Pith. sign in

REVIEW 3 major objections 27 references

Abrupt early swings in deep diffusion latents drive artifacts; a shared EMA detect-suppress step removes them without training.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 16:27 UTC pith:HZQ3N7F7

load-bearing objection Solid training-free artifact fix with real multi-backbone evidence; the causal story for the detector is the soft spot, not the results. the 3 major comments →

arxiv 2607.09753 v1 pith:HZQ3N7F7 submitted 2026-07-04 cs.CV cs.AIcs.LG

Unified Backbone Refinement for Diffusion Models via Internal-Latent Analysis

Haksoo Lim , Myeongjin Lee , Wonjoon Chang , Jaesik Choi This is my paper
classification cs.CV cs.AIcs.LG
keywords diffusion modelsU-NetTransformerartifact reductiontraining-free refinementinternal latentshallucinationsscore function
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Diffusion image generators often invent extra fingers, warped joints, or other hallucinations. This paper claims those defects track sudden early-stage fluctuations inside the deepest, already-denoised latents of the backbone—h-space in U-Nets and mid-to-deep self-attention maps in Transformers. The authors introduce DUNE, a training-free procedure that watches those latents with a fixed exponential moving average, flags the most deviant entries, and applies a light backbone-specific suppression only to the flagged locations during the first part of sampling. Because the same detect-suppress idea works across architectures and needs no extra training or memory, a reader who cares about reliable generation can improve fidelity and cut anatomical defects on existing models simply by intervening at the right place and time. An optional later phase still lets users tweak brightness and detail without undoing the early correction.

Core claim

Artifact-associated dynamics concentrate in early denoising stages and appear as abrupt deviations inside deep low-noise internal latents (U-Net h-space and mid-to-deep Transformer self-attention). Detecting those deviations with a shared EMA criterion on score-normalized latents and applying masked backbone-specific suppression improves FID and HADM while preserving semantic content across multiple high-resolution models.

What carries the argument

DUNE: a shared EMA detector on score-normalized deep latents that selects the top-percentile log-ratio deviations, followed by a gated suppress operator—channel-aware scaling for U-Net h-space, EMA blending for Transformer self-attention latents—applied only in the early detect phase.

Load-bearing premise

A fixed top-percentile log-ratio of score-normalized latents versus a fixed EMA truly isolates the residuals that cause artifacts rather than merely high-variance but harmless structure.

What would settle it

Replace the EMA detector mask with a density-matched random mask of the same size and apply identical suppression; if FID and HADM improve by the same amount or more, the claimed selective association fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Training-free refinement can raise both overall fidelity and anatomical correctness on U-Net and Transformer diffusion models at once.
  • Control should target deep low-noise latents early in sampling rather than shallow features or the final score output.
  • The same detect-suppress principle transfers from convolutional to attention backbones because both form noise-attenuated semantic spaces.
  • Optional late-phase scaling of skip and upsampling paths can adjust tone and detail without undoing the early correction.
  • Random suppression of the same fraction of deep entries degrades quality, so the gain is detector-selective rather than a generic regularizer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If early deep-latent spikes are causal, similar EMA monitors could flag failure modes in video or multi-modal diffusion without architecture-specific redesign.
  • The fixed percentile may still act partly as a soft regularizer; causal tests that deliberately inject spikes would separate correlation from correction.
  • Production samplers could expose a lightweight deep-latent diagnostic layer as a standard, model-agnostic quality gate.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper analyzes phase- and component-wise behavior of diffusion backbones and argues that abrupt early-stage fluctuations in deep low-noise latents (U-Net h-space; mid-to-deep Transformer self-attention) are strongly associated with visual artifacts. From this, it proposes DUNE, a training-free detect–suppress method: an EMA log-ratio detector on score-normalized deep latents (p=0.9, γ=0.7) selects entries, which are then corrected by backbone-specific operators (channel-aware scaling for U-Net; EMA blending for Transformers) during an early detect phase. Optional later-phase FreeU-style reweighting is excluded from quantitative tables. Theory (Props. 1–2, SNR theorems, Lemma 1) motivates deep intervention and bounds suppressor stability. Experiments on SDXL, LCM, Kandinsky 3, PixArt-Σ, and Hunyuan-DiT report modest FID gains, large HADM drops, high original-to-refined CLIP similarity, acceleration diagnostics (NAA/EAGR), layer/random-mask ablations, and a user study.

