Fast and Useful: Adapting Distilled Generators for Visual Learning

1. Introduction

• Synthetic images can expand training sets without new image collection or annotation, but their value depends on helping classifiers recognize real images.Text prompts provide class control and allow datasets at different scales, for synthetic-only training or supplementation of limited real data.Source: §1 · p. 1
• Distillation speeds dataset generation by reducing denoising predictions, yet classifiers trained on released distilled generators can underperform those trained on their slower full-step counterparts.The stated aim is to retain the speed advantage while improving the student images’ usefulness for learning.Source: §1 · p. 1
• At full ImageNet-1k scale with template prompts, adaptation beats fixed base-first samplers by +3.5 points on klein and +5.4 on SD-Turbo.Matched ablations test target guidance, jump endpoint, and update placement.Source: §1 · p. 2
• The approach adapts three distilled generator families and compares them with released students and base-first hybrids under the same classifier-training procedure.The evaluation spans datasets, learners, prompt regimes, and preparation and generation costs.Source: §1 · p. 2
• Fixed adapters improve full-scale ImageNet and fine-grained recognition, while prompt, learner, and real-data comparisons delimit the tested settings where adaptation helps.The paper also tests adaptation alongside diversity methods and evaluates additional classifier architectures and training recipes.Source: §1 · pp. 1–2

Synthetic data for visual learning

• Prior synthetic-data work uses text-to-image generation for classifier training and synthetic-data supplementation, while lower guidance and richer descriptions improve supervised learning and StableRep uses same-caption positives for self-supervised learning.Adamkiewicz et al. [1] identify the complementary issue that newer and distilled generators can yield weaker classifiers than older models.Source: §2 · p. 2

Diffusion acceleration and distillation

• Distillation reduces model evaluations for faster generation, but visual quality does not guarantee equal usefulness for classifier training.The paper adapts released students rather than training another distilled generator from its base.Source: §2 · p. 2

Diversity preservation and adaptation

• Diversity methods include base-first sampling, teacher-supervised first-step training, trajectory-preserving objectives, and teacher-guided directional alignment.Base-first sampling uses the base at inference, whereas this paper adapts a released image student and leaves its completion steps unchanged.Source: §2 · p. 2
• The study tests how unguided teacher targets and first-prediction adaptation affect classifier accuracy while the released student’s completion steps remain fixed.Source: §2 · p. 2

3.1. Problem and approach

• The first student prediction is adapted to approximate an intermediate state from the corresponding full-step base, after which the released student completes generation.The base prepares targets but is not loaded for inference.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2. Constructing the teacher target

• Unguided targets retain the base’s text-conditioned prediction without classifier-free guidance amplification, corresponding to guidance weight one.Comparisons hold prompts, image seeds, and classifier training fixed; reducing guidance improves classifier accuracy in tested configurations.Source: §3.2 · p. 3

3.3. Correcting the early generation steps

• Replacing early student predictions with an unguided base jump improves accuracy more than replacing later predictions, motivating a jump that learns the teacher’s early computation.The endpoint is chosen from the released student’s noise levels so its unchanged schedule can continue; later evaluations use previously fixed adapters and samplers.Source: §3.3 · p. 3, Table 1 · p. 3

3.4. Training the adapter

• The adapter freezes released weights and fits low-rank updates by minimizing squared error between the adapted latent and the teacher target.The loss uses flow-matching coordinates; SD-Turbo uses the corresponding Euler update, and primary adapters use rank-32 updates trained on 16,000 noise/target pairs for 3,000 steps.Source: §3.4 · p. 3

3.5. Sampling with the adapted student

• At inference, the adapter is active only for the first prediction, then the unchanged student completes the schedule; a jump can replace several student steps with one.Generation loads only the student and adapter; target preparation and fitting add one-time costs reported separately from per-image generation time.Source: §3.5 · p. 3, Figure 2 · p. 4

4.1. Experimental setup

• The primary comparison uses matched prompts and seeds and a fixed classifier-training procedure to compare released students, bases, base-first samplers, and adapters.The study also uses fixed adapters for full-scale ImageNet and fine-grained dataset evaluations.Source: §4.1 · p. 4

