acceptodds
Desk rejected at ICLR 2027

PaRDi: Learning to Preserve and Revise with Discrete Diffusion

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing discrete diffusion methods for image editing often encode the source as a reference sequence and generate a separate target from a fully masked canvas—a paradigm we call reference-based regeneration. This forces even content meant to remain unchanged to be generated again, potentially introducing unwanted drift, such as background changes during foreground edits or facial changes during clothing edits. We introduce PaRDi, a discrete diffusion framework that starts from the source canvas and models preservation and repeated revision within the same dynamics. Each token can retain its value or pass through <MASK> before taking a new value, and newly generated tokens remain eligible for further revision. These asynchronous updates produce partially revised canvases in which preserved source tokens and completed revisions jointly guide subsequent generation, supporting both local and image-wide edits. Endpoint-conditioned bridges, combined with constant paths at unchanged positions, analytically provide training-state distributions and transition targets, yielding training canvases and transition-matching objectives without hand-designed remasking trajectories. PaRDi achieves 6.98 on GEdit-EN, 3.95 on ImgEdit, and 92.27% background retention on PIE-Bench. Extending in-place editing to visual reasoning, PaRDi attains 87.29% average accuracy, outperforming DiffThinker's 75.36%.

Primary area:
generative models
Keywords:
Discrete Diffusion, Image Editing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7WwvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

PaRDi studies content preservation and repeated revision in discrete image editing. It adapts pretrained model using LoRA and a time-conditioning module, starting editing directly from the source token canvas. The method constructs a MASK-hub Markov process and conditions it on source–target endpoints to derive intermediate training states and transition supervision, while assigning constant paths to unchanged positions. The model predicts target-token probabilities, which an analytic mapping converts into keep, remask, and reveal transitions. Experiments evaluate image editing, background preservation, and visual reasoning.

Critical strengths

The paper unifies intermediate training states, transition supervision, and inference rules within a single bridge process, forming a coherent methodological contribution. Ablations of transition structure, state construction, and training objectives under shared training settings, together with revision-budget experiments under a fixed update schedule, provide useful evidence for the individual components. The appendix also visualizes actual revision actions and sampling trajectories, helping explain how the model progressively modifies the source canvas.

Critical weaknesses
  1. The ImgEdit training subset, benchmark exclusions, and deduplication procedure are unspecified, leaving evaluation independence unclear.

  2. The fraction of residual MASK tokens and the effect of terminal completion—which excludes source tokens—on editing quality and preservation are not reported.

Questions for authors
  1. The training–evaluation data split is insufficiently documented. The paper trains on ImgEdit and evaluates on ImgEdit-Bench Basic, but does not specify the exact training subset, benchmark exclusions, or deduplication procedure. These details are needed to confirm evaluation independence; there is currently no evidence of data leakage.

  2. The impact of terminal completion is unclear. After 30 updates, the method fills remaining MASK tokens while excluding the original source token at each position. However, the proportion of tokens requiring completion and its effect on the results are not reported. Please quantify this proportion and clarify whether the rule affects editing quality and background preservation.

3 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer 7Wwv

Thank you for recognizing the coherent unification of training states, transition supervision, and inference rules, as well as the controlled ablations and sampling visualizations. We address the two documentation gaps with explicit data-use details, an image-overlap audit, and direct terminal-completion experiments.

W1. ImgEdit training subset, benchmark exclusions, and deduplication

To ensure reproducibility, we train on the full publicly released ImgEdit training set and evaluate on the official ImgEdit-Bench Basic, reported as ImgEdit-Bench in the original ImgEdit paper's main results table.

ImgEdit [1] appeared in the NeurIPS 2025 Datasets and Benchmarks Track. Sections 3.2 and 4.1 describe the construction of its official training set and benchmark. We follow the official training/evaluation separation without additional benchmark-specific deduplication.

We further audited image overlap using the publicly available, widely used Python ImageHash library. No duplicates were detected under the pixel-SHA or pHash-distance ≤4 checks. We compared source and raw target images from the ImgEdit training set against the original source images of ImgEdit-Bench, GEdit-EN, and PIE-Bench. Exact matching uses the Secure Hash Algorithm (SHA) on pixel data obtained using a consistent decoding procedure and on image dimensions; perceptual hashing (pHash) screens image similarity.

