acceptodds
Withdrawn from ICLR 2027

Beyond Faster Fields: Auditing Drifting Generative Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Drifting generative models train a generator through an attraction–repulsion field that moves samples toward a target distribution. We ask whether cheaper fields preserve the intended update, reduce full training cost, and maintain distributional fidelity. We compare optimized dense drifting with projected, representative-based, and transport alternatives in a 240-run feature benchmark, a 24-run image screen, and a 66-model generator–encoder audit. An exact aggregation identity shows how to preserve normalization and copy-specific self deletion when compressing repeated sources. Compatible operator references change apparent approximation error. A 9.26× field saving yields only 1.013× inclusive wall speedup in the balanced image screen. At 15k updates in the balanced small-generator/ResNet-18 bridge, DriftXpress uses about 26.7% less recorded training-component time, with lower FIDtv and higher recall in all three paired seeds under both tested Inception input paths; earlier recall gains are input-sensitive. All twelve rare-target screen runs have zero undercoverage, yet classifier-estimated rare-mass means are 5.6–9.9× the target. These results locate a positive checkpoint regime without establishing general transfer or exact matched-quality speedup. The practical takeaway is to compare native operators, declare cost scope, and measure quality, diversity, and two-sided mass error.

Primary area:
generative models
Keywords:
drifting models, generative modeling, kernel approximation, training efficiency, generative model evaluation, mode coverage

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

The authors audits whether cheaper alternative of drifting models are actually worth it.

Critical strengths
  • Comparing if methods actually hold off to their promises is interesting and useful
  • I appreciate the authors honesty about reproducibility failures.
Critical weaknesses
  • 45/48 U-Net/DINOv3 runs have zero recall making the conclusions in limited generative domains pretty limited.
  • Dense and DriftXpress differ in normalization, positive-support access, and approximation. The exact controls needed to separate these effects are missing.
  • The paper lacks results in the domain which the original paper was trained: Imagenet 1K
  • Overall some quality advice: Improve the plot quality (for example zoom figure 2a). Also better writing would be appreciated
Questions for authors

Can you provide experiments on imagenet 256.

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

The paper audits approximations to drifting generative models through 240 feature-space experiments, 24 image experiments, and a 66-model extension. It examines operator equivalence, computational efficiency, generation quality, and distributional calibration, alongside an exact duplicate-aggregation identity.

Critical strengths

The study carefully separates field-level acceleration from training cost and generation quality. It reports unfavorable results rather than selectively emphasizing improvements. The favorable checkpoint results are also meaningful: DriftXpress reduces recorded training-component time by 26.7% while improving FID and recall across three paired seeds under both tested evaluation input paths.

The exact treatment of normalization and self-interaction deletion is technically valuable. The analysis of evaluator sensitivity and two-sided class-mass errors also provides concrete lessons for evaluating generative models.

Critical weaknesses

DriftXpress (Falahati et al., 2026) already investigates preprocessing, end-to-end training efficiency, and quality–runtime convergence. Theoretical discussions of differing drifting operators also exist. The new aggregation identity is valid but restricted to coarsened supports, with no demonstrated practical acceleration advantage. The paper's broader recommendations largely reinforce established evaluation principles rather than establish a new predictive framework.

Dense and DriftXpress differ simultaneously in normalization and target-data access. The missing information-matched exact controls (Table 23) prevent attribution of the observed improvements to field approximation.

Furthermore, the favorable configuration has very low absolute recall, while 45/48 U-Net/DINOv3 runs have zero secondary recall. The three-seed checkpoint evidence and lack of exact matched-quality training-time measurements further limit the practical conclusions.

Questions for authors

Can information-matched exact-field controls isolate the effects of approximation, normalization, and target-data access?

Do the findings persist in configurations achieving substantially stronger generative coverage?

What new, generalizable conclusion does this audit establish beyond existing wall-clock and operator analyses?

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

The paper benchmarks various field acceleration methods and feature spaces for dirfting generative model with operator fiedality, training cost, and distributional quality. Demonstrate that faster field computation can acceleraste the drift model in small generator and ResNet-18 conguration significantly. Also, altering the hyperparameter of feature encoders can bring huge impact on the performance.

Critical strengths
  1. The emprical results are reasonable
Critical weaknesses
  1. "A 9.26× field saving yields only 1.013× inclusive wall speedup in the balanced image screen" seems to be a straightforward conclusion from profiling and Amdahl’s law since the field computation only occupies 2.94%. Additionally, DriftXpress can brings significant improvement in small-generator scenario is also trivial since field computation occupies 27.95% computational time.
  2. The contribution and motivation of proposition 1 is unclear. Can you explain? Does it provide better approximation for the field?
  3. Can you highlight the contribution of the paper at the end of introduction? It's not easy to follow the motivation and contribution of the paper.
Questions for authors

Same as weakness

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

Related papers

Discussion (0)

Sign in to comment.