-base points use 10, 20, and 50 steps for klein, 10 and 30 for SD 2.1, and 50 for Z-Image. The adapted samplers require only the student at inference. Preparation is excluded here and included in Tab. S2 ; additional operating points appear in Fig. S2 .
variant
FLUX.2-klein
SD-Turbo
ours: unguided, step 1, rank 32
43.8 ± 0.4
43.7 ± 0.6
same, finer teacher grid
43.2 ± 0.7
–
target at guidance 2
31.5 ± 0.7
41.2 ± 0.9
target at default guidance
20.1 ± 0.4
30.4 ± 0.4
adapter in every step
41.4 ± 0.3
38.8 ± 0.8
every weight fine-tuned
43.2 ± 1.0
43.4 ± 0.3
Table 3: Target choice matters more than parameter count. ImageNet-100 ResNet-50 top-1 (%). Ours and SD-Turbo rank 128 pool three adapter runs; others use one. Klein’s guided-target and endpoint variants use the finer teacher grid (second row); the deeper endpoint is 0.511. Guided targets and short jumps forfeit much of the gain; full fine-tuning and higher rank add little. Bold: best per column. FSG, MoNO (published basis: four images), and skipping step 1 require no training.
Sampler
NFE
Food-101
Cars
Pets
FLUX.2-klein
Released
4
10.5 ± 0.3
18.9 ± 0.4
20.5 ± 1.0
Base-first
4
25.9 ± 0.1
38.8 ± 0.3
45.3 ± 0.5
Ours
2
30.9 ± 0.2
36.2 ± 0.2
46.2 ± 0.8
SD-Turbo
Released
4
10.0 ± 0.3
46.9 ± 0.4
19.5 ± 1.4
Table 4: Adapters transfer unchanged to fine-grained datasets. ResNet-50 top-1 (%), 500 generated images/class; mean ± SD over three classifier seeds. Our adapters use one training run. NFE: denoiser evaluations. Bold: best shown per family. Additional references, gaps, and runtimes: Tabs. S11 and S10 .
learner
FLUX.2-klein
SD-Turbo
Original
Adapted
Original
Adapted
R50 + RA
22.0 ± 0.3
46.7 ± 0.4
23.4 ± 0.7
48.0 ± 0.6
Prior recipe
17.9 ± 1.0
39.0 ± 0.3
21.7 ± 1.2
38.6 ± 0.5
ConvNeXt-T
19.9 ± 0.2
39.0 ± 0.2
20.7 ± 0.5
41.8 ± 1.4
ViT-S/16
15.6 ± 0.1
26.9 ± 0.5
13.8 ± 0.4
26.8 ± 0.3
Table 5: Gains persist across learner architectures and recipes. ImageNet-100 top-1 (%), mean ± sd over three seeds, one adapter run. Learners train from scratch on the same sets. R50: ResNet-50; RA: RandAugment; prior recipe: [ 1 ] . Bold: adapted versus released generator. Full references: Tab. S13 .
Comparison
FLUX.2-klein
SD-Turbo
ImageNet-100: adapted − released
Templates / templates
+26.9 [+23.5, +30.4]
+25.0 [+21.9, +28.1]
Captions / templates
+6.4 [+4.3, +8.6]
+20.7 [+17.7, +23.8]
LLM / templates
+2.1 [+0.2, +4.0]
+16.7 [+14.1, +19.4]
LLM / LLM
+5.4 [+3.4, +7.3]
+15.6 [+12.9, +18.5]
ImageNet-1k: vs. released with LLM prompts
Table 6: The benefit depends on fitting prompts and evaluation regime. Top: ImageNet-100 gains (points [95% class-bootstrap CI]), generation/fitting prompts, 500 images/class. LLM rows pool three datasets; others use one, each with three classifier seeds. Bottom: ImageNet-1k, 100 images/class, one dataset and three classifier seeds, versus released students using LLM prompts. Template-based: template fitting/generation. LLM-prompt: LLM generation, fitting selected on ImageNet-100 (LLM for klein, templates for SD-Turbo). One training run per adapter. Bold: interval above zero.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
(a) Guidance only above or only below a noise level.
(b) Hand-offs between the distilled model and the base.