BenchmarkOriginal imagesExact pixel-SHA matches64-bit pHash distance ≤4Permissive crop-hash candidates
ImgEdit-Bench1710052 (30.41%)
GEdit-EN60600176 (29.04%)
PIE70000256 (36.57%)

Even under permissive crop-hash screening, ImgEdit-Bench has a candidate-match rate comparable to those of GEdit-EN and PIE.

[1] Ye et al., ImgEdit, NeurIPS Datasets and Benchmarks Track (2025).

W2. Residual MASK fraction and the effect of terminal completion excluding source tokens

After 30 updates, residual MASK tokens account for 2.4385% of all image tokens pooled across GEdit, ImgEdit, and PIE. Allowing source tokens in terminal completion, or continuing until no MASK remains, yields editing quality and background preservation close to the original method.

We use a fixed 30 steps to enable fair comparisons under a common inference budget. Terminal completion fills only the remaining MASK positions, selecting the highest-probability token after excluding MASK and the position's original source value.

Sampling / completion settingGEdit-EN ↑ImgEdit ↑Keep (%) ↑Leak (%) ↓Mean revision steps
Fixed 30; terminal completion excludes MASK and source6.983.9592.2714.1930.00
Fixed 30; terminal completion excludes MASK only6.983.9592.2814.2130.00
Continue stochastic revision until no MASK remains6.973.9592.2514.1834.90

Continuing to sample until no MASK remains provides no clear performance gain but requires more sampling steps.

When source values are allowed, substantially more source-valued completions occur in the editing region than in the non-editing region. We therefore exclude both MASK and source values, rather than MASK alone, to favor editing at residual positions. Both columns pool token counts across PIE samples and divide by the total number of image tokens:

PIE regionTokens receiving terminal completion / all tokensSource-valued completions when allowed / all tokens
Editing region2.269%0.141%
Non-editing region0.639%0.024%

Q1. Insufficiently documented training–evaluation split

We use the full official ImgEdit training set and evaluate on official ImgEdit-Bench; our audit finds no pixel-exact or pHash-distance ≤4 matches. See response to R 7Wwv W1 for the official split, evaluation scope, and audit results.

Q2. Proportion and impact of terminal completion

The completion fraction is 2.4385% across the three benchmarks; allowing source tokens in terminal completion or continuing sampling until no MASK remains yields aggregate scores close to the original method. Please see response to R 7Wwv W2 for the exact denominator, PIE region breakdown, and editing/preservation comparisons.

I thank the authors for their detailed response, which has satisfactorily addressed my concerns.

Dear Reviewer,

Thank you for revisiting our response and confirming that your concerns have been satisfactorily addressed. We appreciate your recognition of PaRDi’s unified derivation of training states, transition supervision, and inference rules, supported by controlled ablations.

Your feedback helped us clarify our use of the full publicly released ImgEdit training set and official benchmarks to support reproducible evaluation. The additional audit supports training–evaluation independence, while the completion experiments show stable editing quality and content preservation across completion settings. We will incorporate these details and analyses into the final manuscript.

Thank you again for your careful evaluation and constructive feedback.

Best regards,
The Authors

Reviewer_62ZiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a framework for image editing based on discrete diffusion. In contrast to other work that either does reference-based regeneration and having the source as a reference but working on a at the beginning fully masked target canvas, or methods like DICE/GIDE that do training-free in-place token editing, the proposed method revises the source canvas in place, allowing source, <MASK>, and renewed tokens to coexist.

The paper reports gains over its backbone and over training-free in-place methods on GEdit-EN, ImgEdit, and PIE-Bench preservation (cf. Table 1). The method further transfers to grid-based visual reasoning, where it beats DiffThinker (cf. Table 4).

Critical strengths
  1. Interesting problem; clean, well-motivated formulation; well-justified solution with promising results. The problem is well motivated, all steps in the pipeline are derived in a meaningful way and well justified and the results show real, if moderate, gains over the methods it compares to.

  2. Thorough mathematical derivation and detailed appendix. I could not check the mathematical details because there are just too many, but the authors put a lot of effort to describe all aspects of the proposed method in thorough detail. I would assume that the paper could be re-implemented from the information given in the paper, but I would still hope that the code would be released upon acceptance of the paper.