Significance. If the association and intervention claims hold, the work offers a practical, architecture-spanning, training-free refinement recipe and a useful operational principle—control deep, early, low-noise latents rather than score outputs or shallow branches. Strengths include multi-backbone coverage (U-Net and DiT-style models), external metrics (FID, HADM, CLIP similarity), user preference, intervention-location ablations (h-space vs skip/up; Transformer depth table), random-mask controls, and explicit separation of the optional detail phase from quantitative claims. The theoretical material usefully motivates why deep latents are quieter intervention points, even if it does not fully establish causal selection of artifact residuals. The contribution is incremental but relevant to training-free diffusion refinement and internal-latent analysis.

major comments (3)
  1. §4.1–4.2 and the central claim: the manuscript treats the fixed top-percentile EMA log-ratio Δ=log|z_t/z̄_{t+1}| (p=0.9, γ=0.7) as isolating artifact-causing residuals. Fig. 4/9 show spatial alignment with acceleration spikes and Table 5 shows random h-space suppression degrades FID, which is necessary but not sufficient for causality. The detector may still flag high-variance structure that co-occurs with hallucinations; shrinking it can improve HADM without proving the claimed pathway. Please strengthen causal identification (e.g., controlled injection of known artifacts, counterfactual mask swaps, or correlation of mask mass with HADM boxes beyond acceleration alignment) or restate the claim as association plus effective regularizer rather than causal residual correction.
  2. Table 1 and Table 3: FID improvements are very small (e.g., SDXL 18.86→18.75) while HADM drops are large, and κ is chosen per backbone after a grid search that optimizes FID (App. 7.1). This leaves open whether gains are robust or partly hyperparameter-tuned regularization. Report sensitivity of HADM (not only FID) to (p,κ), a single fixed-κ protocol across backbones where possible, and confidence intervals or multi-seed variance for the ~0.1 FID deltas so the fidelity claim is not overstated relative to the hallucination claim.
  3. §3–4 and App. 16: Props. 1–2 and the SNR theorems justify that deep latents are noise-attenuated and that a given mask’s suppression can raise SNR / bound score perturbation (Corollary 1, Lemma 1). They do not show that the EMA detector selects the noise-dominant subset η_t > ρ_t required by the theorems. Either provide empirical estimates of η_t, ρ_t, and ε_t on detector masks versus random masks, or clearly separate “intervention location is stable” from “this detector finds the right entries.”

Circularity Check

0 steps flagged

No significant circularity: detect–suppress and SNR claims are independent of the external quality metrics they are evaluated against.

full rationale

The derivation chain is self-contained and non-circular. Component-wise phase analysis (§2) and noise-contraction propositions for CNN downsampling and self-attention (Props. 1–2 / 3–4) motivate deep latents as intervention points without defining those points in terms of FID/HADM. Detection (§4.1) uses an EMA log-ratio Δ = log |z_t / z̄_{t+1}| on score-normalized latents with fixed γ = 0.7; the mask is a percentile threshold, not a quantity fitted to the reported metrics. Suppression operators and the SNR theorems (Thm. 3/5, Cor. 1, Lem. 1/2) are conditional implications: if the mask concentrates noise relative to signal (or EMA error is small), post-suppression SNR rises and score perturbation is Lipschitz-bounded. These do not tautologically equal the quality claims. Hyperparameters p and κ are chosen by FID sweeps on two backbones then frozen (App. 7.1, Table 3); that is ordinary selection, not a fitted input renamed as prediction of the same quantity. No load-bearing uniqueness theorem or ansatz is imported solely from overlapping authors; external baselines and metrics (FID, HADM-L, CLIP similarity) are independent. Random-mask ablation (Table 5) further shows the detector is not vacuous. Score 0 is therefore warranted.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The method rests on standard diffusion score training, two noise-averaging propositions, an SNR signal/noise split of scaled latents, and several hand-chosen control constants (percentile, EMA rate, per-backbone κ, depth, phase boundary). No new physical entity is postulated; DUNE is an algorithmic construct. Empirical success depends on the detector truly marking artifact-linked residuals.

