acceptodds
Under review as a conference paper at ICLR 2027

Generate to Teach: Adapting Flow Models with Class-Contrastive LoRA for Few-Shot Classification

Abstract

For synthetic data to help few-shot classification, generated images must preserve the distinctions among the target classes. Yet standard generator adaptation fits each support image only to its assigned prompt, leaving these relative distinctions implicit. We introduce TaskRel, a class-contrastive LoRA method for flow-matching generators that turns the generator's own conditional prediction errors into task supervision. For each support example, TaskRel evaluates the correct and competing class prompts on the same noisy latent and target velocity, so the class condition is the only variable. A finite-margin hinge favors the correct condition alongside positive flow-matching fitting. An evolving relation map selects hard competitors, while a scheduler-derived constant controls comparison strength independently of selection. The resulting shared generator is trained solely from labeled supports and its own predictions, without downstream-classifier feedback or inference-time filtering. Across six benchmarks with Stable Diffusion 3.5 Medium, TaskRel reaches 67.50% mean accuracy with a scratch ResNet-50, improving over the matched flow-matching control by 3.36 points and LoFT, the strongest external comparator, by 3.76 points. Its frozen synthetic pools retain gains across scratch ViT and CLIP learners and across 25–100 generated images per class. Thus, class relations provide a direct adaptation signal for reusable synthetic training data while preserving ordinary class-conditioned sampling.

Primary area:
generative models
Keywords:
few-shot classification, generative data augmentation, flow matching, low-rank adaptation, contrastive learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rox3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The manuscript introduces TaskRels, a framework for adapting flow-matching generative models to generate synthetic data for few-shot classification problem. THe motivation of the method is to preserve the distinction between classes in synthesized images, which is a weakness of previous methods where they usually rely on prompt input to do so. TaskRels combines positive fitting for flow models with a contrastive loss, an adaptive competitor sampling, and scheduler-based loss weighting to encourage preservation of class distinctions. Experiments across six datasets show improved average classification accuracy over the baselines and reusability of the generated data across different architectures.

Critical strengths
  1. Promising results. TaskRel improves over the matched positive control on all datasets and achieves the strongest average performance among the evaluated methods. However, the gains vary by datasets, and the baselines methods are not evaluated in their strongest settings (smaller number of synthetic data per class).

  2. Reusability of synthetic datasets across classifier architectures. The frozen synthetic pools retain average improvements when used to train other (downstream) classifiers, including scratch ViT and pretrained CLIP. This supports their model-agnostic property, not requiring additional generator adaptation or classifier feedback during generation. This could reduce the overall cost of synthesis process in practice.

Critical weaknesses
  1. The writing are hard to follow in main method, and the motivation and the necessity of individual components are insufficiently established. The writing, especially in Section 3, makes it quite difficult to understand the motivation and distinguish the contribution of each proposed component. It is unclear for the reviewer why each component was designed and its relationship with the main motivation of preserving the class distinctness. Also, it's hard to see why the proposed components are complementary to each other.

  2. The evaluation lacks some important validation. The accuracy alone does not provide good evidence of improved class distinctness. Figure 3 presents selected qualitative examples and class-level recall, but they were only done for 2 datasets. It does not show for the failure cases of Caltech and Flowers (the datasets on which TaskRels performs worse) to explain the weaker results, and identify whether increased distinctness could come at the expense of within-class diversity.

The computational analysis is also incomplete relative to external methods. The appendix reports some comparison but lacks accuracy/cost comparisons against baselines such as LoFT and ChimeraLoRA. Also, as mentioned before, these baselines use shortened training recipes compared to their original settings.

Questions for authors
  1. Could the authors provide a more detailed analysis of distinctness between classes, together with qualitative comparisons containing both success and failed datasets?
  2. Could the authors compare accuracy against total adaptation and synthesis time for TaskRel, FM, LoFT, and ChimeraLoRA?
  3. What motivates the choice adaptive classes multiplier of 0.03? How sensitive are the results to these choices?
1 comment

W1/Q3. What does each component do, and why 0.03?

The components control different parts of class-relative supervision. Shared corruption and target make class condition the comparison variable. The hinge requires a true-class error advantage and contributes zero once the current gap meets the margin. The relation map chooses the three competitors. The coefficient sets their influence on positive fitting. Generation then uses the ordinary sampler.

At fixed hinge, margin, comparison count and coefficient, the four-domain mean is 54.18% for FM, 57.48% for uniform competitors, 58.27% for frozen soft and 58.68% for refreshed soft. The independent EuroSAT comparison gives refreshed 68.638±0.215% and frozen 67.605±1.267%, three students per pool. Allocation adds to the main comparison effect.

