acceptodds
Withdrawn from ICLR 2027

Image Quality Emerges from the Gaussianity of Diffusion Inversion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion models learn to generate natural images by transforming Gaussian noise into samples from a complex image distribution. While previous studies mainly exploit this forward generation process, we ask whether the inverse trajectory can reveal information about the observed image itself. We discover that diffusion inversion provides a measurable signal of observation–prior compatibility: high-quality images that better agree with the learned image distribution tend to recover terminal latents with Gaussian-like statistics, whereas low-quality images induce increasingly atypical latent distributions. Based on this finding, we introduce Diffusion Inversion Gaussianity Score (DIGS), a training-free and quality-label-free measure that evaluates image quality by quantifying the deviation of inversion latents from the intrinsic Gaussian prior. Beyond marginal statistics, we further characterize channel-wise and spatial deviations to obtain DIGS. Extensive experiments demonstrate that diffusion inversion statistics correlate strongly with human perceptual judgments across authentic and synthetic distortions, while requiring no quality annotations, reference images, or additional training. Our results reveal diffusion inversion as a new measurement space for probing how observations interact with learned generative priors.

Primary area:
generative models
Keywords:
Non-reference quality assessment, diffusion inversion, Gaussian Distribution

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

The paper proposes a training-free IQA metric based on diffusion inversion. The central idea is that high-quality images invert to latents that more closely resemble Gaussian noise. The proposed DIGS demonstrates a meaningful correlation with human perceptual judgments, although CLIP-IQA consistently outperforms the two proposed variants.

Critical strengths
  1. New empirical insight: The Gaussianity of inverted diffusion latents correlates with image quality and potentially reflects compatibility between observations and learned generative priors.

  2. The authors measure deviations in the skewness and kurtosis of standardized inversion latents from their Gaussian reference values, and consider other statistical aspects to formulate a training-free IQA metric.

Critical weaknesses

I appreciate the authors’ investigation and insight of the statistical properties of diffusion inversion and associate it to the image quality. However, as an IQA metric, the proposed DIGS (and DIGS*) has several limitations:

  1. Inferior performance. Other methods, such as CLIP-IQA, consistently outperform both proposed variants on SROCC and PLCC across all three datasets. CLIP-IQA also uses a pretrained prior and does not require image-quality labels or additional training. This questions the effectiveness of DIGS, particularly given the computational overhead associated with diffusion inversion.

  2. Sensitive to diffusion backbones and inversion setting. The authors report in Table 4 that for different backbones and operators, the SROCC and PLCC can differ a lot. This dependence raises concerns about the robustness and generalizability.

Questions for authors

Why is the performance of DIGS related to the inversion operators? The authors say that the choice of inversion trajectory affects the quality-related statistics extracted from terminal latents. How is this correlation explained?

In Figure 6, how are the distributional divergences computed for the KonIQ-10k dataset? This should be clarified.

Reviewer_1X9zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores the information revealed by the inverse trajectory of diffusion models, moving beyond the traditional forward generation process. The authors discover that diffusion inversion serves as a measurable indicator of observation–prior compatibility: high-quality images adhering well to the learned prior recover terminal latents with Gaussian-like statistics, whereas low-quality images result in increasingly atypical latent distributions. Building on this insight, the paper introduces the Diffusion Inversion Gaussianity Score (DIGS), a training-free and quality-label-free metric that assesses image quality by measuring the deviation of inversion latents from an intrinsic Gaussian prior. Furthermore, the authors propose DIGS by incorporating channel-wise and spatial deviations. Extensive experiments demonstrate that these diffusion inversion statistics strongly correlate with human perceptual judgments across both authentic and synthetic distortions, without requiring quality annotations, reference images, or additional model training.

