acceptodds
Withdrawn from ICLR 2027

STRIDE: Training-Free Diversity Guidance via PCA-Directed Feature Perturbation in Single-Step Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Distilled one-step (T=1) or few-step (T4) diffusion models enable real-time image generation but often exhibit reduced sample diversity compared to their multi-step counterparts. In multi-step diffusion, diversity can be introduced through schedules, trajectories, or iterative optimization; however, these mechanisms are unavailable in the few-step or single-step setting, limiting the effectiveness of existing diversity-enhancing methods. A natural alternative is to perturb intermediate features, but naive feature perturbation is often ineffective, either yielding limited diversity gains or degrading generation quality. We argue that effective diversity injection in few-step models requires perturbations that respect the model's learned feature geometry. Based on this insight, we propose STRIDE, a training-free method that requires no per-image optimization or backward passes and operates in a single forward pass. STRIDE injects spatially coherent (pink) noise into intermediate transformer features, projected onto the principal components of the model's own activations, ensuring that perturbations lie on the learned feature manifold. This design enables controlled variation along meaningful directions in the representation space. Extensive experiments on FLUX.1-schnell and SD3.5 Turbo across COCO, DrawBench, PartiPrompts, and GenEval show that STRIDE consistently improves diversity while maintaining strong text alignment. In particular, STRIDE reduces intra-batch similarity with minimal impact on CLIP score, and Pareto-dominates existing training-free baselines on the diversity-fidelity frontier. These results highlight that, in the absence of iterative refinement, improving diversity in few-step and one-step diffusion depends not on increasing perturbation strength, but on aligning perturbations with the model's internal representation structure.

Primary area:
generative models
Keywords:
Diffusion models, few-step generation, Inference-time, Sample diversity

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

This paper proposes an inference-time method to improve the generation diversity of distilled single- and few-step diffusion models. It injects spatially coherent noise to the intermediate features by projecting the noise onto the principal components of the model’s activations. Experiments are conducted on the MM-DiT-based FLUX.1-schnell and SD3.5 Large Turbo models across diverse datasets, demonstrating that the proposed method successfully improves the diversity of generated images and largely preserves the text alignment capability, while requiring no additional training or model forward passes.

Critical strengths
  1. The paper investigates a practical setting in which distilled single- and few-step diffusion models can be guided at inference-time for improved generation diversity. The method is well-motivated, intuitive, and simple by design.

  2. Extensive empirical evaluations are presented across different models and datasets. Thorough ablative analyses and computational cost reports are also provided.

Critical weaknesses
  1. The authors describe the PCA subspace of spatial patches as the “learned feature manifold” or “local tangent space”, but what these subspaces really mean semantically remains unclear and is not supported by further evidence.

  2. Some related methods, such as [1-3], are not discussed. TPG is described as a close neighbor, but its perturbation mechanism is not included in the evaluation. SSG [4] is another close neighbor to STRIDE but is neither discussed nor evaluated.

  3. The choices of layers and steps at which to apply STRIDE appear highly heuristic, which limits the generalizability and applicability of the method.

  4. The computational cost does not appear favorable, despite the method requiring a single model forward pass. Section 4.2 and Table S11 show that the configurations used for the main experimental results may incur up to 150-200% latency increases. This would call the practical significance and the “single-forward-pass” efficiency claim into question.

[1] Park et al. “Understanding the Latent Space of Diffusion Models through the Lens of Riemannian Geometry”. NeurIPS 2023.

[2] Mo et al. “FreeControl: Training-Free Spatial Control of Any Text-to-Image Diffusion Model with Any Condition”. CVPR 2024.

[3] Kwon et al. “Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation”. ICML 2026.

[4] Zhang et al. “Guiding a Diffusion Model by Swapping Its Tokens”. CVPR 2026.

Questions for authors
  1. Could the authors evaluate the effect of token shuffling, as used in TPG, and token/channel swapping, as in SSG, as alternative perturbation mechanisms?

  2. How well does the proposed mechanism generalize to U-Net-based diffusion models?

  3. Why do Tables 1/S17 and S13/S10 report notably different diversity and CLIP results for the seemingly similar FLUX settings?

Reviewer_nZ72OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper propose STRIDE, a training-free method to improve diversity of one or few-step models. It's known that these few-step suffer from decreased diversity, and prior works mostly focus on multi-step diffusion setting. Harrington et al is the closest work where they iteratively optimize noise using gradient from diversity and human-preference objective, but it requires multiple optimization steps (forward + backward, around 80 iterations). The proposed method in this paper improve diversity by add noise to the hidden states of the intermediate layers of the diffusion model, with some tricks like adding pink noise (low-frequency noise) and project it to the subspace spanned by principle components of spatially grouped patches. This method is training-free, saves computational cost compared to other related methods since it does not involve iterative optimization steps. They show that the proposed method achieves Preto frontier in terms of CLIP similarity (prompt alignment) and diversity in FLUX.1-schnell (1 step) and SD3.5 Turbo (4 steps).