The fixed-state optimizer diagnostic isolates loss scale. Across 32 identical evaluation batches, raw NLL's median gradient norm is 147.90 times the unit-weight hinge's, triggering clipping in 32 versus 3 batches. Generator-derived NLL scaling reduces clipping to 5 batches. The final scheduler rule c=E[σ~2]c=\mathbb E[\widetilde\sigma^2] is a separate calibration against the positive loss's mean weight.

The factor 0.03 sets m=0.03median⁡(ey)m=0.03\operatorname{median}(e_y), a dimensionless fraction of warm-up fitting error. It is a margin multiplier, not a class-sampling multiplier. Changing it changes the active comparisons. We tested half and twice the default with supports, warm-up, graph, horizon and synthesis fixed, using 100 images/class.

EuroSAT marginAccuracy, mean±SDGain over FM
FM64.48±0.39—
0.01568.53±1.57+4.05
0.0369.32±1.26+4.84
0.0674.64±0.80+10.16

All tested margins outperform matched FM at all three student seeds. The benefit survives a factor-of-two change in either direction, and 0.06 adds 5.32 points over the default without extra updates or images. The submitted default remains unchanged. SD describes student variation on fixed generated pools.

Q1/W2. Distinctness, diversity and the weaker domains

We tested discrimination directly on 148 held-out Pets images and 40 EuroSAT images, with all conditions and four shared corruptions per image. Minimum-average-error class accuracy rises 58.11→67.57% and 47.50→65.00%, with hinge activity falling 25.89/25.76 points. The gain extends beyond the supports.

The image audit covers all 396 classes and 100 generated images/class. Entries below are FM→TaskRel class averages. Agreement uses frozen CLIP. P/R/C use DINOv2-L precision/recall/coverage at k=5k=5 against held-out real images [1].

DatasetAgreementPRC
Pets0.865→0.9130.707→0.8520.368→0.0920.836→0.869
DTD0.518→0.5990.899→0.9640.239→0.1410.832→0.865
EuroSAT0.397→0.4750.420→0.4820.277→0.1390.064→0.078
Aircraft0.172→0.2010.755→0.8170.751→0.7000.752→0.792
Caltech0.918→0.8780.778→0.7170.352→0.2910.882→0.800
Flowers0.658→0.5540.151→0.0540.412→0.1720.211→0.068

Four domains improve agreement, precision and coverage. Aircraft additionally improves unbiased polynomial-kernel MMD² from 3.406 to 2.966. Downstream recall improves on 76/100 Aircraft and 38/47 DTD classes across three students.

Caltech/Flowers have weaker feature alignment and lower recall. Caltech's synthetic dispersion rises 0.322→0.338, so reduced diversity alone cannot explain every result. TaskRel still improves scratch accuracy over matched FM by 0.257/1.868 points there. With pretrained CLIP, its 96.47%/96.75% exceeds every compared synthetic source on both domains. Feature geometry and classifier utility measure distinct properties.

The accompanying fixed-selection panels cover all six domains, including Caltech and Flowers. Two classes per domain and a common within-class synthetic index were selected with seed 91023, independently of performance. Distributional conclusions use the all-class metrics.

Q2/W2. Cost and external methods

Calibrated FM-50/TaskRel adaptation hours are 1.504/8.284 on Pets and 0.415/1.978 on EuroSAT, including probing. FM extended to those budgets gives scratch accuracy 60.898/68.786%, versus TaskRel's 66.485/69.321%. Pets FM-50 is stronger at 65.595% and remains the relevant comparator. These are fixed terminal points.

Both methods generate at 1.484–1.487 s/image and 16.46 GiB, batch 2, 24 steps and CFG 2, excluding loading/writing. Extra adaptation is paid once, and pools are reused across learners. EuroSAT banks contain 15.811M learned parameters for FM/TaskRel, 187.761M for LoFT and 199.496M for Chimera, counting shared factors once. These are bank sizes, not active capacity or time. Comparable isolated LoFT/Chimera timings are unavailable. Shared-server times are excluded from the calibrated profiles. External rankings concern the disclosed SD3.5 recipes.

Reference

[1] M. F. Naeem et al. Reliable Fidelity and Diversity Metrics for Generative Models. ICML, 2020.

Reviewer_yTAbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes TaskRel, a method for adapting pretrained flow-matching generators to produce synthetic data for few-shot classification. Its main idea is a contrastive objective that encourages lower prediction error under the correct class label than under incorrect labels, using the same noisy input and prediction target. The method also incorporates calibration, an evolving class-relation matrix for selecting competing labels, and a scheduler-derived loss coefficient. It is evaluated on six classification benchmarks with multiple downstream classifiers.

Critical strengths

The evaluation covers six datasets and includes comparisons with simpler sampling and weighting variants. The results provide useful evidence that class-comparison supervision can improve synthetic-data utility over positive-only flow matching at matched training epochs.

