{"id":"6e07c519-570f-407b-aebf-cd89321f4b63","arxiv_id":"2507.13401","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A masking-augmented diffusion objective plus pause-token inference scaling modestly improves instruction adherence and source preservation for OmniGen-based image editing.","lead":"MADI adds random masking to noisy image tokens during diffusion training and inserts pause tokens at inference to improve text-guided editing with OmniGen. The reported gains on editing benchmarks are small and partly confounded by best-of-n selection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inference-time scaling results are confounded by best-of-n selection: gains attributed to pause tokens may come from sampling more candidates, not from added capacity.","rationale":"The paper's central claim is that MAgD improves editing via a masking-augmented dual-corruption objective and that pause tokens provide inference-time capacity scaling. The most load-bearing condition for the second contribution is that observed gains are caused by pause tokens rather than by the extra sampling that the protocol implicitly performs. Table 4 and Figs. 6–8 select the best CLIP-Dir across pause-token counts {0,8,16,32} per sample; without a pause=0 multi-seed control, the improvement is exactly what best-of-n selection would produce. If a best-of-4 pause=0 baseline reproduces the reported gains, the inference-time scaling claim collapses to a sampling artifact. I agree with the reader's weakest_assumption. The MAgD training-objective claim is less affected by this confound but rests on a single seed and a small CLIP-Dir delta (0.004), so it also needs error bars and public code for confidence. I additionally note an internal inconsistency: Eq. (6) applies masking when t<tau_MAgD, while the text (Sections 3.1.3 and 7) repeatedly states masking is restricted to t>=tau_MAgD; Table 9's 'Timestep conditioning T-0.3' is ambiguous. This does not by itself refute the empirical results but must be resolved for the mechanism to be reproducible. Because the reader's CONDITIONAL verdict already captures these needs, no verdict change is required.","tokens_in":17737,"tokens_out":9341,"duration_ms":91679,"concrete_test":"Run the inference-time scaling evaluation with four random seeds at pause 0 (that is, no pause tokens) and select the best CLIP-Dir per sample among those four; compare the resulting CLIP-Dir and DINO against the reported best-of-0-8-16-32 selection. If the pause 0 best-of-4 matches or exceeds the reported pause-token best-of-4 results, the gains attributed to pause tokens are a best-of-n artifact. As a secondary check, inspect the released training code to verify whether masking in Eq. (6) is applied for t less than tau_MAgD or t greater than or equal to tau_MAgD, resolving the inconsistency in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's inference-time scaling claim (Section 4.1, Table 4, Figs. 6–8) is established by selecting, per sample, the best CLIP-Dir among generations with pause tokens 0, 8, 16, and 32, and comparing against a single generation at pause 0. This protocol confounds pause-token capacity with best-of-n selection: the expected maximum of n i.i.d. samples increases with n, so any improvement over a single sample could arise from sampling multiple candidates, not from the pause tokens themselves. The causal role of pause tokens is the load-bearing assertion for the inference-time capacity scaling contribution, yet no pause 0 multi-seed baseline is reported. The MAgD training-objective gains (Table 3) are not affected by this confound but are reported from a single seed with no error bars, and the CLIP-Dir gain (0.122 to 0.126) is small. Additionally, Eq. (6) and Section 7 contain an internal inconsistency: the equation applies masking when t is less than tau_MAgD while the text says masking is restricted to t greater than or equal to tau_MAgD, making the implemented corruption schedule ambiguous. This does not by itself refute the training-objective results, but it needs correction for reproducibility.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework, MADI, built on top of the OmniGen in-context diffusion architecture for instruction-based image editing. It introduces MAgD, a training objective that adds a random token-masking corruption on the Gaussian-noised input and supervises the denoiser on the masked input, and it introduces inference-time pause tokens inserted into the prompt to provide extra computational capacity without retraining. On Emu-Edit, MagicBrush, Complex-Edit, IdeaBench, and GenEval, the authors report that MAgD improves both source faithfulness (DINO 0.882 to 0.927) and instruction adherence (CLIP-Dir 0.122 to 0.126) over the OmniGen baseline, and that pause-token inference scaling further shifts CLIP-Dir to 0.134.","tokens_in":18008,"tokens_out":8423,"duration_ms":88404,"significance":"If the training-objective results are reproducible, MAgD is a simple, computationally cheap modification to the standard diffusion objective, and the paper's explicit treatment of the faithfulness-versus-instruction-following trade-off is useful to the editing community. The pause-token mechanism is an interesting transfer of an LLM idea to diffusion-based visual editing, but its causal contribution is not yet established. Strengths include the breadth of benchmarks, component-level ablations in Tables 3, 7, and 