free parameters (5)
  • detection percentile p
    Fixed to 0.9 after FID heatmaps on LCM and Kandinsky 3; selects top 10% log-ratio outliers for the mask.
  • suppression strength κ (per backbone)
    Hand-set per model (e.g. 0.3 SDXL/LCM, 0.5 Kandinsky 3, 0.95 PixArt-Σ, 0.85 Hunyuan-DiT) after sweeps; directly scales correction strength.
  • EMA coefficient γ
    Fixed at 0.7 for all experiments without per-backbone tuning; defines the reference trajectory for Δ.
  • detect-phase boundary (≈0.6T) and warm-up steps
    Early window and short warm-up (3 steps long samplers / 1 step LCM) chosen by design, not derived.
  • Transformer target depth (3/4 stack)
    Chosen via FID ablation on PixArt-Σ; default intervention layer for all Transformer runs.
axioms (5)
  • domain assumption DDPM/score-matching formulation: ϵ_θ estimates noise/score of the reverse process.
    Background §2; entire detect–suppress pipeline assumes standard score-parameterized sampling.
  • standard math Downsampling convolutions scale isotropic noise by the kernel Frobenius norm (Prop. 1 / App. Prop. 3).
    Used to argue h-space is lower-noise and safer to edit than shallow skips/ups.
  • standard math Self-attention averages token noise by √(Σ w_j²) for fixed weights (Prop. 2).
    Justifies mid-to-deep attention latents as Transformer analogue of h-space.
  • ad hoc to paper Scaled latent z_t = s_t + n_t with slowly varying s_t and zero-mean n_t (Assumptions 1–2).
    SNR analysis §4.3; needed for theorems claiming post-suppression SNR gain.
  • domain assumption Artifact regions exhibit abnormal temporal score acceleration, so latent log-ratio outliers track them.
    Motivated by ASCED/hallucination literature and Fig. 4; load-bearing for the detector.
invented entities (2)
  • DUNE detect–suppress operator (shared EMA criterion + backbone-specific S_bb) no independent evidence
    purpose: Training-free correction of detector-selected deep latents during early denoising.
    Core algorithmic object; validated only by the paper’s experiments, not an external physical entity.
  • NAA and EAGR acceleration diagnostics no independent evidence
    purpose: Quantify that suppression reduces excess acceleration in detector masks vs random masks.
    Paper-defined metrics supporting mechanism; not independently standardized.

pith-pipeline@v1.1.0-grok45 · 30532 in / 3264 out tokens · 35122 ms · 2026-07-14T16:27:24.770303+00:00 · methodology

0 comments
read the original abstract

Diffusion models have achieved remarkable success across diverse domains, with performance closely related to the denoising backbones that parameterize the score function. In this paper, we present a systematic, phase-aware analysis of diffusion components and show that abrupt, early-stage fluctuations in deep latents are strongly associated with artifacts. Guided by these findings, we introduce DUNE (Diffusion Unified Network refiNEr), a training-free refinement framework that detects abrupt deviations in deep low-noise internal latents using a shared EMA-based criterion, and applies backbone-specific suppression to the detector-selected entries. Although derived from U-Net, the same detect-suppress principle extends naturally to Transformer-based diffusion models by acting on the latents of deep self-attention blocks. Extensive experiments across multiple backbones indicate that DUNE improves fidelity while reducing hallucinations, offering new insight into where and when diffusion backbones should be controlled.

Figures

Figures reproduced from arXiv: 2607.09753 by Haksoo Lim, Jaesik Choi, Myeongjin Lee, Wonjoon Chang.