4.2. Classifier accuracy from adapted images

• On ImageNet-100, adaptation raises classifier accuracy from 17.1% to 43.8% on klein and from 18.3% to 43.7% on SD-Turbo; Z-Image-Turbo reaches 35.1%.Adapters outperform base-first samplers; the unguided bases retain higher accuracy in these comparisons.Source: §4.2 · p. 4, Figure 3 · p. 5
• On ImageNet-1k, ResNet-50 top-1 rises from 8.4% to 34.2% for klein and from 12.7% to 32.1% for SD-Turbo.At 1,281 images per class, adapters exceed base-first sampling by +3.5 points for klein and +5.4 for SD-Turbo.Source: §4.2 · pp. 4–5, Table 2 · p. 5
• At full ImageNet scale, gains remain positive on 900 classes absent from adapter training, and larger datasets benefit the adapted generator more than the unadapted klein generator.The same classifier epoch count means larger datasets receive more optimization updates.Source: §4.2 · p. 5

4.3. Generation and preparation costs

• Adapted sampling matches or improves on released-student generation costs while increasing classifier accuracy; preparation costs make adaptation more expensive at 50,000 images in the measured template settings.Adaptation becomes cheaper overall than base-first after approximately 399,000 images for klein and 807,000 for SD-Turbo.Source: §4.3 · pp. 5–6, Figure 4 · p. 6

4.4. Evidence for the design choices

• Early-step comparisons favor an unguided intermediate target and an update used only for the initial prediction over guided targets, shorter jumps, deeper endpoints, or keeping the update active.On klein, the unguided target reaches 43.2% versus 20.1% for the guidance-four target.Source: §4.4 · p. 6, Table 3 · p. 6

4.5. Generalization across datasets and learners

• ImageNet-trained adapters transfer unchanged to Food-101, Cars, and Pets, and improvements persist across learner architectures and training recipes, though gain sizes vary.Klein’s adapter trails base-first on Cars.Source: §4.5 · pp. 6–7, Table 4 · p. 7, Table 5 · p. 7

4.6. Prompt regimes and real-data availability

• Template-based adaptation remains useful against richer-prompt alternatives, but the benefit of fitting and generating with LLM prompts varies by generator and evaluation regime.At ImageNet-1k with 100 images per class, template-based procedures exceed released students using image-free LLM descriptions by +2.7 points for klein and +7.1 for SD-Turbo.Source: §4.6 · p. 7, Table 6 · p. 7

• Unguided intermediate targets outperform guided targets and the original first endpoint, while further adapting DP-DMD’s student improves accuracy but leaves a gap to its unguided base.The evaluated comparisons do not isolate the source of that remaining gap.Source: §5 · pp. 7–8
• The useful prompt and sampler choice depends on generator and dataset: templates help both models in specified comparisons, while base-first exceeds klein’s adapter on Cars.Real-image captions retain gains but require access to those images.Source: §5 · p. 8
• Adaptation improves accuracy at comparable or lower generation time without the base at inference, but preparation can make base-first cheaper at 50,000 images; full-scale costs differ by family.At full ImageNet scale, klein’s estimated preparation-plus-generation cost is 41.4 H100-hours versus 54.5 for base-first; SD-Turbo costs are approximately equal.Source: §5 · p. 8
• Real-data recall and CLIP Fréchet distance track classifier accuracy across studied configurations, but shared settings limit independence and the associations establish neither causality nor reliable classwise prediction.Training classifiers remains the direct evaluation of usefulness.Source: §5 · p. 8

6. Limitations & Future Work

• No rule is established for choosing a useful endpoint on a new generator; the endpoint is selected empirically using downstream classifier results.Applying the method also requires a compatible base model and an intermediate state the student can complete.Source: §6 · p. 8
• Full-scale ImageNet-1k and fine-grained evaluations use fixed adapters, but classes absent from adapter fitting may still have appeared during development.ImageNet-100, the 200-class protocol, and the small ImageNet-1k pilot informed method choices; more independent dataset regenerations would better characterize uncertainty.Source: §6 · p. 8
• Evaluation covers classification and feature transfer, not detection or segmentation, and aggregate gains do not establish balanced performance across demographic groups or rare conditions.The rich-prompt ImageNet-1k evaluation uses 100 images per class and one generated dataset per sampler, so it does not establish outcomes at full synthetic-dataset scale.Source: §6 · p. 8
• Adaptation inherits the biases and coverage limitations of the released generators.Separate evaluation is required for balanced performance across demographic groups and rare visual conditions.Source: §6 · p. 8