8, and a candid discussion of the limitations of CLIP-based metrics. The paper does not, however, provide code, checkpoints, or repeated-seed variability for its headline editing numbers, which matters because several of the key reported differences are small.","major_comments":[{"comment":"The inference-time scaling evidence is confounded with best-of-n selection. The protocol in Fig. 6 selects, per sample, the generation with the best CLIP-Dir among pause-token counts {0, 8, 16, 32}, while the comparison baseline is a single generation at pause=0. Under this protocol, the pause-0 arm samples one image while the pause-token arm samples up to four, so the observed gain can be explained by the maximum of a larger sample rather than by the added capacity of the pause tokens. Table 4's 'Best CLIP' and 'Best DINO' columns are selected in the same way. A valid test requires a matched multi-sample baseline: for example, generate K images with pause=0 and select the best under the same criterion, then compare best-of-K at pause=0 with best-of-{0, 8, 16, 32}; the paper should also report results under a fixed, non-oracle pause count. Without such a baseline, the causal claim that 'pause tokens provide additional capacity' in Sections 3.3 and 4.1 is not supported.","section":"Section 4.1, Table 4, Fig. 6"},{"comment":"The masking schedule is internally inconsistent. Section 3.1.3 states that masking is restricted to high noise levels, t greater than or equal to tau_MAgD, but Eq. (6) applies the masked loss when t is less than tau_MAgD; Section 7's Eq. (9) repeats the same inconsistency, and Table 7's caption uses 't < tau_MAgD' while the surrounding text says 't >= tau_MAgD'. This is not a typographical detail: it changes which training steps receive the dual corruption and therefore changes the trained model. The authors should choose one condition and align the equation, the prose, and the table.","section":"Eq. (6) / Eq. (9), Sections 3.1.3 and 7"},{"comment":"The text claims that 'MAgD alone boosts both CLIP-Dir and DINO over OmniGen and finetuned omnigen.' The table contradicts this for the fine-tuned comparison: row 4 (MAgD without expressive prompting) has DINO 0.913, which is below the 0.917 DINO of the fine-tuned baseline in row 2, even though CLIP-Dir improves from 0.111 to 0.127. The claim should be restricted to the comparison with the untuned OmniGen checkpoint, or new data are needed to support the original statement.","section":"Table 3 and text after it"},{"comment":"All headline editing results are single-run point estimates. The main claimed CLIP-Dir improvement from 0.122 to 0.126 is a difference of 0.004, and the inference-scaling gain is 0.008; without standard deviations or confidence intervals, these differences may be within run-to-run or evaluation noise. Table 5 averages GenEval over five seeds, so the authors already have the infrastructure for seed-averaged reporting; the editing metrics should be reported with the same practice.","section":"Tables 1, 3, and 5"}],"minor_comments":[{"comment":"The product m circle x_t circle empty-set is not well-defined because the empty-set is described as a mask token (a vector) while m circle x_t is a vector; please specify how the mask embedding replaces the masked entries rather than multiplying them.","section":"Section 3.1.3, Eq. (4)"},{"comment":"The baselines paragraph cites Show-O as [13], but reference [13] in the bibliography is the scratchpad paper; Show-O is reference [24]. Please correct the citation.","section":"Baselines paragraph, Section 4"},{"comment":"The naming of rows in Table 1 ('MAgD-I' vs. 'w/o Inference Scaling') is inconsistent with Figure 2 ('MAgD-I (w/o inference scaling)' and 'MAgD-I'), making it difficult to tell which model is the training-objective ablation; please use a single naming convention throughout.","section":"Table 1 and Figure 2"},{"comment":"The column header 'Best CLIP' is ambiguous: the first block reports DINO and CLIP-Dir under a selection rule that likely means best CLIP-Dir; please clarify which metric is used for selection and which metric is reported.","section":"Table 4"},{"comment":"Table 9 lists 'Timestep conditioning T-0.3', which is inconsistent with the tau >= 0.7 description in Sections 3.1.3 and 7; this should be updated after the masking-schedule inconsistency in the major comments is resolved.","section":"Section 9, Table 9"},{"comment":"The paper states that MLLM(Aggregate) is a weighted average, but the exact weighting and the exclusion rule described for Table 6 are not summarized in the main text; readers should be able to interpret the headline MLLM values without consulting the appendix.","section":"Section 4 and supplementary Table 6"}],"recommendation":"major_revision","confidential_remarks":"The editor may wish to ask the authors for code and checkpoints, and for a precise description of the inference protocol used for Tables 1 and 4. The central training-objective claim is plausible but not yet convincing given the single-seed reporting and the small metric gaps; the inference-scaling section requires new matched-baseline experiments. I see no evidence of fabrication, but the internal inconsistencies should be corrected before a further round of review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it. First, the MAgD objective—adding a random mask on top of the noisy latent during training—is a sensible idea, and the ablations show it consistently beats a vanilla OmniGen finetune on DINO and CLIP-Dir. Second, the inference-time scaling story is weaker than it looks: the paper selects the best CLIP-Dir among generations with 0, 8, 16, and 32 pause tokens, and compares against a single generation at 0, so the gain could just be best-of-n. There is no pause=0 multi-sample baseline.