Critical strengths
  1. The paper introduces a fresh and insightful conceptual shift by investigating how diffusion inversion trajectories can be leveraged to probe the interaction between observations and learned generative priors, rather than merely focusing on image generation.
  2. The proposed DIGS and DIGS metrics operate without the need for human quality annotations, reference images, or any auxiliary training, making them highly efficient, practical, and generalizable for assessing image quality in diverse scenarios.
  3. The authors provide comprehensive experimental evaluations showing that diffusion inversion statistics correlate robustly with human perceptual judgments across a wide range of both authentic and synthetic distortions, confirming the effectiveness and reliability of the proposed measurement space.
Critical weaknesses
  1. This method has a fatal flaw: despite being a training-free approach, its performance falls short of that of CLIP-IQA—a method that requires training—rendering it of little practical value.
  2. The method relies on a Gaussian prior but lacks validation for other types of priors, such as the Laplace distribution.
  3. There is a lack of sufficient validation; existing validation data is inadequate for verifying the correctness of the training-free method across the full data distribution.
  4. The proposed method lacks a comparison with recent state-of-the-art approaches, such as DP-IQA,LGDM, PFD-IQA
Questions for authors

see weakness

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

The paper proposes Diffusion Inversion Gaussianity Score (DIGS), a training-free, no-reference image quality measure based on the observation that higher-quality images tend to produce more Gaussian-like inversion latents. DIGS measures skewness and kurtosis deviations, while DIGS incorporates channel-wise and spatial statistics. Experiments demonstrate correlations with human quality judgments, although the explanation linking Gaussianity to perceptual quality remains less established than the empirical association.

Critical strengths
  1. The paper investigates diffusion inversion statistics as a quality signal, rather than training a quality predictor on diffusion features. This offers an appealing alternative use of pretrained generative models.
  2. The method uses a frozen diffusion model and fixed statistical computations, without fitting an additional quality regressor or requiring reference images during inference.
  3. DIGS combines deviations in skewness and kurtosis, making its computation transparent. Component ablations show that retaining both statistics performs better than using either alone.
  4. Evaluation covers KonIQ-10k, HRIQ, and SPAQ, complemented by controlled degradation experiments involving blur, compression, noise, and jitter. This provides broader support than a single-benchmark evaluation.
  5. The paper examines alternative statistical measures, inversion operators, diffusion backbones, and inversion-step counts, helping characterize when the proposed signal is effective.
Critical weaknesses
  1. Equations (2)–(4) describe a conditional latent distribution, whereas deterministic inversion produces a single latent for a fixed encoded input. The implemented score instead measures empirical coefficient moments; matching skewness and kurtosis alone does not establish Gaussianity.
  2. Performance varies substantially across inversion operators and backbones—for example, SROCC drops from 0.558 with DDIM to 0.295 with Euler. Additional controls are needed to distinguish perceptual-quality information from inversion error and content-dependent effects.
  3. DIGS rewards spatial energy dependence, whose connection to Gaussian-prior compatibility is unclear. Moreover, Table 3 mainly evaluates additive combinations rather than removing individual components from the exact final log-ratio formulation.
  4. CLIP-IQA outperforms both variants on all three datasets, with a substantial KonIQ-10k SROCC gap of 0.685 versus 0.558 for DIGS The extension also provides negligible improvement on HRIQ, increasing SROCC only from 0.612 to 0.613.
  5. The default 20-step configuration is motivated by its peak correlation on a KonIQ-10k subset. The manuscript should clarify whether this selection used a separate development set, distinguishing label-free scoring from label-free hyperparameter selection.
Questions for authors
  1. How should q(z^T∣x)q(\hat z_T\mid x) be interpreted under deterministic inversion, and what controlled evidence separates image-quality effects from inversion error and image-content effects?

  2. Why does greater spatial energy dependence increase DIGS, and can you provide leave-one-component-out ablations of the exact final formula?

  3. Were inversion steps, backbone, and statistical settings selected using an independent development set, and was the identical scoring procedure then applied unchanged across all evaluation datasets?

  4. Also address the weakness

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

Related papers

Discussion (0)

Sign in to comment.