Critical strengths
  • The paper is easy to follow, and it cites the related papers extensively with detailed discussion.
  • The proposed method are well-motivated, and save the cost a lot.
Critical weaknesses
  • The main complaint about the paper is that it shows any qualitative results only in the teaser, and does not show any more results in the main paper, although qualitative results are critical in the papers about image generation. The final sample quality is the most important thing people care about. In the main paper, it only shows samples from one prompt, with only baselines, not other method competing with. I think it shows qualitative comparison in the main paper.
  • If I see the Figure S3 and S4 in Appendix, the samples looks not that great. For example, in Fig. S3 "A pink colored girraffe", STRIDE No-PCA and STRIDE results looks quite bad compared to other baselines, exhibiting distorted face (double-face issues in No-PCA case), and weird structure in STRIDE. The quality from "A shark in the desert" is also not great. If this is cherry-picked samples, I assume this method simply doesn't work. My read from the results in the paper is that the proposed method achieves more diversity by just making random artifacts in the images, which misses the assumption of methods image generation to preserve the quality.
Questions for authors
  • Could authors show the qualitative results from the proposed method and other methods in comparison, both the curated and uncrated results? For the uncurated results, I would encourage the authors to fix the seed to 0 for noise generation and use the prompts in first 10 prompts in COCO, DrawBench or PartiPrompts to make sure it's uncurated.
  • Could authors provide other human preference metrics, like PickScore v2, LAION Asthetic, ImageReward following the main table or GenEval table? The proposed method would be more convincing if these metrics are consistently comparable to other methods.
  • Could authors show the diversity-quality graph with HPS (or PickScore) vs. similarity score? Currently it's showing CLIP score, which measure prompt alignment, but the quality is usually different from prompt alignment. In my opinion, if the authors want to claim Pareto friontiers for diversity-quality tradeoff, the paper should show human-preference metrics rather than CLIP score.
  • Why are some metrics in Table 4 blank? For example, i.i.d and GI, NoiseOpt misses InBSim and CLIP metrics.
  • I think PCA in spatial group is less motivated. People usually do feature-PCA to all the tokens globally, and it works. Why should we avoid this? If one want to capture the most variant direction in feature space, why should they group locally? Doesn't using the global features provide consistent and robust estimate of the most sensitive direction? Wouldn't this just a way to add more randomness in the features, which may results in worse quality? Could the author provide Table 5 with HPS and PickScore results? (also, the bold on the numbers is misleading, since in SD3.5 case the CLIP score is the best in the baseline, but it use bold in P=2 case)

some minor questions

  • L128: In my understanding, the diversity increase in the cited papers (See Table 1 (conditional, recall) and Table 3 in perturbed-attention guidance, and Table 1 LPIPS in smoothed energy guidance paper)
  • L459: The sentence "We present STRIDE, a training-free method for improving output diversity in distilled few-step diffusion models through PCA-directed spatial feature perturbation." seems pasted here wrong.
Reviewer_pC5jOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes STRIDE, a training-free approach for improving sample diversity in distilled one-step and few-step diffusion models. STRIDE computes PCA over spatially grouped intermediate image features, projects spatially correlated noise into the resulting principal subspace, reweights the perturbation using singular values, and injects it into early transformer blocks. It requires neither per-image optimization nor backward passes.

Critical strengths
  • The paper discusses an important problem in distilled generative models which is the reduction of diversity in the few-step diffusion model.

  • STRIDE introduces PCA-directed feature perturbations without retraining, backpropagation, or an auxiliary model. This offers a practical way to adjust output diversity in existing generators.

Critical weaknesses
  • The claim that 'PCA subspace is a learned manifold' is too strong. This lacks necessary theoretical support and proof.

  • Visible artifacts in the qualitative results. STRIDE samples in Fig. 1(a) exhibit conspicuous grainy or grid-like textures in the sky, building facades, and road surfaces. Although these examples have more diverse compositions, they also indicate the significant cost of decreased generation quality.

  • The qualitative results presented in the main text are too few.

  • Despite requiring no additional model, STRIDE incurs substantial runtime overhead, taking approximately 2.7–3.0× the baseline inference time at its default settings (Table S11).

Questions for authors
  • Please check the inconsistent FLUX results in Tables S13 and S17: 0.510/26.64 versus 0.590/28.61 for InBSim/CLIP. Were these results obtained under the same experimental settings?

  • see weakneses.

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

Related papers

Discussion (0)

Sign in to comment.