7. Conclusion

• Fixed klein and SD-Turbo adapters benefit full-scale ImageNet training and fine-grained recognition, with improvements extending across three generator families.The approach uses a low-rank update to reach an unguided intermediate base-model state, then lets the unchanged student finish generation without the base at inference.Source: §7 · p. 8
• Controlled comparisons establish the importance of target guidance, endpoint, and update placement.Source: §7 · p. 8
• The supplement covers protocols, target design, baseline comparisons, generalization, prompt and real-data regimes, and supporting diagnostics.Its figures and tables provide additional experiments, numerical results, reference accuracies, and diagnostic sensitivity checks; the main comparison plots are not reproduced.Source: §7 · p. 9

A. Reproduction and evaluation protocols

• The protocol section specifies dataset roles, classifier training, generator settings, adapter fitting, and cost accounting.Later fixed-setting evaluations test development choices; classes unseen during adapter training may still have participated in method development.Source: A. Reproduction and evaluation protocols · p. 9, Table S1 · p. 9

A.1. Datasets, classifiers, and repetition

• The default ResNet-50 is trained from scratch for 100 epochs, while the alternative recipe uses Adam, 80 epochs, CutMix, and held-out generated-image checkpoint selection.The protocol also describes ViT-S/16 training, evaluation splits and real-data references, feature transfer, and repetition/reporting procedures.Source: §A.1 · pp. 9–10

A.2. Sampler development and adapter selection

• ImageNet-100 sampler selection and reported accuracy use the same 5,000 real validation images; the 200-class protocol and ImageNet-1k pilot also informed development.The table distinguishes these development protocols from later evaluations using previously selected adapters and samplers.Source: Table S1 · p. 9, §A.2 · p. 10
• Endpoint and rank choices are selected using downstream accuracy criteria, while adapter optimization uses held-out noise-target pairs to choose between full-update learning rates.For klein and SD-Turbo, no tested alternative passed the stated replacement criterion; standard low-rank adapters use the final 3,000-step checkpoint.Source: §A.2 · p. 10

A.3. Feature metrics and cost measurement

• Feature metrics use normalized DINOv3 and CLIP embeddings; recall measures the fraction of real features inside generated-feature balls.Precision, recall, density, and coverage use within-class Euclidean distances and fifth-nearest-neighbor radii; FID uses Inception-v3 features.Source: §A.3 · p. 10
• Cost estimates count adapter preparation once and use measured per-image generation times, assuming constant throughput.Against base-first, estimated crossovers are approximately 399,000 images for klein and 807,000 for SD-Turbo with template prompts; estimates exclude sampler development and classifier evaluation.Source: §A.3 · pp. 10–11, Table S2 · p. 11, Table S3 · p. 12

A.4. Generation and adapter implementation

• The adapter uses rank-32 low-rank updates trained on pairs whose targets are unguided-base states, with training seeds held out from evaluation.Target construction runs the base sampler on a fine grid containing the target noise level.Source: §A.4 · p. 12
• For SD-Turbo, the adapted call jumps from timestep 999 to 499, then the unchanged student finishes at 249 and zero in three U-Net calls.The target state is produced by unguided SD 2.1 using the same initial noise, text encoder, and VAE.Source: §A.4 · p. 13

A.5. Data provenance and impact

• The evaluation uses existing recognition datasets, released generators, and COCO captions, while noting that generated data can inherit model biases and omit rare conditions.Aggregate accuracy or recall does not establish fairness or deployment safety; downstream uses require validation on intended populations and conditions.Source: §A.5 · p. 13

B.1. Early guidance and base/student handoffs

• 17.5 of the 21.4 points lost with full-trajectory guidance are lost by guiding only klein’s first 10 of 50 steps; late-only guidance preserves unguided value (+0.3).For SD 2.1, guidance 7.5 in the first 9 of 30 steps costs 16.5 of the 19.5 points lost with guidance 7.5 throughout, relative to guidance 2.Source: §B.1 · p. 14, Figure S1 · p. 14
• 43.5 on klein is reached by using the unguided base before handing off to the student at noise 0.85, versus 26.1 when their order is reversed.On the distilled four-step schedule, replacing the first two predictions with the unguided base raises klein from 17.1 to 39.9 and SD-Turbo from 18.3 to 36.4.Source: §B.1 · p. 14, Figure S1 · p. 14
• 45.9 on klein but 40.7 on SD 2.1 are reached with transferred interval guidance, compared with 49.8 for unguided SD 2.1.The transferred interval still guides SD 2.1 within its sensitive early range; schedule effects depend on the noise levels covered.Source: §B.1 · p. 14