Critical weaknesses
  1. The main comparisons do not provide a fair assessment under a common computational budget. Matching training epochs gives TaskRel substantially more adaptation compute than the FM baseline: Table 16 reports approximately 4.8–5.5× greater cost in the two profiled settings. When FM receives comparable compute on EuroSAT, TaskRel’s advantage falls to 0.54 points with a scratch classifier and essentially disappears with a pretrained classifier. On Pets, its scratch-classifier advantage over the best reported FM result within the allowed budget is only 0.89 points. These results substantially weaken the practical significance of the headline gains. The main comparisons do not establish that TaskRel offers a better accuracy–compute tradeoff.

  2. The benefit beyond simpler class-contrastive adaptation is insufficiently established. The core objective follows a familiar margin-ranking formulation, and the paper acknowledges related class-contrastive supervision in MBTI without providing a numerical comparison. The additional adaptive design also has limited support: in Table 2, updating the relation matrix improves average accuracy by only 0.41 points over keeping it fixed, while reducing EuroSAT accuracy by 3.97 points. The results support the basic use of class comparisons more clearly than they establish the necessity of the full procedure. It remains unclear whether the proposed design provides a sufficiently substantial advance over simpler alternatives.

  3. Figures 1 and 2 are too small and dense to read comfortably at the intended printed size. Essential labels and connections are difficult to follow, making the central method unnecessarily hard to understand. The figures should be enlarged and simplified, with implementation details separated from the main idea.

Questions for authors
  1. Do the main accuracy gains remain when FM receives the same adaptation compute as TaskRel? Table 16 shows that FM largely catches up on EuroSAT. Please provide a comparable evaluation on another main benchmark, such as Aircraft, allowing FM to use its best validation-selected checkpoint within the same compute budget.

  2. Does updating the class-relation matrix reliably improve performance over keeping it fixed? The average improvement in Table 2 is small, and the frozen matrix performs substantially better on EuroSAT. Please show whether the improvement persists with another independently trained generator and newly generated synthetic dataset.

1 comment

Q1/W1. Does the benefit remain at comparable compute?

Matching adaptation compute retains positive mean gains, including a near tie for pretrained EuroSAT. The complete comparison retains both FM budgets. Hours are calibrated from CUDA-event profiles, including relation probing and excluding loading/checkpointing. Accuracy is mean±SD across three students on each fixed pool.

Dataset/armHoursScratch RN50Pretrained RN50
Pets/FM-501.50465.595±0.86988.217±0.103
Pets/FM-extended8.28460.898±0.93188.035±0.992
Pets/TaskRel-508.28466.485±0.46989.579±0.247
EuroSAT/FM-500.41564.477±0.39186.000±0.504
EuroSAT/FM-extended1.97868.786±1.92887.930±0.616
EuroSAT/TaskRel-501.97869.321±1.25887.947±0.236

Against the better reported FM point, paired TaskRel gains are 0.890±1.103/1.363±0.310 points on Pets and 0.535±0.989/0.016±0.851 on EuroSAT for scratch/pretrained RN50. The comparison shows the residual effect after increasing FM compute, alongside the larger fixed-horizon gains. The completed control compares terminal checkpoints on Pets and EuroSAT. A validation-selected Aircraft frontier remains a separate empirical question.

The practical return is the utility of a reusable training set. The same pools improve six-domain means for scratch RN50, scratch ViT and CLIP. The expanded CLIP comparison gives TaskRel 84.57%, FM 82.52%, DataDream 81.98%, LoFT 80.49%, Chimera 80.19% and DA-Fusion 68.34%, at seed 0 under one student recipe. TaskRel leads all five synthetic comparators on every domain. At 25/50/100 images per class, its scratch means are 54.73/62.51/67.37%, versus FM's 52.32/60.13/64.47%. Those are fixed-100-epoch, seed-0 operating points.

The additional comparison cost is paid during adaptation. FM and TaskRel both sample at 1.484–1.487 s/image and 16.46 GiB, batch 2, 24 steps and CFG 2, excluding loading/writing. Reuse is shared by frozen-pool methods. TaskRel's distinction is the measured usefulness of its pool across learners, with ordinary generation after adaptation.

Q2/W2. Does refreshing survive an independent pipeline?

We completed the direct frozen/refreshed comparison on the independent EuroSAT support/generator/generation realization. Both arms share its warm-up, initial graph, coefficient, horizon and generation settings. Each supplies 100 images/class to scratch-RN50 seeds 0/1/2.

AllocationMean±student SD
Frozen67.605±1.267
Refreshed68.638±0.215

Refreshing gains 1.033 points on the new realization. The paired-difference SD is 1.482, with two positive seeds out of three. The result reverses the original EuroSAT mean ordering, where frozen led by 3.97 points. It supplies the independent comparison requested, while retaining the distinction between student variation and generator replication.