Figure 1
Figure 1. Figure 1: We propose DUNE, a method that freely improves image quality, meaning no additional training or increase of memory. DUNE mitigates visual artifact while preserving semantic consistency. Abstract. Diffusion models have achieved remarkable success across di￾verse domains, with performance closely related to the denoising back￾bones that parameterize the score function. In this paper, we present a systematic,… view at source ↗
Figure 2
Figure 2. Figure 2: Effects of scaling U-Net components across denoising steps. Upsampling blocks affect brightness and tone, while skip connections enhance edge details. In early phases, however, scaling primarily leads to semantic changes. operating on deep self-attention latents that satisfy the same intervention crite￾rion as U-Net h-space: semantic integration with reduced stochastic fluctuation. These improvements incur… view at source ↗
Figure 3
Figure 3. Figure 3: Suppression Results with Different Components in U-net Controlling. The change of the score function according to denoising time step is called acceleration [4]. This inherent denoising capability early in the reverse trajectory clarifies where to intervene. However, directly correcting these anomalies in shallow lay￾ers poses challenges, as simultaneous adjustments of image content and added noise are nee… view at source ↗
Figure 4
Figure 4. Figure 4: We visualize how detector-selected regions align with abrupt score-acceleration patterns. The plot shows the averaged score function across diffusion steps, with “Tar￾get” and “Non-target” representing the inner and outer regions of the red bounding contour, respectively. softmax. As the attention distribution becomes more diffuse (higher entropy; e.g., larger temperature), P j w 2 ij ↓ toward 1/N and the … view at source ↗
Figure 5
Figure 5. Figure 5: Overall framework of DUNE. The h-space is directly matched with self￾attention latent in Transformer-based diffusion models. into two phases: the detect phase (1.0T–0.6T) and the detail phase (0.6T–0.0T). This division reflects that refinements in early and later stages predominantly affect global structure and local details, respectively. Note that the detect phase constitutes the core of DUNE and is suff… view at source ↗
Figure 6
Figure 6. Figure 6: An experimental result on high-resolution text to image generation. We give remained Hunyuan-DiT figures in Appendix 15. distorted joints). For HADM analysis, we generate 5,000 images from human￾related prompts by using the human-related annotation files from official COCO resource, following HADM paper [25]. Lastly, since we observed existing refine￾ment methods fail to preserve consistent semantic (e.g.,… view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison of ASCED and DUNE on PixArt-Σ. FreeU ASCED PAG DUNE 0 2 4 6 8 10 12 14 16 Sampling time (s) 9.55s9.59s 9.55s 9.55s 9.55s 13.37s 14.38s 9.71s Sampling Time Comparison Original Other Methods DUNE (Ours) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sampling time comparison on SDXL. 5.2 Enhancing Quality with DUNE We first verify that DUNE suppresses detector-selected unstable regions. Fig￾ure 9 (extending [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Before/after score acceleration. Target denotes the detector-selected mask and non-target denotes its complement. Curves omit the final 10 denoising steps and are normalized only for visualization. settings for Kandinsky 3, PixArt-\Sigma , and Hunyuan-DiT, while PAG lacks offi￾cial/default settings for Kandinsky 3 and Hunyuan-DiT. DUNE also maintains high original-to-refined CLIP similarity, indicating low… view at source ↗
Figure 10
Figure 10. Figure 10: Heatmap of FID across different percentages and kappas. 7 Hyperparameter Setting and Miscellaneous environments 7.1 Sensitivity Analysis on Percentiles and Kappas In this section, we investigate the sensitivity of DUNE with respect to its two hyperparameters: the detection percentile p and the suppression scaling factor κ (see Section 4.1). Specifically, we generate 1000 images using LCM and Kandin￾sky 3 … view at source ↗
Figure 11
Figure 11. Figure 11: Self-attention latent maps over depth and time for PixArt-Σ and Hunyuan-DiT. For each model, we visualize token maps obtained by channel-averaging the post-attention activations at four depths of the Transformer stack—0/4, 1/4, 2/4, and 3/4 of the total attention depth—sampled every 10 denoising steps. The terminal layers are omitted from visualization because they primarily output the directional noise (… view at source ↗
Figure 12
Figure 12. Figure 12: Investigation on applying DUNE to non-hallucinated images. dle layers, while deeper layers remain semantically sharper. We further validate the practical target-layer choice on PixArt-Σ by applying DUNE at four rep￾resentative depths (1/4, 1/2, 3/4, and the last layer) while keeping all other settings fixed. As shown in [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Representative examples of detected samples in SDXL through HADM. in Figure 12a, when no artifacts are present, the detected areas (marked in red dots) do not concentrate on any specific region, but instead spread evenly throughout the latent space. Therefore, to further test the robustness of DUNE, we applied our suppres￾sion method randomly across the h-space, using the same correction scaling factor (κ… view at source ↗
Figure 14
Figure 14. Figure 14: Visual comparison between latest unsupervised refinement methods. 10 Metrics Also, we use HADM-L [25] to measure the efficacy of DUNE in mitigating lo￾calized anatomical distortions. By employing a ViTDet-based object detection architecture, HADM-L explicitly localizes and classifies fine-grained structural defects. Specifically, it predicts bounding boxes around implausible body parts (e.g., six-fingered… view at source ↗
Figure 15
Figure 15. Figure 15: Sampling results of different correction interventions across various time steps. As depicted in [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative comparison between FreeU and DUNE (ours). we emphasize that FreeU exclusively focuses on reducing skip connections to remove noise and increasing upsampling blocks to highlight original semantics. Ironically, these modifications often eliminate essential details from the origi￾nal images and result in overly saturated outputs. Conversely, DUNE maintains original semantic content while enabling… view at source ↗
Figure 17
Figure 17. Figure 17: Visualization of controlling different components of the U-Net architecture. lier, we identified that skip connections primarily carry high-frequency informa￾tion (e.g., edges), while upsampling blocks handle lower-frequency content (e.g., overall color or contents). Intuitively, downsampling blocks integrate comprehen￾sive information by combining multiple components. Figure 17a illustrates that amplifyi… view at source ↗
Figure 18
Figure 18. Figure 18: Ablation study depth-wise control of each component. adjustments typically remain visually acceptable. This aligns with the intuition that manipulating downsampling blocks simultaneously influences skip connec￾tions and upsampling blocks, thus requiring more careful handling. Moreover, we investigate alternative scaling approaches for skip connections and upsampling blocks. Contrary to FreeU, which conven… view at source ↗
Figure 19
Figure 19. Figure 19: Comparison between visual results and user preference scores in unconditional generation tasks. Training-free refinement methods differ in where they intervene. Score-level approaches—TAG [6] (tangential amplification of sampling increments), Dy￾namic Guidance [24] (selective score sharpening), and ASCED [4] (corrective noise injection)—modify the output or trajectory without explicitly accessing internal… view at source ↗
Figure 20
Figure 20. Figure 20: Qualitative analysis on Hunyuan-DiT. training datasets for these models, we omitted their formal quantitative eval￾uations in the main paper. Nonetheless, qualitative improvements are strongly supported by user preference data. As summarized in [PITH_FULL_IMAGE:figures/full_fig_p031_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Figures for User Survey. SDXL is used to generate these images [PITH_FULL_IMAGE:figures/full_fig_p043_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Figures for User Survey. Kandinsky 3 is used to generate these images [PITH_FULL_IMAGE:figures/full_fig_p044_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Figures for User Survey. DDPM is used for CelebA-HQ and Church, while VE is used for FFHQ and Bedroom [PITH_FULL_IMAGE:figures/full_fig_p045_23.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 14 linked inside Pith