Critical weaknesses
  1. While I really appreciate all the mathematical details, the paper is dense in notation and I found it quite hard to comprehend. This in turn made it somehow impossible for me to completely verify their correctness. Some terms are not well introduced and make it very difficult to understand their intuition or meaning, e.g. RτR_{\tau} defined in Equation 1 or rsr_s.

  2. Masked positions cannot return to their source value and it somehow remains completely unclear what implications this has on the results. It seems that in the current formulation tokens that are masked cannot ever get their original values back.

Questions for authors
  1. Does the fact, that the method cannot get a masked value back have any substantial influence?
4 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer 62Zi - Part 1/2

Thank you for recognizing the clean, well-motivated formulation, the meaningful derivation, and the thorough mathematical appendix. We appreciate your requests for clearer notation and a precise account of source-token recovery.

W1. Dense notation and insufficient intuition for the rate matrix and proposal

We have revised Section 3.2 to explain the rate matrix, the proposal distribution, and the relation between rates and jump probabilities.

At a single image position, the state can be any of the VV visual-token values or an additional MASK state, which represents a masked token. For example, with a two-value vocabulary A,B\\{A,B\\}, the state set is A,B,MASK\\{A,B,\mathrm{MASK}\\}. A token can remain in its current state or change through MASK: A→MASK→BA\to\mathrm{MASK}\to B. The rate matrix specifies how quickly these transitions occur, while the proposal distributes the total reveal rate across visual-token values.

For this position, ss is its source token value, yy its target token value, and XτX_\tau its state at normalized time τ∈[0,1]\tau\in[0,1]. The source endpoint is X1=sX_1=s at τ=1\tau=1; the target endpoint is X0=yX_0=y at τ=0\tau=0. Forward modeling uses target →\to source (τ:0→1\tau:0\to1), while reverse generation uses source →\to target (τ:1→0\tau:1\to0).

  1. Forward token updates: when does MASK appear, and how is it filled? The nonnegative rate λτ\lambda_\tau controls how quickly a non-MASK token becomes MASK. The nonnegative rate μτ\mu_\tau is the total rate at which MASK becomes any non-MASK token: a larger rate gives a higher chance of filling MASK over the same short forward-time interval.

    For a particular non-MASK value vv, such as A or B, the proposal probability rs(v)r_s(v) specifies its share of that total rate. Thus, MASK becomes vv at rate μτrs(v)\mu_\tau r_s(v). Our implementation uses the same fixed uniform proposal rs(v)=1/Vr_s(v)=1/V for every source value ss. For example, if the vocabulary contains only A and B, their reveal rates are each μτ/2\mu_\tau/2. The proposal defines the forward process; endpoint conditioning and model predictions determine the actual generation rates.

  2. The rate matrix: how are these rules represented? RτR_\tau collects the forward transition rates, with rows indexing the current state and columns the next state. Let IVI_V be the V×VV\times V identity matrix, 1V\mathbf{1}_V the all-ones column of length VV, and rs\boldsymbol r_s the column containing the VV proposal probabilities; ⊤\top denotes transpose. Ordering the non-MASK values before MASK gives Eq. (1):

    Rτ=(−λτIVλτ1Vμτrs⊤−μτ).R_{\tau}=\begin{pmatrix} -\lambda_{\tau} I_{V} & \lambda_{\tau}\mathbf{1}_{V} \\\\ \mu_{\tau}\boldsymbol{r}_{s}^{\top} & -\mu_{\tau} \end{pmatrix}.

    Diagonal entries are negative outgoing-rate sums, so each row sums to zero. To relate rates to probabilities, let u,vu,v be any two distinct token states, including MASK, and Δτ>0\Delta\tau>0 a short time interval. Writing Pr⁡R\Pr_R for probabilities under this forward process,

    Pr⁡R(Xτ+Δτ=v∣Xτ=u)≈Rτ(u,v)Δτ.\Pr_R(X_{\tau+\Delta\tau}=v\mid X_\tau=u)\approx R_\tau(u,v)\Delta\tau.

    In words, the probability of changing from uu to vv over a short interval is approximately the corresponding rate times the interval length.