The original controlled allocation ablation gives

RuleFour-domain meanGain over FM
FM54.180.00
Uniform57.48+3.30
Frozen soft58.27+4.09
Refreshed soft58.68+4.50

The hinge supplies the larger gain. Relation-guided allocation improves on uniform selection, and refreshing adds 0.41 points over frozen soft on average. The new independent result supports an additional mean effect across a second generated pool. The method's contribution combines useful class-relative supervision with an allocation rule evaluated against simpler alternatives.

MBTI's core textual inversion separates predictions while retaining fitting and regularization [1]. TaskRel compares their errors against the same target. Writing r=vy−ur=v_y-u and δ=vj−vy\delta=v_j-v_y, the gap is (∥δ∥2+2r⊤δ)/d(\|\delta\|^2+2r^\top\delta)/d. Predictions u+au+a and u−au-a have positive separation but equal error. The target-relative criterion consequently enforces a different requirement. A matched MBTI numerical comparison is not supplied. New held-out probes verify the intended discrimination, with minimum-error class accuracy rising 58.11→67.57% on Pets and 47.50→65.00% on EuroSAT using four paired corruptions per image.

W3. Figures

We will enlarge the shared-state comparison path in Figure 1 and show warm-up, probing and continuation as three readable stages in Figure 2.

Reference

[1] B. Chae et al. MBTI: Metric-Based Textual Inversion for Fine-Grained Image Generation. WACV, 2026.

Reviewer_EMWjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose TaskRel, a LoRA-based method for adapting a flow-matching generator, specifically Stable Diffusion 3.5 Medium, to synthesize training data for few-shot image classification tasks. Its adapter architecture largely follows that of DataDream, an earlier baseline. The main difference is an additional class-contrastive hinge loss, which encourages the generator to favor the correct class over competing classes, an idea closely related to the class comparisons used in MBTI and Diffusion Classifier. When the generated images are used (together with a small set of real support images) to train a ResNet-50 model from scratch, TaskRel achieves higher average classification accuracy than the evaluated baselines, including a matched positive-only, DataDream-style control that omits the class-contrastive loss.

Critical strengths

[S1] Empirical Benefit of Class-Contrastive Supervision

Adding class-contrastive hinge supervision to DataDream-style fine-tuning yields improvements over the matched positive-only control on all six datasets in the main evaluation.

Critical weaknesses

[W1] Limited Justification for Using Flow Matching

In the submitted paper, I could not find a clear justification for (i) what makes the class-contrastive objective in TaskRel specifically tailored to flow matching, and (ii) why flow matching is particularly advantageous for synthetic data generation.

The paper motivates its use of flow matching by pointing to the availability of a known target velocity (lines 65–68). However, conventional diffusion training also regresses toward a known target, such as the added noise ϵ\epsilon [1] or a vv-prediction target [2], and closely related comparisons of class-conditional errors on a shared noisy input are already used by Diffusion Classifier [3] and MBTI [4] with conventional denoising diffusion models. As a result, it is unclear which part of TaskRel relies on flow matching. Given the shared LoRA adaptation used in DataDream and the class comparisons already explored in prior work, the additional methodological contribution of TaskRel remains unclear.

Furthermore, the paper does not explain why a flow-matching generator is a better choice for synthetic data generation than conventional diffusion models such as Stable Diffusion (SD) 2.1 [5], on which prior methods, including DataDream, LoFT, and ChimeraLoRA, were developed. In addition, although the baselines are ported to SD 3.5 Medium to enable a common-backbone comparison, they were originally developed for conventional diffusion models. Even though the matched control supports the benefit of class-contrastive training on SD 3.5, it does not resolve concerns about the comparison with the ported external methods. This makes it difficult to determine whether the advantage of TaskRel over the external baselines reflects the proposed method itself or differences in how well the compared methods transfer to SD 3.5.

Finally, the evaluation is limited to SD 3.5 Medium, as mentioned in Section F.2. Consequently, the evaluation establishes neither whether TaskRel is particularly suited to flow matching nor whether its benefits generalize across flow-matching generators.

[W2] Concerns about Experimental Comparability

First of all, the paper provides limited justification for using a randomly initialized classifier as the primary evaluation setting. In a few-shot setting, strong representations from pre-trained models such as CLIP and ImageNet pre-trained ResNet are available, and prior synthetic-data generation methods such as DataDream, LoFT, and ChimeraLoRA primarily evaluate whether generated data can improve such pre-trained models. In contrast, the main results in this paper use a ResNet-50 trained from scratch. While this setting may help isolate the information contained in the generated dataset, the paper does not clearly explain why pre-trained representations are excluded from the primary evaluation, especially in a setting motivated by data scarcity. Although Tables 3 and 8 include pre-trained classifiers, the CLIP comparison (Table 3) includes only LoFT and the matched FM control and omits a real-only baseline, and the pre-trained ResNet-50 comparison (Table 8) includes no external method. This leaves the benefit over real-only few-shot adaptation and the ranking against the other external methods unclear in the pre-trained setting.