\n\nWhat's genuinely new: the specific combination of dual-corruption masking of noisy latents with pause-token inference scaling for diffusion editing, and the synthetic expressive-prompt decomposition for training. The MAgD idea is well motivated by MAE/DINO literature and is cheap to try. The GenEval results are also a nice sanity check that the objective doesn't hurt generation.\n\nThe soft spots are real but mostly fixable. The headline improvements are small (CLIP-Dir 0.122 to 0.126) and come from a single seed with no error bars; that is the biggest problem for the training-objective claim. The pause-token section is more seriously confounded: without a pause=0 multi-sample baseline you cannot attribute the gains to added capacity. On top of that, Eq. (6) and the text contradict each other: the equation applies masking when t < τ_MAgD, while the text says masking is restricted to t ≥ τ_MAgD. That needs to be resolved. Also, the paper says ~450K training samples in one place and 400K in Table 9, and 4000 vs 3200 gradient steps; these inconsistencies matter for reproducibility.\n\nI would send this to peer review, but only with the expectation of major revision. The core training objective is worth testing, and the pause-token idea is interesting enough to warrant a proper control. A serious referee should ask for multi-seed statistics, the missing baseline, and a corrected objective definition. The paper is not ready to be taken at face value, but it is not a waste of anyone's time.","headline":"MAgD is a plausible training trick with consistent but small gains, but the pause-token scaling claim is confounded by best-of-n, and the paper needs error bars and a fix to its own objective equation.","tokens_in":18545,"tokens_out":3973,"would_cite":false,"duration_ms":41349,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that masking-augmented diffusion training plus inference-time pause tokens jointly improve visual editing faithfulness and instruction adherence.","keywords":["visual editing","diffusion models","masked reconstruction","inference-time scaling","pause tokens","instruction following","image faithfulness"],"falsifier":"Fine-tune the same base architecture on the same corpus twice, once with the MAgD objective and once with the standard denoising objective, holding data, steps, and seeds fixed; if the DINO and CLIP-Dir gaps disappear, the masking objective is not the cause. For pause tokens, generate multiple samples per prompt with 0, 8, 16, and 32 pause tokens and apply the paper's best-CLIP-Dir selection within each condition; if the pause=0 best-of-n matches the best pause-token CLIP-Dir, the scaling effect reduces to best-of-n sampling.","tokens_in":17506,"feed_emoji":"🎨","tokens_out":6466,"duration_ms":68187,"temperature":0.7,"pith_summary":"This paper tries to establish that diffusion models for visual editing can be made more controllable by changing both how they are trained and how they are run. It introduces MAgD, a training objective that adds a masked-reconstruction term to standard denoising score matching by masking the noisy input during the forward process at high noise levels. It also introduces an inference-only mechanism in which pause tokens are inserted into the prompt, giving the model extra computation before it generates the edited image. The paper reports that these mechanisms together improve both source-image faithfulness and instruction adherence, and that pause-token scaling lets practitioners trade one against the other without retraining. A sympathetic reader would care because editing is an inverse problem where faithfulness and instruction-following usually compete, and the paper claims to relax that trade-off.","feed_headline":"Masked training plus pause tokens improves diffusion editing","feed_subtitle":"Dual-corruption training boosts source-image fidelity; pause tokens steer the faithfulness-versus-instruction trade-off.","key_machinery":"The central object is the MAgD objective, which applies a random binary mask to the noisy latent $x_t$ from the forward process, replacing masked entries with a mask embedding, and trains the denoiser with the standard score-matching loss on this dual-corrupted input. Masking is applied stochastically (with probability $p_{\\mathrm{magd}}$) and only at high noise levels ($t \\ge \\tau_{\\mathrm{MAgD}}$), which the paper argues targets the regime where the network builds global structure. The second mechanism is the pause token $\\langle\\mathrm{pause}\\rangle$, inserted at inference time between the instruction or reference image and the generation segment, which effectively grants extra transformer computation before output. Together, the masking objective is intended to instill localized, compositional representations, and the pause tokens are intended to let the model spend more compute on hard edits.","core_discovery":"MAgD's dual-corruption objective—standard Gaussian denoising plus stochastic masking of the noisy input at high noise levels—produces visual representations that are more