  3. How does this produce a specified edit? For source s=As=A and target y=By=B, a possible reverse-generation path is A→MASK→BA\to\mathrm{MASK}\to B. During training, conditioning on the paired source and target yields intermediate states and target-directed transition supervision. During inference, the model predicts target-token probabilities from the current canvas, instruction, and time; the analytic bridge mapping converts these predictions into transition rates that determine whether tokens are kept, remasked, or revealed. Sections 3.3–3.5 give the corresponding derivation and sampling rules.

We will release the code and models as soon as the double-blind review concludes.

AuthorsOpenReview ↗

Authors' Response to Reviewer 62Zi - Part 2/2

W2. Whether masked positions can return to their source value, and the practical implications

Masked positions can return to their source tokens during regular revision. The existing sampler's proposal covers the entire visual-token vocabulary, including the source token, which remains a valid sampling destination. Returning to it requires no additional recovery operation.

In PaRDi, constant paths model preservation, while training perturbations improve robustness to masking errors. For positions with identical source and target tokens, the desired behavior is to preserve the correct content throughout editing. During training, we independently mask each such position with probability 10%, retaining the original source token as the target for the existing Target CE. Unperturbed positions learn to preserve, while perturbed positions learn to recover.

We compare full PaRDi with variants that remove these recovery perturbations or add an independent source-reference image:

SettingGEdit-EN ↑ImgEdit ↑Keep (%) ↑Leak (%) ↓PIE Target CLIP ↑BG MSE ↓BG SSIM ↑
PaRDi6.983.9592.2714.1927.190.0047090.8861
w/o non-editing-position recovery perturbations6.8273.5843691.222618.930726.47510.0047900.883788
PaRDi + independent source-reference image6.9943.977191.8214.4725.21540.0058350.8882

Even without recovery augmentation, PaRDi achieves strong editing and preservation performance, demonstrating the effectiveness of its base formulation. Recovery augmentation further improves instruction following and background preservation. Adding an independent source reference yields similar general editing performance and higher background structural similarity, while the current-canvas model achieves better token preservation, edited-region alignment, and background pixel fidelity.

The predictor infers masked source content from visible tokens, the editing instruction, and learned visual priors. Neighboring tokens retained during asynchronous updates provide context. In the extreme case where a distinctive detail is represented by a single token, masking that token may leave insufficient context for exact recovery. An independent source reference provides continued access to the original content, but doubles the number of image tokens and increases computational cost, with self-attention scaling quadratically in the total sequence length.

The revised manuscript specifies the training-time recovery augmentation in Section 3.4 and Appendix C.5, with the corresponding update to Algorithm 3. Appendix A.7 clarifies how the existing sampler permits returns to source values.

Q1. Does inability to recover a masked source value have a substantial influence?

Masked positions can return to their source tokens. PaRDi achieves strong editing and preservation performance even without the recovery augmentation; adding it further improves both. An independent source reference provides only limited gains in general editing performance on the evaluated benchmarks. See response to R 62Zi W2 for the mechanism and ablation results.

Reviewer_62ZiOpenReview ↗

Thanks

I thank the authors for the detailed response.

*) Notation: The added explanation of RτR_\tau and rsr_s, including the small example with the two-token vocabulary, makes Section 3.2 much easier to follow. This addresses my concern.

*) Masked positions: Thank you for clarifying that masked tokens can return to their source value during regular revision, and that only the terminal completion excludes the source value. The new ablation is helpful. I note that removing the recovery perturbations has a clear effect (e.g., ImgEdit 3.95 -> 3.58, Leak 14.19 -> 18.93), so I am glad this is now properly described in the paper.

My concerns are addressed and I keep my positive score. I would still strongly encourage the authors to release the code as promised.

Dear Reviewer,

Thank you for confirming that your concerns have been addressed and for maintaining your positive assessment. We appreciate your recognition of PaRDi’s clean, well-motivated formulation and detailed derivations, and your feedback has helped us present the method more clearly.

As promised, we will release the code and model checkpoints as soon as the double-blind review concludes to support reproducibility and further research.

Thank you again for your time and constructive feedback.