Next, the methods differ substantially in their adapter configurations and training procedures, but neither model capacity nor computational cost is fully accounted for in the comparison. TaskRel applies rank-16 LoRA to both the transformer and two CLIP text encoders (CLIP-L and CLIP-G in SD 3.5), whereas the ported LoFT and ChimeraLoRA keep the text encoders frozen. At the same time, ChimeraLoRA uses multiple instance-specific LoRA heads, so the overall parameter budget is not obvious from the reported ranks alone. Furthermore, parameter counts alone do not capture adaptation cost, since TaskRel requires additional computation for class comparisons and relation-map construction. Table 16 estimates the adaptation time for TaskRel at approximately 4.8–5.5 times that of the 50-epoch FM control, but corresponding costs for the external baselines are not provided. The same table also includes a compute-matched FM control, though only on Pets and EuroSAT. On EuroSAT, it closes most of the gap to TaskRel. On Pets, extending the same FM training recipe lowers test accuracy, leaving open whether a better-tuned recipe could use the same computational budget more effectively. Reporting total trainable parameter counts and adaptation time on the same hardware, together with compute-matched comparisons against external baselines, would clarify whether the accuracy gains persist at comparable computational cost.

Moreover, several baselines are evaluated far from their original operating regimes. For example, DataDream was originally trained substantially longer and evaluated with 500 synthetic images per class. LoFT similarly evaluates 500 images per class and uses substantially longer per-image adaptation. ChimeraLoRA was primarily developed under a 4-shot, 500-synthetic-image setting. In contrast, the main comparison here uses 16 real images and only 100 synthetic images per class, while also shortening the adaptation schedules of these baselines. The authors justify these schedules by matching total support-image exposure across methods, as discussed in Section C.2. However, equal exposure does not establish that the shortened schedules provide sufficient training for each method. Since Table 4 shows substantial accuracy changes across synthetic-data budgets, it is unclear whether the reported ranking persists under their established settings.

Finally, the evaluation coverage is narrower than that of prior work. The six datasets considered in this paper include several fine-grained benchmarks, but the main experiments are restricted to a uniform 16-shot setting. (A separate five-shot study is reported on Pets in Table 13, but it does not establish robustness across shot counts and datasets.) Broader prior evaluations include substantially lower-shot and long-tail settings, as well as specialized domains such as medical images. Consequently, it remains unclear whether the gains extend consistently to lower-shot settings, long-tailed class distributions, or specialized domains.

[1] Denoising Diffusion Probabilistic Models (NeurIPS 2020; Ho et al.)

[2] Progressive Distillation for Fast Sampling of Diffusion Models (ICLR 2022; Salimans and Ho)

[3] Your Diffusion Model is Secretly a Zero-Shot Classifier (ICCV 2023; Li et al.)

[4] MBTI: Metric-Based Textual Inversion for Fine-Grained Image Generation (WACV 2026; Chae et al.)

[5] High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2022; Rombach et al.)

Questions for authors

I have several questions, and I list them below in descending order of importance to my assessment.

[Q1] Evaluation across Generator Backbones

As discussed in [W1], the class-contrastive objective does not appear to rely specifically on flow matching. Could the authors evaluate TaskRel on SD 2.1 against a corresponding positive-only control and, if feasible, against the external baselines in their native SD 2.1 setting? If applying the objective to SD 2.1 requires modification or does not yield comparable gains, could the authors explain which properties of flow matching are important for its effectiveness?

Furthermore, the evaluation is limited to SD 3.5 Medium. Could the authors evaluate TaskRel on another flow-matching generator to assess whether the gains over positive-only adaptation generalize across flow-matching generators?

[Q2] Characterization of the Minimizer of the TaskRel Objective (Eq. (7))

From Eq. (1), u=(zσ−z)/σu = (z_\sigma - z)/\sigma. Therefore, for h=(zσ,t)h = (z_\sigma, t), μj(h)−μy(h)=E[Z∣h,y]−E[Z∣h,j]σ,\mu_j(h) - \mu_y(h) = \frac{\mathbb{E}[Z \mid h, y] - \mathbb{E}[Z \mid h, j]}{\sigma}, and, by Eq. (4), at the unrestricted population optimum of positive-only regression, E[Δyj∣h,y]=∥E[Z∣h,y]−E[Z∣h,j]∥2σ2d.\mathbb{E}[\Delta_{yj} \mid h, y] = \frac{\lVert \mathbb{E}[Z \mid h, y] - \mathbb{E}[Z \mid h, j] \rVert^2}{\sigma^2 d}. Consequently, the expected gap attained by the positive-only flow-matching solution is determined by the data distribution and varies across class pairs and noise levels.

