Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Identity-First Selection Boosts Subject Fidelity but Shrinks Diversity

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 · glm-5.2

2026-07-09 18:16 UTC pith:OVSICJ42

load-bearing objection DCAL shows a real identity-diversity trade-off, but the headline comparison is confounded by best-of-K sampling, and the titular SPaRa module has no training-side validation. the 4 major comments →

arxiv 2607.07173 v1 pith:OVSICJ42 submitted 2026-07-08 cs.CV

Stage-Aware Adaptation and Distribution Calibration for Subject-Driven Personalized Text-to-Image Generation

Wenyan Xu , Alizer Wong This is my paper
classification cs.CV
keywords text-to-image diffusionsubject-driven personalizationlow-rank adaptationcandidate selectionidentity-diversity trade-offdiffusion denoising stages
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.

This paper argues that personalizing text-to-image diffusion models is a multi-objective trade-off among identity consistency, text alignment, and sample diversity, and that current methods distort this trade-off in two complementary ways. On the training side, uniform low-rank adapter strength across all diffusion denoising timesteps fails to account for the fact that different stages of denoising handle different aspects of image formation—global structure versus fine texture. On the inference side, selecting the best candidate image primarily by identity similarity mathematically forces the selected samples into a shrinking ball around the reference image in feature space, compressing visual diversity. The paper formalizes both problems with bounds and propositions, then proposes SPaRa (stage-aware adapter scaling that varies strength by timestep) and DCAL (a candidate selection rule that weights identity, text alignment, and diversity jointly). The completed experiments validate only DCAL applied to an existing LoRA checkpoint: it improves identity and text metrics but measurably decreases pairwise diversity, confirming the predicted trade-off.

Core claim

The central discovery is a quantified trade-off: when candidate selection for personalized image generation is driven by identity similarity, it provably restricts selected features to a Euclidean ball of radius sqrt(2(1−η)) around the reference center, where η is the identity threshold. Empirically, the DCAL selector confirms this by improving identity fidelity (CLIP-I, DINO-I, 1-LPIPS) and text alignment (CLIP-T) while simultaneously degrading CLIP diversity, DINO diversity, and pairwise LPIPS. This demonstrates that identity metrics alone are a misleading evaluation axis for personalization.

What carries the argument

The paper introduces two mechanisms. SPaRa replaces the constant LoRA scaling factor α with a timestep-dependent function α(t), redistributing perturbation budget across high-noise (structure-forming) and low-noise (texture-refining) denoising stages without changing the rank or shape of the adapter matrices. DCAL replaces single-metric candidate selection with a weighted score S = λ_I·I + λ_T·T + λ_D·D, combining identity consistency (averaged CLIP and DINO similarity to reference center), text alignment (CLIP image-text similarity), and a diversity reward that penalizes candidates too close to already-selected images. The theoretical machinery includes a perturbation bound showing |α(t)|/r

Load-bearing premise

The necessity of stage-aware scaling (SPaRa) depends on the assumption that different diffusion timesteps have non-overlapping acceptable perturbation-budget intervals, meaning no single uniform scaling factor can satisfy all stages simultaneously. These intervals are defined abstractly but never empirically measured, so the core motivation for the training-side contribution rests on an unverified premise.

What would settle it

If the acceptable perturbation-budget intervals [L_{ℓ,t}, U_{ℓ,t}] across timesteps are found to have a nonempty intersection in practice, then Proposition 4.4 no longer necessitates stage-aware scaling, and SPaRa reduces to conventional LoRA with no theoretical advantage. Additionally, if DCAL's diversity loss were shown to vanish with larger candidate pools or different base models, the identity-diversity trade-off would be an artifact of insufficient sampling rather than a structural property of identity-biased selection.

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

If this is right

  • Evaluation protocols for personalized generation should adopt multi-axis reporting (identity + text + diversity) as standard, since identity-only metrics can mask distributional collapse.
  • The identity-threshold radius bound suggests that any selection method using a hard identity cutoff will compress diversity, which applies beyond diffusion to any retrieval or reranking pipeline that filters by similarity to a reference.
  • Stage-aware adapter scaling (SPaRa) is proposed but only validated at inference time; if the training-side version works, it implies that adaptation capacity should be scheduled along the denoising chain like a curriculum.
  • The diversity penalty term in DCAL (Eq. 16) is a greedy anti-redundancy mechanism that could be swapped for submodular optimization or determinantal point processes for more principled diversity maximization.