B.2. Target design and stability

• Guided targets lose 23.1 points on klein and 13.2 on SD-Turbo, retaining only 11% and 48% of the unguided-target gain, respectively.Across all three families, default-guidance targets also retain at most 32% of the Vendi gain and 25% of the recall gain.Source: §B.2 · p. 15
• 32.1 on klein and 32.2 on SD-Turbo result when targets end at the student’s original first level, versus 40.0 and 41.2 for the deeper tested jumps.The selected intermediate targets perform better than either tested endpoint, but the comparisons do not define a universal structural threshold.Source: §B.2 · p. 15
• Seedwise target correspondence has a cosine-similarity advantage of 0.32 for klein and 0.33 for SD-Turbo, above base-first and distilled references.The comparisons support the interpretation that the adapter learns the target trajectory.Source: §B.2 · p. 15

C.1. Guidance and distillation across generator families

• SD 2.1 reaches 49.5 at guidance 2, compared with 30.0 at guidance 7.5; guidance changes improve tested models but do not identify one universal weight.Removing guidance improves tested klein and Z-Image bases, while reducing it to 2 improves SD 1.5, SD 2.1, and SDXL.Source: §C.1 · pp. 15–16, Table S5 · p. 16
• The evaluated distilled releases perform at or below their tested guided bases, but SRPO improves on FLUX.1-dev by +3.1.Three of four SDXL students use the base’s resolution, so resolution alone cannot explain that pattern.Source: §C.1 · p. 16, Table S5 · p. 16

C.2. Additional generation-cost comparisons

• Transferred interval guidance performs well on klein but remains below unguided sampling on SD 2.1; FSG, MoNO, and skipping the first step remain below the klein adapter.These findings concern the transferred implementations and settings, not optimized limits for each method.Source: §C.2 · p. 17, Figure S2 · p. 17

C.3. Additional primary results and alternative adaptation prompts

• Table S6 compares primary adapters and COCO-caption adapters across ImageNet-100, 200-class development, and full ImageNet-1k protocols.Its caption specifies 500, 500, and 1,281 images per class, respectively, but the supplied passages do not state the table’s cell values.Source: §C.3 · p. 17, Table S6 · p. 18

C.4. DP-DMD’s released student

• On DP-DMD’s SD3.5-Medium student, the adapter adds +7.7 at the first student level and +13.4 at the second, exceeding base-first at the first level by +5.9.The comparison establishes an advantage over the coarse base-first sampler but does not isolate teacher integration, hand-off mismatch, and student continuation.Source: §C.4 · p. 18, Table S7 · p. 18

D.2. Additional generator comparisons on fine-grained datasets

• Our adapter exceeds SDA by 6.5 points on Food-101, 5.3 on Cars, and 11.6 on Pets; all twelve class-bootstrap intervals across the reported comparisons lie above zero.It also improves DP-DMD’s student by 15.0, 10.8, and 34.1 points on those datasets, respectively.Source: §D.2 · p. 20, Table S10 · p. 20
• The reported class-bootstrap intervals condition on the generated sets and fitted adapters, not variation across independent regenerations or adapter-training runs.The gains are mean-class accuracy differences, which weight classes equally, unlike Table 4’s image-weighted top-1.Source: Table S10 · p. 20, §D.2 · p. 21

D.3. Transfer of learned representations

• Adapter-trained ImageNet classifiers yield frozen features that gain 10.7–28.5 points over distilled-release features on Food-101, Cars, Pets, and the 900 remaining ImageNet classes.Klein’s adapted-image features reach accuracy comparable to its unguided base across all four targets, using two denoiser evaluations per image.Source: §D.3 · p. 21, Table S12 · p. 21

D.4. Reference accuracies for other learners

• The learner comparison uses matched generated sets, while its released-student and adapter results are reported in Table 5.The supplied passages do not state the alternative learners’ comparative outcomes.Source: §D.4 · p. 21, Table S13 · p. 21