  1. [1]

    Ahn, D., Cho, H., Min, J., Jang, W., Kim, J., Kim, S., Park, H.H., Jin, K.H., Kim, S.: Self-rectifying diffusion sampling with perturbed-attention guidance (2025), https://arxiv.org/abs/2403.17377

    Pith/arXiv arXiv 2025

  2. [2]

    Advances in Neural Information Pro- cessing Systems37, 134614–134644 (2024)

    Aithal, S.K., Maini, P., Lipton, Z., Kolter, J.Z.: Understanding hallucinations in diffusion models through mode interpolation. Advances in Neural Information Pro- cessing Systems37, 134614–134644 (2024)

    2024

  3. [3]

    arXiv preprint arXiv:2410.21061 (2024)

    Arkhipkin, V., Vasilev, V., Filatov, A., Pavlov, I., Agafonova, J., Gerasimenko, N., Averchenkova, A., Mironova, E., Bukashkin, A., Kulikov, K., et al.: Kandinsky 3: Text-to-image synthesis for multifunctional generative framework. arXiv preprint arXiv:2410.21061 (2024)

    Pith/arXiv arXiv 2024

  4. [4]

    arXiv preprint arXiv:2503.16218 (2025)

    Cao, Y., Zhao, Z., Patras, I., Gong, S.: Temporal score analysis for understanding and correcting diffusion artifacts. arXiv preprint arXiv:2503.16218 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Chen, J., Ge, C., Xie, E., Wu, Y., Yao, L., Ren, X., Wang, Z., Luo, P., Lu, H., Li, Z.: Pixart-σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation (2024),https://arxiv.org/abs/2403.04692

    Pith/arXiv arXiv 2024

  6. [6]

    Cho, H., Ahn, D., Hong, S., Kim, J.E., Kim, S., Jin, K.H.: TAG: Tangential am- plifying guidance for hallucination-resistant diffusion sampling (2025)