Where Pith is reading between the lines

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

  • If the acceptable perturbation-budget intervals for different timesteps do overlap in practice (which the paper never empirically verifies), then SPaRa's necessity collapses—uniform scaling would be sufficient. The entire motivation for stage-aware training-side adaptation rests on an unverified assumption about interval disjointness.
  • The trade-off between identity and diversity may be fundamental to few-shot personalization rather than an artifact of the selection rule: with only a few reference images, any method that faithfully reproduces identity is implicitly biased toward the reference neighborhood.
  • The DCAL diversity cost could potentially be mitigated by generating candidates from multiple adapter checkpoints or varying guidance scales, effectively enlarging the candidate pool before selection rather than reweighting within a fixed pool.

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

4 major / 5 minor

Summary. This paper studies the trade-off among identity consistency, text alignment, and generation diversity in subject-driven personalized text-to-image generation. It proposes two modules: SPaRa, a training-side stage-aware low-rank adaptation mechanism that introduces timestep-dependent scaling α(t) into the LoRA adapter, and DCAL, an inference-side distribution-calibrated candidate selection rule that combines identity, text, and diversity scores. The theoretical analysis (Section 4) derives bounds on low-rank perturbation magnitudes and candidate feature radii. The experimental evaluation (Section 6) reports Full30 results for DCAL on a LoRA baseline, showing improvements in identity and text metrics at the cost of diversity. However, the paper explicitly states that training-side SPaRa and the combined SPaRa-DCAL framework lack complete Full30 experimental validation, and the PaRa baseline is missing a same-protocol comparison.

Significance. The paper addresses a genuine and important problem: the tension between identity preservation, text alignment, and diversity in personalized diffusion models. The framing of personalization as a multi-objective trade-off rather than a single identity-maximization problem is well-motivated. The theoretical analysis, while straightforward, is correct and provides useful formal grounding for the design choices. The DCAL selection rule is a reasonable inference-time mechanism. The paper is commendably transparent about its experimental limitations—it explicitly lists missing ablations and incomplete results in Appendix B and Table 1, which is good scientific practice. However, the significance is substantially undermined by the gap between what is claimed in the title/abstract (SPaRa, SPaRa-DCAL) and what is actually validated (DCAL on LoRA only).

major comments (4)
  1. Title-claim mismatch: The title and abstract prominently feature SPaRa and SPaRa-DCAL, but the paper itself admits (Table 1, Section 6.2, Section 7) that 'Strict training-side Full30 missing; no numerical claim reported' for SPaRa, and 'Same-protocol Full30 missing; no numerical claim reported' for SPaRa-DCAL. The only completed Full30 experiment is DCAL on a LoRA baseline (Table 2). This is a fundamental mismatch between the paper's framing and its evidence base. The title should be revised to reflect the actual validated contribution, or the missing experiments must be completed.
  2. Missing best-of-K control (Table 2, Section 6.2): The headline DCAL result compares K=4 DCAL selection against a K=1 LoRA baseline. This conflates two effects: (1) the statistical benefit of drawing 4 samples and keeping the best (a mechanical best-of-K effect), and (2) the specific contribution of DCAL's weighted selection rule. The paper does not include a K=4 random-selection or K=4 identity-only-selection baseline on Full30. Without this control, one cannot determine whether DCAL's selection rule adds value beyond simply sampling more candidates. The Heldout9 sensitivity table (Table 4) varies K, CFG, and weights but never isolates the pure K effect with a trivial selection rule. This is the most critical missing ablation for the paper's central empirical claim.
  3. Unverified necessity assumption for SPaRa (Section 4.3, Proposition 4.4): The theoretical motivation for stage-aware scaling rests on the assumption that acceptable perturbation-budget intervals [L_{ℓ,t}, U_{ℓ,t}] for different timesteps do not overlap. The paper states: 'estimating such intervals would require additional sensitivity experiments.' Since these intervals are never empirically estimated, the necessity of SPaRa is motivated by an unverified assumption about capacity mismatch. The stage-aware inference scaling experiment (Table 2) does not validate this, as it only changes α(t) at sampling time on a LoRA checkpoint and does not show stable identity improvements. The theoretical analysis is correct but trivially derived from spectral norm submultiplicativity, and the key premise remains empirically ungrounded.
  4. Missing PaRa baseline (Section 6.5, Table 5): SPaRa is motivated by the parameter rank-reduction perspective of PaRa, but the available PaRa data contain only a five-subject rank sweep (Table 5) that is 'not comparable with the Full30 main table in subject count, evaluation purpose, or protocol.' Without a same-protocol PaRa baseline, the paper cannot substantiate claims about the relative merits of its rank-constrained approach. The paper acknowledges this gap but it remains a load-bearing omission for a paper whose central training-side contribution is defined relative to PaRa.
minor comments (5)
  1. Section 3.1: The notation θ_s for subject-adapted parameters and θ_0 for pretrained parameters is introduced but θ_s is not used consistently in later sections; consider unifying.
  2. Figure 4: The qualitative comparison figure is described as 'schematic' and the rightmost column shows 'failure cases' that appear to be illustrative rather than real outputs. If these are not actual generated images, this should be stated more prominently in the figure caption.
  3. Table 4: The DCAL parameter sensitivity analysis is only on Heldout9 (9 subjects). The paper correctly notes this should not be extrapolated to Full30, but it would strengthen the paper to at least note whether the trends are consistent with the Full30 results in Table 2.
  4. Section 5.1, Eq. (11)-(12): The hard and smooth schedules for α(t) are defined, but the specific values of τ_lo and τ_hi used in experiments are not clearly stated in the main text. The stage-aware inference scaling mentions α_hi=14 and α_lo=16, but the threshold values are not specified.
  5. The paper would benefit from a clearer statement upfront (e.g., in the abstract or introduction) that only DCAL-on-LoRA has complete Full30 validation, to set reader expectations appropriately.