However, Eq. (3) imposes the same fixed margin m=0.03⋅median(ey)m = 0.03 \cdot \mathrm{median}(e_y) on all pairs and noise levels. Whenever this population gap is below mm, the expected hinge loss is positive at the positive-only solution. Since the positive term has zero gradient there, the gradient of Eq. (7) reduces to that of the hinge. For a pair active for all target draws at hh, the expected hinge gradient vanishes with respect to vyv_y and is proportional to −(μj(h)−μy(h))-(\mu_j(h) - \mu_y(h)) with respect to vjv_j. For such a pair, gradient descent pushes vjv_j away from μy(h)\mu_y(h), increasing the expected competing-condition error. Whether the positive-only flow remains stationary for the full objective depends on the aggregate gradient across pairs.

As also acknowledged in Section B.2, this margin can be satisfied by increasing the competing-condition error eje_j, rather than improving the correct-condition fit eye_y. Moreover, the hinge is applied to individual noisy samples rather than to the expected gap. By Jensen’s inequality, E[[m−Δyj]+]≥[m−E[Δyj]]+,\mathbb{E}\big[[m - \Delta_{yj}]_+\big] \geq \big[m - \mathbb{E}[\Delta_{yj}]\big]_+, so the hinge loss may remain positive even when the expected gap exceeds mm. This raises the question of whether the proposed objective trades off fidelity to the positive-only flow-matching solution for discriminability. Downstream accuracy with a classifier trained from scratch cannot by itself distinguish faithful samples from samples that deviate from the real data distribution but remain highly discriminative. Could the authors (i) characterize the population minimizer of Eq. (7), or clarify when the positive-only flow remains a minimizer or stationary point; (ii) quantify whether the hinge is satisfied mainly by decreasing eye_y or increasing eje_j; and (iii) report a distributional or perceptual fidelity metric in addition to downstream classification accuracy?

[Q3] Baseline Comparisons with Larger Synthetic Datasets and Longer Training

As discussed in [W2], the external baselines are evaluated with fewer synthetic images and shorter adaptation schedules than in their original evaluations. Could the authors repeat the comparisons on SD 3.5 Medium using 500 synthetic images per class and longer adaptation schedules consistent with the original baseline recipes? This would clarify whether the reported ranking persists when the external baselines receive more training and larger synthetic datasets. In addition, could the authors report a real-only baseline and zero-shot CLIP for the CLIP setting in Table 3?

[Q4] Combining TaskRel with Existing Adaptation Methods

TaskRel can be viewed primarily as an additional class-contrastive training objective on top of shared generator adaptation. In this sense, it appears largely orthogonal to some of the compared methods. For example, combining TaskRel with ChimeraLoRA seems plausible, since ChimeraLoRA changes the adapter parameterization while TaskRel changes the training objective. Could the authors evaluate at least one such combination under the protocol of Table 1 on a subset of datasets? This would help clarify whether TaskRel constitutes a generally useful objective that complements existing synthetic-data generation methods, or whether its gains are specific to the DataDream-style shared-LoRA setup.

[Q5] Hyperparameter Sensitivity

Several hyperparameters in TaskRel appear to be fixed without sensitivity analyses, including the margin coefficient (m=0.03⋅median(ey)m = 0.03 \cdot \mathrm{median}(e_y)), the number of competing classes (Kn=3K_n = 3), the number of probe states per support image (Kg=8K_g = 8), and the 25% positive-only warm-up ratio. Could the authors provide sensitivity analyses for these choices to show that the reported gains are not sensitive to these specific settings?

2 comments

W1/Q1. Why flow matching?

(i) TaskRel is tailored through velocity-space supervision and scheduler calibration. With zσ=(1−σ)z+σϵz_\sigma=(1-\sigma)z+\sigma\epsilon and u=ϵ−zu=\epsilon-z, the gap measures how class conditioning changes vector-field fitting. Our clean-latent loss satisfies z^y=zσ−σvy,ℓ+=∥z^y−z∥2/d=σ2ey.\hat z_y=z_\sigma-\sigma v_y,\qquad \ell_+=\|\hat z_y-z\|^2/d=\sigma^2e_y. We set c=E[σ~2]≃0.542c=\mathbb E[\widetilde\sigma^2]\simeq0.542 under our scheduler and timestep law, matching the mean positive-loss weight without noise-weighting individual hinges. Margin and competitors use velocity errors. Diffusion applications require recalibrated margins and weights.

(ii) We do not regard adopting newer technology as a drawback of our paper. Flow matching underlies mainstream image generators, including SD3 [1] and FLUX [2]. Esser et al. [1] report better quality retention at fewer sampling steps. Our aim is not to obtain better scores from a stronger generator. We seek to advance how generators learn task-relative supervision from their own errors and produce reusable data with 24-step synthesis. We welcome further research and hope to contribute to the continued development of the field.

The SD3.5 control matches capacity, supports, positive loss, synthesis and student settings.