    2025

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Choi, J., Lee, J., Shin, C., Kim, S., Kim, H., Yoon, S.: Perception prioritized training of diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11472–11481 (2022)

    2022

  8. [8]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  9. [9]

    In: WACV (2024)

    Jeong, J., et al.: Training-free content injection usingh-space in diffusion models. In: WACV (2024)

    2024

  10. [10]

    arXiv preprint arXiv:2210.10960 (2022)

    Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960 (2022)

    Pith/arXiv arXiv 2022

  11. [11]

    Neurocomputing 479, 47–59 (2022)

    Li, H., Yang, Y., Chang, M., Chen, S., Feng, H., Xu, Z., Li, Q., Chen, Y.: Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing 479, 47–59 (2022)

    2022

  12. [12]

    arXiv e-prints pp

    Li, S., Hu, T., Shahbaz Khan, F., Li, L., Yang, S., Wang, Y., Cheng, M.M., Yang, J.: Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv e-prints pp. arXiv–2312 (2023)

    2023

  13. [13]

    Li, Z., Zhang, J., Lin, Q., Xiong, J., Long, Y., Deng, X., Zhang, Y., Liu, X., Huang, M., Xiao, Z., Chen, D., He, J., Li, J., Li, W., Zhang, C., Quan, R., Lu, J., Huang, J., Yuan, X., Zheng, X., Li, Y., Zhang, J., Zhang, C., Chen, M., Liu, J., Fang, Z., Wang, W., Xue, J., Tao, Y., Zhu, J., Liu, K., Lin, S., Sun, Y., Li, Y., Wang, D., Chen, M., Hu, Z., Xia...

    Pith/arXiv arXiv 2024

  14. [14]

    CoRRabs/1405.0312(2014),http://arxiv.org/abs/1405.0312

    Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. CoRRabs/1405.0312(2014),http://arxiv.org/abs/1405.0312

    Pith/arXiv arXiv 2014

  15. [15]

    Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthe- sizing high-resolution images with few-step inference (2023),https://arxiv.org/ abs/2310.04378

    Pith/arXiv arXiv 2023

  16. [16]

    arXiv preprint arXiv:2502.16872 (2025) DUNE: Unified Backbone Refinement for Diffusion Models 17

    Ozbulak, U., et al.: Mitigating hallucinations in diffusion models through adaptive attention modulation. arXiv preprint arXiv:2502.16872 (2025) DUNE: Unified Backbone Refinement for Diffusion Models 17

    Pith/arXiv arXiv 2025

  17. [17]

    Peebles, W., Xie, S.: Scalable diffusion models with transformers (2023),https: //arxiv.org/abs/2212.09748

    Pith/arXiv arXiv 2023

  18. [18]

    arXiv preprint arXiv:2307.01952 (2023)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    Pith/arXiv arXiv 2023

  19. [19]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  20. [20]

    In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18. pp. 234–241. Springer (2015)

    2015

  21. [21]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Si, C., Huang, Z., Jiang, Y., Liu, Z.: Freeu: Free lunch in diffusion u-net. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 4733–4743 (2024)

    2024

  22. [22]

    Advances in neural information processing systems32(2019)

    Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems32(2019)

    2019

  23. [23]

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations (2021),https: //arxiv.org/abs/2011.13456

    Pith/arXiv arXiv 2021

  24. [24]

    arXiv preprint arXiv:2510.05356 (2025)

    Triaridis, K., Graikos, A., Chatziagapi, A., Chrysos, G.G., Samaras, D.: Mit- igating diffusion model hallucinations with dynamic guidance. arXiv preprint arXiv:2510.05356 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    Wang, K., Zhang, L., Zhang, J.: Detecting human artifacts from text-to-image models (2025),https://arxiv.org/abs/2411.13842

    Pith/arXiv arXiv 2025

  26. [26]

    Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., Yang,M.H.:Diffusionmodels:Acomprehensivesurveyofmethodsandapplications (2023)

    2023

  27. [27]

    A person jumping with arms raised

    Yang, X., Zhou, D., Feng, J., Wang, X.: Diffusion probabilistic model made slim. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 22552–22562 (2023) DUNE: Unified Backbone Refinement for Diffusion Models 1 Fig. 10:Heatmap of FID across different percentages and kappas. 7 Hyperparameter Setting and Miscellaneous en...

    2023