E.1. Rich captions

• With real-image captions, SD 2.1’s guidance gap is 13.8 points versus 19.5 with templates, and klein’s distillation gap is 8.0 versus 27.9.The first training runs of the adapters retain 81% and 98% of unguided-base gains under captions.Source: §E.1 · p. 22, Table S14 · p. 22
• For klein, base-first has the highest observed accuracy in the caption comparison, while the template-trained adapter uses two student evaluations and released generation needs no adapter preparation.Real-image captions require access to training images; image-free descriptions use class names and seeded scene/viewpoint hints.Source: §E.1 · p. 22

E.2. Fitting adapters on image-free LLM descriptions

• Each image-free-description adapter was fit on 16,000 descriptions, with one initial noise sample per description; evaluation used three independently sampled datasets and three classifier seeds.Pooled intervals condition on the three generated datasets and fitted checkpoints, not adapter-training variation or new dataset realizations.Source: §E.2 · pp. 22–23, Table S15 · p. 23
• The preferred fitting prompt depends on generator: the LLM-fitted klein adapter beats its template-fitted counterpart on every dataset, while template-fitted SD-Turbo remains better under both prompt styles.The klein adapter has no resolved advantage over base-first; one post-hoc sensitivity analysis also leaves that advantage unresolved.Source: §E.2 · p. 23, Table S16 · p. 23

E.3. Frozen-adapter evaluation with LLM prompts on ImageNet-1k

• On the 900 classes outside fitting, the selected klein adapter gains −1.6 points and SD-Turbo gains +6.9 over released students under LLM prompts.The evaluation uses 1,000-way predictions; the selected SD-Turbo adapter improves on both references, while the selected klein adapter falls below both under matched LLM prompts.Source: §E.3 · p. 24, Table S17 · p. 24
• Klein’s LLM-fitted adapter changes from −4.6 points versus the released student at 100 images/class to +5.2 at 500, while both sizes use 100 epochs.The larger set therefore also receives more optimizer updates, so the comparison does not separate data-size effects from optimization budget; a post-hoc template-adapter check found neither tested klein adapter improved over the released student under these prompts.Source: §E.3 · p. 24

E.4. Augmenting real training data

• With 50 real images/class, adapted images beat released-student images by 12.1 points for klein and 13.1 for SD-Turbo; at full real-data scale, adapted augmentation helps shifted tests but slightly lowers in-distribution accuracy.Released klein augmentation lowers all three reported accuracies at full real-data scale.Source: §E.4 · pp. 24–25, Table S18 · p. 25

E.5. Accuracy on shifted test sets

• Shifted-test classifiers were evaluated without retraining, while the feature-space diagnostics cover a fixed collection of 186 template-prompt sets and do not establish causal mediation.Later evaluations do not expand that diagnostic collection.Source: §E.5 · p. 25, F. Diagnostic metrics and their scope · p. 25, Table S18 · p. 25

F.1. Metric comparisons and collection sensitivity

• DINOv3 recall, CLIP recall, and CLIP Fréchet distance have similar aggregate associations with accuracy, while coverage, Vendi, and FID are weaker.Recognizability, Inception Score, and preference scores correlate negatively with accuracy.Source: §F.1 · p. 26, Table S19 · p. 26
• The ranking protocol tests ranking rather than calibration, and its intervals condition on the configuration collection without accounting for shared prompts, seeds, or neighboring plans.Held-out-family checks assess transfer only across studied families; SDXL and FLUX.1-dev remain uncertain.Source: §F.1 · p. 26, Table S20 · p. 27, Table S21 · p. 27

F.2. Adapted-set diagnostics

• Adaptation improves FID in all three families while reducing recognizability, but classwise recall changes do not reliably explain accuracy gains.For the first klein and SD-Turbo adapters, classwise recall changes correlate inconsistently with accuracy gains.Source: §F.2 · pp. 27–28, Table S22 · p. 27

G. Supporting accuracy contrasts

• Table S23 collects paired intervals for generator, data-scaling, and augmentation comparisons, including comparisons with matched classifier-training steps.Its estimates support Sections 4, D.1, and E.4.Source: Table S23 · p. 28

LLM Usage Statement

• The authors report using Claude Code for experiment and analysis code, GPT-6.1-Sol for manuscript revision, and Qwen2.5-VL-7B-Instruct for captions and class descriptions.They state that the authors reviewed the resulting code and text and take responsibility for the reported results and conclusions.Source: LLM Usage Statement · p. 29