Best regards,
The Authors

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

This paper proposes PaRDi, a discrete diffusion framework for instruction-guided image editing that operates directly on the source image's visual-token canvas. Unlike reference-based approaches that regenerate a separate target sequence from a fully masked canvas, PaRDi preserves unchanged tokens while allowing other tokens to undergo repeated masking and renewal. The main methodological contribution is an endpoint-conditioned Markov bridge formulation that analytically derives intermediate training distributions and target-directed transition rates from paired source and target images. These are used to train a target-token predictor with a combination of cross-entropy and transition-rate matching. The authors do experiments which demonstrate improvements in image-editing quality and preservation on GEdit-Bench, ImgEdit-Bench, and PIE-Bench. The authors additionally evaluate visual reasoning tasks, showing that repeated token revision improves solution accuracy under a fixed inference schedule.However, I am not fully convinced how much of the improvement comes from the proposed bridge formulation compared to simpler preservation and remasking strategies. In particular, a more direct comparison with Reflective Masking and an ablation of the constant-path preservation mechanism would help establish the novelty of the approach.

Critical strengths

Technical contribution. The endpoint-conditioned MASK-hub bridge provides a principled formulation of preservation and revision, rather than relying entirely on manually designed remasking trajectories.

Strong controlled ablations. Comparisons between absorbing diffusion, direct-transition bridges, and MASK-hub bridges isolate the effect of the transition topology. Additional experiments separate bridge-derived training states from transition-matching supervision.

Convincing preservation results. The method improves both editing quality and retention of unchanged image content, with substantial reductions in unwanted edits outside target regions.

Interesting extension to visual reasoning. The revision-budget experiments provide evidence that repeated revisions are useful for tasks requiring global consistency, particularly Sudoku and TSP.

Critical weaknesses

My main concern is the novelty relative to existing iterative editing and remasking approaches, particularly Reflective Masking. While the bridge formulation is different, the paper does not fully establish whether this leads to improvements over a comparable learned revision strategy.

The ablation results suggest that a significant part of the improvement comes from the bridge-derived training states rather than the transition-matching objective itself. For example, using bridge states with target CE already achieves 6.33 on GEdit-EN compared to 6.98 for the complete approach. It would be helpful to better isolate these contributions.

The method explicitly uses constant paths for unchanged source-target tokens. This seems to provide a strong preservation prior, but its contribution is not independently quantified. It is unclear how much of the preservation improvement comes from this mechanism versus the broader bridge dynamics.

The evaluation would benefit from a more controlled comparison against simpler source-initialized editing methods under the same backbone, training data, and inference budget.

Questions for authors

How does PaRDi compare with Reflective Masking under a matched experimental setup? In particular, what additional benefit does the Markov bridge provide over learned remasking when both approaches start from the source canvas?

Can the authors isolate the effect of constant-path preservation? How much of the improvement in background retention comes from keeping unchanged tokens fixed versus learning the proposed transition dynamics?

What is the main benefit of transition-rate matching beyond bridge-derived training states? Table 3 suggests that changing the training-state distribution already provides substantial improvements. Could the authors discuss this distinction more clearly?

How does the method behave on edits that require large structural changes? It would be interesting to understand whether the strong preservation prior can become restrictive when a significant portion of the image needs modification.

5 comments
AuthorsOpenReview ↗

Authors' Response to Reviewer PnQg - Part 1/4

Thank you for recognizing the principled MASK-hub formulation, the controlled ablations, the preservation results, and the extension to visual reasoning. Your questions help us clarify the contribution beyond learned remasking and quantify the roles of training states, transition supervision, and constant-path preservation.

W1. Novelty relative to iterative editing and remasking, particularly Reflective Masking

PaRDi formulates editing as a unified discrete diffusion process that jointly determines training states, transition supervision, and inference rules. Reflective Masking (RM) [1] instead learns revision decisions from hand-designed trajectories using tokenwise CE, without such a unified formulation. PaRDi's endpoint-conditioned bridges yield intermediate states and transition rates; the same analytic mapping converts target-token predictions into editing updates. The original submission illustrates this distinction in Figure 11 of Appendix D.6.