Circularity Check

0 steps flagged

No circularity: theoretical results are straightforward derivations from stated assumptions; no fitted parameters are disguised as predictions; no self-citation chain is load-bearing.

full rationale

The paper's theoretical results (Propositions 4.1–4.8) are straightforward mathematical derivations from explicitly stated assumptions (spectral norm submultiplicativity, Pythagorean identity, cosine similarity algebra). No step reduces to its own inputs by construction. Proposition 4.2 (adapter perturbation bound) follows directly from standard matrix norm inequalities. Proposition 4.6 (identity-threshold radius) is a direct algebraic identity: for unit vectors, the squared Euclidean distance equals 2(1−cosine similarity), so a threshold on cosine similarity mechanically bounds the radius — but this is explicitly acknowledged as a conditional statement, not a fitted-then-predicted result. Proposition 4.4 (uniform-scaling feasibility) is a set-intersection argument with no empirical fitting. The DCAL selection rule (Eq. 17) is a weighted sum of three explicitly defined scores; the corollary (4.8) that weights determine ranking is trivially true by construction of the weighted sum, but this is not presented as a prediction — it is presented as a design rationale. No self-citation is load-bearing: PaRa [9] is cited as a conceptual baseline but the paper's own results do not depend on PaRa's claims being true. The empirical results compare DCAL against a K=1 LoRA baseline, which is a methodological confound (best-of-K effect) but not a circularity issue — the paper does not claim that DCAL's gains are predicted by its theory, only that they are observed. The theory is low-information but non-circular.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces standard adaptation hyperparameters and relies on basic linear algebra axioms. The key domain assumption (perturbation-budget intervals) is an abstract construct used for motivation rather than an empirically grounded entity.

free parameters (3)
  • alpha_hi, alpha_lo = 14, 16
    Scaling values for stage-aware inference scaling, chosen for the boundary experiment (Sec. 6.1)
  • lambda_I, lambda_T, lambda_D = 1.0, 0.35, 0.15
    DCAL selection weights for identity, text, and diversity scores (Sec. 6.1)
  • tau (threshold) = Not specified
    Threshold for two-stage hard schedule, mentioned but not given a specific value in main text
axioms (4)
  • standard math Spectral norm submultiplicativity
    Used in proof of Lemma 4.2 (Sec. A.2)
  • standard math Pythagorean identity for orthogonal projections
    Used in proof of Proposition 4.5 (Sec. A.4)
  • domain assumption Existence of acceptable perturbation-budget intervals [L, U]
    Introduced in Sec. 4.3 as an abstract object to analyze capacity mismatch; never empirically validated
  • domain assumption Local linear layer approximation inside U-Net
    Stated in Sec. 4.2 to derive perturbation bounds

pith-pipeline@v1.1.0-glm · 22219 in / 1858 out tokens · 222381 ms · 2026-07-09T18:16:55.554281+00:00 · methodology