compositional and more finely grounded in text, so that a unified in-context diffusion model performs better at instruction-guided editing. The paper also claims that pause tokens, special placeholder tokens inserted into the prompt only at inference, give the model additional computational capacity that improves instruction adherence, and that by varying the number of pause tokens a user can navigate the trade-off between source-image faithfulness and instruction following without retraining. Empirical support comes from fine-tuning a single base architecture on about 400K samples and reporting that MAgD raises DINO from 0.882 to 0.927 and CLIP-Dir from 0.122 to 0.126 on Emu-Edit, with pause-token scaling further raising CLIP-Dir to 0.134 at a small DINO cost.","pith_inferences":["Beyond the paper: if pause tokens add real capacity, then holding total compute fixed, a run with pause tokens should beat a run with extra denoising steps of the same cost; this comparison is not in the paper.","Beyond the paper: the high-noise-only masking schedule suggests MAgD's benefit should concentrate on edits that require global restructuring, such as background or global-style changes, and be smaller for localized texture swaps; a task-sliced ablation would test this.","Beyond the paper: because pause-token scaling is inference-only, it could be stacked with other zero-retraining controls such as guidance-scale adjustment to trace a full Pareto frontier of faithfulness versus instruction adherence per prompt."],"forward_implications":["If MAgD's representation gains are real, the same training objective should transfer to any unified in-context diffusion architecture, improving image editing without architectural changes.","Pause-token scaling would let practitioners select an operating point on the faithfulness-instruction trade-off at inference, with no retraining, and the paper reports higher recall of edits meeting a DINO $\\ge 0.91$ target when scaling is used.","Fine-tuning with expressive, step-wise prompts appears to further improve instruction adherence, and the gains are larger when combined with MAgD than with the base objective.","The dual-corruption objective also improves scores on compositional text-to-image benchmarks such as GenEval, suggesting the representation benefit is not limited to editing."],"supporting_citations":[{"why":"Provides the base in-context diffusion architecture that is fine-tuned; all comparisons inherit from it.","marker":"[23]"},{"why":"Establishes masked reconstruction as a representation-learning strategy that MAgD adapts to the diffusion forward process.","marker":"[8]"},{"why":"Supplies the 250K-sample high-quality editing training data used to build the 400K corpus.","marker":"[29]"},{"why":"Supplies a manually annotated editing dataset used both for training and for benchmark evaluation.","marker":"[28]"},{"why":"Provides the Emu-Edit benchmark and task suite used for the main quantitative comparisons.","marker":"[17]"},{"why":"Defines the Complex-Edit benchmark with instruction-following and identity-preservation protocol reused in evaluation.","marker":"[26]"},{"why":"Supplies the GenEval compositional text-to-image benchmark used to test generalization beyond editing.","marker":"[6]"},{"why":"Introduces pause tokens, the mechanism adapted for inference-time capacity scaling.","marker":"[7]"}],"fun_headline_variants":["Masking and pause tokens boost diffusion editing","Dual corruption training improves edit fidelity","Pause tokens let users steer edit faithfulness","MAgD: masking-augmented diffusion for better edits","Inference-time scaling enhances diffusion editing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim would collapse if the measured improvements came from the extra fine-tuning data or from picking the best of several random samples, rather than from the masking objective and pause tokens themselves.","fun_headline_variants_meta":{"raw":{"variants":["Masking and pause tokens boost diffusion editing","Dual corruption training improves edit fidelity","Pause tokens let users steer edit faithfulness","MAgD: masking-augmented diffusion for better edits","Inference-time scaling enhances diffusion editing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1505,"prompt_tokens":960,"completion_tokens":545,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":477}},"tokens_in":576,"tokens_out":545,"duration_ms":5918,"temperature":1.0,"reasoning_tokens":477,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:42:14.496863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune the same base architecture on the same corpus twice, once with the MAgD objective and once with the standard denoising objective, holding data, steps, and seeds fixed; if the DINO and CLIP-Dir gaps disappear, the masking objective is not the cause. For pause tokens, generate multiple samples per prompt with 0, 8, 16, and 32 pause tokens and apply the paper's best-CLIP-Dir selection within each condition; if the pause=0 best-of-n matches the best pause-token CLIP-Dir, the scaling effect reduces to best-of-n sampling.","supporting_citations":[{"cited_title":"Complex-Edit: Cot-like instruction generation for complexity-controllable image editing benchmark, 2025","cited_arxiv_id":null,"evidence_quote":"Defines the Complex-Edit benchmark with instruction-following and identity-preservation protocol reused in evaluation."}],"review_version":1}