RM's public release lacks fine-tuned checkpoints, processed data, and the preprocessing pipeline; its image-editing code also differs from the paper in revision rules and auxiliary losses. The paper evaluates on a custom held-out ImgEdit subset rather than the official ImgEdit-Bench released with the training dataset (NeurIPS 2025). It neither evaluates on any public image-editing benchmark nor reports reproducible scores under those protocols. Our emails requesting materials and clarification remain unanswered, as do related GitHub requests.

We trained and verified two RM implementations on the same backbone and editing data. RM (paper) matches PaRDi's inference budget; RM (released code) retains the released sampling configuration. Their settings are:

SettingRM (paper)RM (released code)PaRDi
History ReferenceEarlier editing states, aggregated causallyNoneNone
Intermediate training statesRandomly draw wrong tokens excluding the target; sample boundaries for wrong → MASK → target or MASK → target paths; keep correct positions fixed; sample from 0…T−1Same as the paper, but sample from 0…TSample time and token states from the analytic source–target bridge distribution; equal endpoints use constant paths with additional recovery perturbations
Supervision targetsWrong visible token → MASK; masked or correct token → targetSame as the paperTarget token at each position, together with analytic transition-rate targets
Training lossTarget CE + Keep/remask CE (equal keep/mask/renew weights) + non-edit-position unlikelihood loss + stage-2 ordering lossTarget CE + Keep/remask CE (equal keep/mask/renew weights)Target CE + PaRDi rate loss
Revision ruleNo separate masking and unmasking phasesTwo phases: masking first, then unmaskingPer-token keep, remask, and reveal transitions determined by bridge-induced rates
CFG scale1.54.01.5
Update schedule30 revision stepsAt most 8 masking + 8 unmasking steps; early stopping allowed30 stochastic revision steps

We additionally searched the keep/mask/renew CE weights for the released-code implementation. The table reports editing quality on GEdit-EN and ImgEdit-Bench, and preservation on PIE-Bench:

MethodGEdit-EN ↑ImgEdit ↑PIE Keep (%) ↑PIE Leak (%) ↓
RM (paper)1.3671.78381000
RM (released code)1.4341.76891000
RM (released code, tuned CE weights)2.451.8284.2722.36
PaRDi + RM training states (Table 3, row 3)2.571.9385.4521.49
PaRDi + RM objective (Table 3, row 5)6.743.7989.6518.75
PaRDi6.983.9592.2714.19

Under their original configurations, both RM reproductions produce nearly unchanged outputs. Image editing often requires preserving much of the source image; with equal per-token CE weights, a predominance of keep-supervised positions can bias learning toward keep. Tuning these weights improves editing quality, but PaRDi still achieves substantially better editing quality and content preservation than the tuned RM baseline.

The controlled variants in Table 3 further establish the value of PaRDi's unified formulation: replacing bridge states with RM-style hand-designed states or rate matching with Keep/remask CE reduces both editing quality and content preservation. These comparisons demonstrate the benefits of the derived states and supervision; see response to R PnQg W2 for the theoretical relationship between the objectives.

We will release the data, preprocessing and trajectory-construction code, training and evaluation code, configurations, and model checkpoints for PaRDi and both RM implementations to make these comparisons reproducible.

[1] Zhang et al., Reflective Masking, arXiv (2026).

AuthorsOpenReview ↗

Authors' Response to Reviewer PnQg - Part 2/4

W2. Bridge-derived training states versus the transition-matching objective

PaRDi contributes a unified editing process that jointly determines training states, transition supervision, and inference rules. The bridge-state gains are a core contribution of this process, and rate matching provides additional benefits with those states held fixed. As in Gaussian diffusion, the state process defines the learning problem, while different objectives can be used to learn its reverse dynamics. The bridge-state CE baseline uses alternative supervision within the PaRDi process. Its gains over training on hand-designed states under the same CE objective demonstrate the contribution of PaRDi's derived training states.

Appendix B.5–B.6 shows how Keep/remask CE can serve as a classification surrogate for PaRDi rate loss. Under the analyzed shared-predictor parameterization, with the same conditioning information and conditional target distribution, the two objectives share the optimal calibrated probability, yet their loss functions and gradient forms remain non-equivalent.

With bridge states fixed, Table 3 shows that the full PaRDi supervision improves instruction following and content preservation over both Target CE alone and Target CE + Keep/remask CE.