0 comments
read the original abstract

Subject-driven personalized text-to-image generation requires a pretrained diffusion model to acquire a specific subject from a few reference images while preserving subject identity, following novel text prompts, and maintaining sample diversity. Existing optimization-based methods instantiate subject adaptation through full fine-tuning, textual embedding optimization, or low-rank parameter updates; PaRa further constrains personalization from the perspective of parameter rank reduction. However, a uniform low-rank constraint or a uniform adapter strength cannot explicitly distinguish the capacity requirements of different denoising stages. Moreover, inference-time candidate selection driven mainly by identity similarity may compress the selected samples in the visual representation space. We decompose the problem into two complementary components: SPaRa denotes training-side stage-aware low-rank adaptation, DCAL denotes inference-side distribution-calibrated candidate selection, and SPaRa-DCAL denotes the combined framework. Theoretical analysis shows that timestep-dependent scaling controls the effective perturbation magnitude of a low-rank adapter, while identity-biased candidate selection restricts the radius of selected features around the reference center under explicit conditions. Auditable experiments under the SDXL and DreamBooth 30-subject protocol show that DCAL improves 1-LPIPS, CLIP-I, DINO-I, and CLIP-T on a fixed LoRA candidate pool, while revealing a clear trade-off with CLIP/DINO pairwise diversity and pairwise LPIPS. These results indicate that personalized generation should be evaluated through identity consistency, text alignment, and representation diversity rather than identity metrics alone.

Figures

Figures reproduced from arXiv: 2607.07173 by Alizer Wong, Wenyan Xu.

Figure 1
Figure 1. Figure 1: Motivation. Stage-agnostic adaptation and identity-only [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SPaRa–DCAL. Given reference images and a text prompt, SPaRa introduces timestep-dependent low-rank adaptation [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Full30 identity-diversity trade-off. DCAL on LoRA im [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic qualitative comparison. Each row shows reference images, a text prompt, LoRA output, SPaRa output, DCAL selected [PITH_FULL_IMAGE:figures/full_fig_p013_4.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

23 extracted references · 23 canonical work pages · 11 internal anchors

  1. [1]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,

    work page

  2. [2]

    High-Resolution Image Synthesis with Latent Diffusion Models

    [Online]. Available: https://arxiv.org/abs/2112.10752

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,” arXiv preprint arXiv:2307.01952, 2023. [Online]. Available: https://arxiv.org/abs/2307.01952

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion

    R. Gal, Y . Alaluf, Y . Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, and D. Cohen-Or, “An image is worth one word: Personalizing text-to-image generation using textual inversion,”arXiv preprint arXiv:2208.01618, 2022. [Online]. Available: https://arxiv.org/abs/2208.01618 12 Figure 4. Schematic qualitative comparison. Each row shows reference images, a...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  5. [5]

    DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation

    N. Ruiz, Y . Li, V . Jampani, Y . Pritch, M. Rubinstein, and K. Aberman, “Dreambooth: Fine tuning text-to- image diffusion models for subject-driven generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. [Online]. Available: https://arxiv.org/abs/2208.12242

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  6. [6]

    Multi-Concept Customization of Text-to-Image Diffusion

    N. Kumari, B. Zhang, R. Zhang, E. Shechtman, and J.-Y . Zhu, “Multi-concept customization of text-to-image diffusion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. [Online]. Available: https://arxiv.org/abs/2212.04488

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    LoRA: Low-Rank Adaptation of Large Language Models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,”arXiv preprint arXiv:2106.09685, 2021. [Online]. Available: https://arxiv.org/abs/2106.09685

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  8. [8]

    Low-rank adaptation for fast text-to-image diffusion fine-tuning,

    cloneofsimo, “Low-rank adaptation for fast text-to-image diffusion fine-tuning,” https://github.com/cloneofsimo/lora, 2022

    work page 2022

  9. [9]