Diffusion Classifier uses errors for classification and fine-tuning [3]. MBTI combines prediction separation, fitting and regularization [4]. TaskRel requires a shared-target advantage. For r=vy−ur=v_y-u and δ=vj−vy\delta=v_j-v_y, ej−ey=(∥δ∥2+2r⊤δ)/d.e_j-e_y=(\|\delta\|^2+2r^\top\delta)/d. Predictions u+au+a and u−au-a are separated yet equally accurate. TaskRel couples the target-relative criterion to generator-derived competitors and calibrated strength.

Q2. Population behavior, discrimination and fidelity

Positive fitting bounds departure from the conditional mean while the hinge supplies class supervision. For the analysis, fix h=(zσ,t)h=(z_\sigma,t), σ>0\sigma>0, finite conditional second moments and a sampler independent of the target draw conditional on class. Let πy=Pr⁡(y∣h)\pi_y=\Pr(y\mid h), μy=E[U∣h,y]\mu_y=\mathbb E[U\mid h,y], w=σ2w=\sigma^2, ek=∥ak−U∥2/de_k=\|a_k-U\|^2/d, and ryj=Pr⁡(j∈Ny)/Knr_{yj}=\Pr(j\in N_y)/K_n. With Ey=E[⋅∣h,y]\mathbb E_y=\mathbb E[\cdot\mid h,y], write J=L++HJ=L_++H, where L+=w∑yπyEyeyL_+=w\sum_y\pi_y\mathbb E_y e_y and

H(a)=c∑yπy∑j≠yryjEy[(m+ey−ej)+].H(a)=c\sum_y\pi_y\sum_{j\ne y}r_{yj}\mathbb E_y[(m+e_y-e_j)_+].

For c>0c>0, at a=μa=\mu the positive gradient vanishes. With Iyj=1{m+ey−ej>0}I_{yj}=\mathbf1\{m+e_y-e_j>0\} evaluated there, stationarity requires, for every kk,

=\sum_{y\ne k}\pi_yr_{yk}\mathbb E_y[I_{yk}(\mu_k-U)].$$ Activity remains inside the expectation. Use admissible subgradients at hinge boundaries. If $H(\mu)=0$, then $J(a)\ge L_+(a)\ge L_+(\mu)=J(\mu)$. Any global minimizer obeys $$\frac{w}{d}\sum_y \pi_y \left\lVert a_y^* - \mu_y \right\rVert^2 \le H(\mu) - H(a^*) \le H(\mu).$$ The bound follows from $J(a^*)\le J(\mu)$ for the frozen-sampler population objective. We tested 148 held-out Pets and 40 EuroSAT images using four shared corruptions and all class conditions. FM-relative changes in units of $10^{-3}$ are |Dataset|Own error|Competing error|Gap| |---|---|---|---| |Pets|+1.071|+5.411|+4.340| |EuroSAT|+0.321|+1.476|+1.156| The gap grows mainly through competing error. Native positive-loss increases are 0.35%/0.54% relative to FM. Minimum-average-error class accuracy rises 58.11→67.57% on Pets and 47.50→65.00% on EuroSAT, while hinge activity falls 25.89/25.76 points. **The learned change improves discrimination on unseen images.** All-class DINOv2-L metrics against held-out references give FM→TaskRel values. P/R/C are precision/recall/coverage at $k=5$ [5]. MMD² is the unbiased polynomial-kernel estimate. |Dataset|P|R|C|MMD²| |---|---|---|---|---| |Pets|0.707→0.852|0.368→0.092|0.836→0.869|3.582→4.570| |DTD|0.899→0.964|0.239→0.141|0.832→0.865|4.876→5.988| |EuroSAT|0.420→0.482|0.277→0.139|0.064→0.078|5.021→5.199| |Aircraft|0.755→0.817|0.751→0.700|0.752→0.792|3.406→2.966| |Caltech|0.778→0.717|0.352→0.291|0.882→0.800|3.944→5.306| |Flowers|0.151→0.054|0.412→0.172|0.211→0.068|12.332→16.762| Four domains improve precision and coverage. Aircraft also improves MMD², with class-bootstrap intervals excluding zero for all three changes. Recall and Caltech/Flowers outcomes distinguish task utility from uniform distributional improvement. Downstream gains establish useful supervision across all six domains. ### References [1] P. Esser et al. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. ICML, 2024. [2] Black Forest Labs et al. FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space. arXiv:2506.15742. [3] A. C. Li et al. Your Diffusion Model is Secretly a Zero-Shot Classifier. ICCV, 2023. [4] B. Chae et al. MBTI: Metric-Based Textual Inversion for Fine-Grained Image Generation. WACV, 2026. [5] M. F. Naeem et al. Reliable Fidelity and Diversity Metrics for Generative Models. ICML, 2020.

W2/Q3. Pretrained learners and fair comparisons

We found no highly relevant SD3.5 baseline when designing the study. The closest works, DataDream, LoFT and ChimeraLoRA, used SD2.1. Porting them was therefore the only viable way to make the intended comparison on a common SD3.5 backbone.