W3. Independent contribution of constant-path preservation

The constant-path ablation shows that constant paths reduce leakage and background pixel error, while the learned bridge dynamics sustain high editing scores and token retention.

SettingGEdit-EN ↑ImgEdit ↑Keep (%) ↑Leak (%) ↓Target CLIP ↑BG MSE ↓BG SSIM ↑
PaRDi w/o constant paths6.953.908392.470516.134726.070.0055100.888161
PaRDi6.983.9592.2714.1927.190.0047090.8861

The variant without constant paths preserves slightly more background tokens and achieves a slightly higher SSIM, while full PaRDi provides better editing alignment and background pixel fidelity. Keep measures unchanged tokens among background tokens; Leak measures background changes among all changed tokens.

Constant paths supervise preservation where paired training endpoints agree. At inference, the model selects what to preserve using the current canvas, instruction, and time. response to R PnQg Q4 reports the structural-editing evaluation.

AuthorsOpenReview ↗

Authors' Response to Reviewer PnQg - Part 3/4

W4. Controlled comparison against simpler source-initialized editing methods

Under the same backbone, training data, and inference budget, PaRDi improves both editing and content preservation over the RM-style source-initialized baseline trained with hand-designed edit states and Target CE + Keep/remask CE (Table 3).

We additionally compare source reference + full-MASK regeneration (Table 2) with source reference + PaRDi source-canvas editing, using the same backbone and training data. Both methods retain access to the original source reference, isolating the comparison between regeneration and source-canvas editing:

Editing processGEdit-EN ↑ImgEdit ↑Keep (%) ↑Leak (%) ↓
Source reference + full-MASK regeneration (Table 2)6.613.5284.5231.84
Source reference + PaRDi source-canvas editing6.9943.977191.8214.47

With source-reference conditioning retained, PaRDi source-canvas editing also achieves better editing and preservation.

The table below presents controlled comparisons addressing several key questions and summarizes the corresponding conclusions.

Question addressedControlled comparisonWhat the results establish
Is source initialization plus learned remasking sufficient?Hand-designed states trained with Target CE or Target CE + Keep/remask CE vs. PaRDi (Table 3); matched RM (paper) comparison (response to R PnQg W1)PaRDi unifies source initialization and learned remasking within a discrete diffusion process, deriving training states, transition supervision, and inference rules from the same formulation to provide a systematic basis for designing methods in this editing paradigm.
How does the bridge process support learned remasking?Hand-designed states vs. bridge states, holding the training objective fixed (Table 3)Bridge states define a remasking mechanism grounded in discrete diffusion theory that is easier for the model to learn.
What does the full PaRDi supervision add?With bridge states fixed, compare Target CE, Target CE + Keep/remask CE, and Target CE + PaRDi rate loss (Table 3)PaRDi rate loss directly matches the transition rates derived from the process, aligning more closely with the process underlying the training states and achieving better empirical results than alternative CE supervision.
Is preservation mainly due to constant paths?Full PaRDi vs. PaRDi w/o constant paths (response to R PnQg W3)Preservation primarily comes from PaRDi's overall formulation; constant paths further reduce leakage and background pixel error.
Does the preservation prior hinder large changes?CompBench Action, Location, and Viewpoint, including the constant-path ablation (response to R PnQg Q4)PaRDi performs the tested structural edits effectively; removing constant paths produces small, mixed effects across categories.

Q1. Matched comparison with Reflective Masking and benefit over learned remasking

PaRDi improves editing quality over the matched RM adaptation. Controlled comparisons further establish the benefits of bridge-derived training states and transition supervision. Please see response to R PnQg W1 for the matched comparison and response to R PnQg W2, response to R PnQg W3, and response to R PnQg W4 for the controlled attribution.

Q2. Constant-path preservation versus learned transition dynamics

Removing constant paths retains high editing quality and background token retention, but increases the leakage fraction and background MSE. Please see response to R PnQg W3 for the ablation and the distinction between training preservation supervision and inference-time decisions.

Q3. Main benefit of transition-rate matching beyond bridge states

With bridge states fixed, rate matching improves instruction following and content preservation together. See response to R PnQg W2 for the controlled comparison and the theoretical relationship to CE.

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

Related papers

Discussion (0)

Sign in to comment.