    SVDiff: Compact Parameter Space for Diffusion Fine-Tuning

    L. Han, Y . Li, H. Zhang, P. Milanfar, D. Metaxas, and F. Yang, “Svdiff: Compact parameter space for diffusion fine-tuning,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023. [Online]. Available: https://arxiv.org/abs/2303.11305

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction

    S. Chenet al., “Para: Personalizing text-to-image diffusion via parameter rank reduction,”arXiv preprint arXiv:2406.05641, 2024. [Online]. Available: https: //arxiv.org/abs/2406.05641

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Disenbooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation,

    H. Chen, Y . Zhang, S. Wu, X. Wang, X. Duan, Y . Zhou, and W. Zhu, “Disenbooth: Identity-preserving disentangled tuning for subject-driven text-to-image generation,” inInternational Conference on Learning Representations, 2024

    work page 2024

  12. [12]

    Attndreambooth: Towards text-aligned personalized text-to-image generation,

    L. Pang, J. Yin, B. Zhao, F. Wu, F. L. Wang, Q. Li, and X. Mao, “Attndreambooth: Towards text-aligned personalized text-to-image generation,” inAdvances in Neural Information Processing Systems, 2024

    work page 2024

  13. [13]

    Hollowed net for on-device personalization of text-to-image diffusion models,

    W. Cho, S. Choi, D. Das, M. Reisser, T. Kim, S. Yun, and F. Porikli, “Hollowed net for on-device personalization of text-to-image diffusion models,” inAdvances in Neural Infor- mation Processing Systems, 2024

    work page 2024

  14. [14]

    Instantbooth: Per- sonalized text-to-image generation without test-time finetun- 13 ing,

    J. Shi, W. Xiong, Z. Lin, and H. J. Jung, “Instantbooth: Per- sonalized text-to-image generation without test-time finetun- 13 ing,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024

    work page 2024

  15. [15]

    Ms- diffusion: Multi-subject zero-shot image personalization with layout guidance,

    X. Wang, S. Fu, Q. Huang, W. He, and H. Jiang, “Ms- diffusion: Multi-subject zero-shot image personalization with layout guidance,” inInternational Conference on Learning Representations, 2025

    work page 2025

  16. [16]

    Universalbooth: Model-agnostic personalized text-to-image generation,

    S. Liu, R. Yu, and X. Wang, “Universalbooth: Model-agnostic personalized text-to-image generation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025

    work page 2025

  17. [17]

    Customcontrast: A multilevel contrastive perspective for subject-driven text-to-image customization,

    N. Chen, M. Huang, Z. Chen, Y . Zheng, L. Zhang, and Z. Mao, “Customcontrast: A multilevel contrastive perspective for subject-driven text-to-image customization,” inAAAI Confer- ence on Artificial Intelligence, 2025

    work page 2025

  18. [18]

    Disenvisioner: Disentangled and enriched visual prompt for customized image generation,

    J. He, H. Li, Y . Hu, G. Shen, Y . Cai, W. Qiu, and Y .-C. Chen, “Disenvisioner: Disentangled and enriched visual prompt for customized image generation,” inInternational Conference on Learning Representations, 2025

    work page 2025

  19. [19]

    Dreammatcher: Appearance matching self-attention for semantically-consistent text-to-image personalization,

    J. Nam, H. Kim, D. Lee, S. Jin, S. Kim, and S. Chang, “Dreammatcher: Appearance matching self-attention for semantically-consistent text-to-image personalization,” in Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2024

    work page 2024

  20. [20]

    Learning Transferable Visual Models From Natural Language Supervision

    A. Radfordet al., “Learning transferable visual models from natural language supervision,” inInternational Conference on Machine Learning, 2021. [Online]. Available: https://arxiv.org/abs/2103.00020

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  21. [21]

    DINOv2: Learning Robust Visual Features without Supervision

    M. Oquabet al., “Dinov2: Learning robust visual features without supervision,”Transactions on Machine Learning Research, 2024. [Online]. Available: https: //arxiv.org/abs/2304.07193

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. [Online]. Available: https://arxiv.org/abs/1801.03924

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  23. [23]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004. 14 A Proof Details This appendix provides proof details for the statements in Sec. 4. All variables follow the definitions in Sec. 3. A.1 Proof of Pro...

    work page 2004