With the backbone fixed, TaskRel gains 3.36 scratch-RN50 points over FM on average, positive on all six domains, and 3.76 over ported LoFT.

TaskRel leads all synthetic sources on all six domains with CLIP. Adapted rows use CLIP ViT-B/16, 16 real supports/class, 100 synthetic images/class and 100 epochs, seed 0. Zero-shot uses fixed prompts. Avg. weights domains equally.

SourcePetsDTDEuroSATAircraftCaltechFlowersAvg.
Zero-shot90.5745.9238.2323.5888.1171.0559.58
DA-Fusion78.9658.4584.4249.7443.7794.7268.34
ChimeraLoRA89.2365.4378.2055.4596.1996.6380.19
LoFT89.0763.0681.1957.5295.4696.6780.49
DataDream90.6864.9588.4256.1495.8695.8681.98
FM90.8165.1990.6056.9596.4395.1382.52
TaskRel91.6668.5691.6262.3896.4796.7584.57

TaskRel averages +2.05 points over FM and +2.59 over DataDream. All nine paired CLIP student differences on Pets/DTD/EuroSAT are positive. Scratch RN50 tests data without pretrained student features, also evaluated by LoFT [1]. Reusing the pools with scratch ViT gives 45.75% versus FM's 43.71%.

Aircraft armReal-loss controlFMDataDreamTaskRel
Test accuracy68.4169.1968.0571.11

TaskRel gains 2.70 points over the control and 1.92 over FM with the same pool. Aircraft was selected after its earlier deficit. Weighting and update count change jointly, so the diagnostic remains separate from the original table.

Capacity is identical for TaskRel and FM. Each learns 15,810,560 parameters across the transformer and both CLIPs. EuroSAT external banks contain 1.174M/image and 187.761M total for LoFT, and 19.950M/class and 199.496M total for Chimera, counting shared A once alongside 16 B heads. Stored state differs from active capacity and compute.

RTX-4090 calibrated hours include probing, excluding loading/checkpointing. Accuracy is mean±SD over three students per pool.

Dataset/armHoursScratch RN50Pretrained RN50
Pets/FM-501.50465.595±0.86988.217±0.103
Pets/FM-extended8.28460.898±0.93188.035±0.992
Pets/TaskRel-508.28466.485±0.46989.579±0.247
EuroSAT/FM-500.41564.477±0.39186.000±0.504
EuroSAT/FM-extended1.97868.786±1.92887.930±0.616
EuroSAT/TaskRel-501.97869.321±1.25887.947±0.236

Against the better FM mean, unrounded records give TaskRel gains 0.890/1.363 points on Pets and 0.535/0.016 on EuroSAT for scratch/pretrained RN50. The last difference is effectively a tie. These are terminal operating points, not a validation-selected frontier. Comparable isolated external-method timings are unavailable. Both generate at 1.484–1.487 s/image and 16.46 GiB at batch 2, 24 steps and CFG 2, excluding loading/writing.

Table 4 uses 16 real supports/class and 100 scratch epochs, seed 0.

Synthetic/classLoFT avg.FM avg.TaskRel avg.
2550.3952.3254.73
5058.9860.1362.51
10063.8564.4767.37

TaskRel leads throughout, including one quarter of the images. Fixed epochs also change update count and mixture fraction. The longer local DataDream-200 EuroSAT run reaches 70.63±1.91%, versus TaskRel-50 at 69.32±1.26%. Scheduling and text learning rate also differ, so it is a separate recipe, not a duration-only ablation. A 500-image/native-budget comparison has not been run.

Q4. Combining TaskRel and ChimeraLoRA

Combining the mechanisms means holding the same anchor adapter fixed for both conditions while retaining Chimera’s factorization. Our measured gains concern shared adapters. Table 13's five-shot study tests context conditioning, not Chimera factorization. The main scope is six domains with 16-shot support, plus five-shot Pets and ImageNet100 extensions. Medical/long-tail settings are untested.

Q5. Sensitivity

The EuroSAT gain does not depend on exactly 0.03. All tested factors beat matched FM at every student seed. The sweep retains the same supports, warm-up, graph, horizon, synthesis and 100 images/class.

EuroSAT marginAccuracy, mean±SDGain over FM
FM64.48±0.39—
0.01568.53±1.57+4.05
0.0369.32±1.26+4.84
0.0674.64±0.80+10.16

The multiplier scales median warm-up true-class error to set hinge activity. The default stays fixed. SD is over three students per generated pool. Competitor count 3, probe count 8 and 25% warm-up are held fixed, with no sweeps over those settings. Tables 2/15 separately test allocation/weighting.

Reference

[1] J. M. Kim et al. LoFT: LoRA-fused Training Dataset Generation with Few-shot Guidance. BMVC, 2025.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.