{"id":"216b70a5-9377-4718-8574-94e95d65018d","arxiv_id":"2608.01288","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TurboClear combines region-calibrated distribution matching and learnable spatial fusion to turn a multi-step object-removal diffusion model into a one-step student with comparable or better quality and up to 665x lower FLOPs.","lead":"TurboClear is a one-step image-removal model that erases an object and its shadows or reflections from a photo in a single neural-network pass. It reports matching or better visual quality than multi-step models while cutting inference cost by tens to hundreds of times.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The efficiency speedup is well supported, but the paper never measures removal quality inside the effect mask m_e. The central 'comparable or better' claim is therefore not yet directly verified for the effect component of object-effect removal.","rationale":"The reader's weakest_assumption focused on the training-time dependency on accurate effect masks m_e, which the paper discloses in Limitations H. My concern is the complementary evaluation side: even granting accurate m_e during training, the paper does not report any test-time metric defined against m_e. Because PSNR-M uses only the object mask and local LPIPS uses the object-mask bounding box, the specific quantity the method is designed to improve, removal of visual effects beyond the object mask, is never isolated. This matters for the strongest claim because LSF's background-copying can inflate preservation metrics without improving effect removal. The speedup and latency evidence, plus the paired CIs and user study in Appendix F, are otherwise solid; the additional effect-region metrics would settle whether the central quality claim is truly about effects or mostly about background preservation. I therefore recommend accepting the paper conditionally on this evaluation check, not rejecting it.","tokens_in":18261,"tokens_out":17064,"duration_ms":162452,"concrete_test":"On OBER-Test, compute PSNR and LPIPS restricted to the effect mask m_e, and separately to the annular region of m_e outside the object mask M_o, for TurboClear, ObjectClear, FlashClear, and OmniPaint using the paper's evaluation protocol and the same 163 samples. If TurboClear is not within the paired 95% bootstrap CI of ObjectClear on these effect-region metrics, the 'comparable or better visual removal quality' claim should be narrowed to background-preservation and object-mask fidelity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"TurboClear's method is explicitly calibrated to the object-effect mask m_e: RDM masks and normalizes the distribution-matching gradient inside m_e (Eqs. 6-9), the paired losses in Eq. 10 split on m_e, and LSF supervision and gate regularization in Eqs. 17-18 depend on it. Yet every reported quality metric in the main and supplementary evaluations ignores m_e. PSNR-M (Appendix A, Eq. 20) is computed only inside the object mask M_o; LPIPS-L uses the bounding box B(M_o); PSNR-BG in Table 5 is the complement of M_o; and DISTS-Local uses B(M_o). A model that leaves shadows, reflections, or occlusion traces just outside M_o can therefore score well on PSNR-M, and such residuals are diluted in global metrics. Since the central claim is 'comparable or better visual removal quality' for object-effect removal, not just object-mask inpainting, the evaluation does not directly test the effect-removal component. In particular, LSF learns spatial gates that can copy the original image in regions classified as safe background, so a residual effect outside M_o could be preserved while reported metrics improve because background copying raises PSNR-BG. This is an internal evaluation gap, not an external disagreement; it can be settled with the existing per-sample effect masks on OBER-Test.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"TurboClear proposes a one-step SDXL-based object-effect removal model that combines Region-Calibrated Distribution Matching (RDM) for distillation with a Learnable Spatial Fusion (LSF) module for inference-time blending. The student is distilled from the multi-step ObjectClear teacher, using the object-effect mask m_e as privileged training supervision to mask and normalize distribution-matching gradients and to supervise paired reconstruction and localization losses. LSF learns spatial gates that fuse the one-step removal stream with an identity-preserving reference stream. The paper reports large denoising FLOPs reductions and latency speedups relative to ObjectClear and OmniPaint, with quantitative results on OBER-Test and RORD-Val, ablations, a double-blind user study, and an appendix with paired bootstrap confidence intervals and a function-class analysis for LSF.","tokens_in":18546,"tokens_out":3286,"duration_ms":30172,"significance":"If the central claim is fully established, TurboClear would be a practically significant contribution: it demonstrates that one-step distillation with region-calibrated objectives and learned spatial fusion can retain the asymmetric edit-and-preserve behavior required for object-effect removal while cutting inference cost by one to two orders of magnitude. The paper has notable strengths: the method is clearly motivated, the ablations separate the effects of distillation strategy and fusion strategy, paired bootstrap intervals are provided in the appendix, the user study is double-blind, and the Appendix E analysis is a parameter-free function-class argument. The public code release is also a concrete reproducibility asset. The main weakness is that the evaluation never directly measures removal quality inside the object-effect mask m_e, even though the method's training objectives are explicitly calibrated to that mask; this gap currently prevents the central 'comparable or better visual removal quality' claim from being fully verified for the effect component of the task.","major_comments":[{"comment":"All reported removal-quality metrics are computed on the object mask M_o or its bounding box, never on the object-effect mask m_e that defines the task. Equation (20) computes PSNR-M inside M_o, LPIPS-L uses the bounding box B(M_o), PSNR-BG is evaluated on the complement 1-M_o, and DISTS-Local uses B(M_o). Since RDM masks and normalizes the distribution-matching gradient with m_e (Eqs. 6-9), the paired losses split on m_e (Eq. 10), and the LSF supervision and gate regularization depend on m_e (Eqs. 17-18), the 'effect' component of object-effect removal is never directly measured. A student that leaves shadows, reflections, or occlusion traces just outside M_o could score well on PSNR-M and its residuals would be diluted in global metrics. I request that the authors add effect-mask metrics on OBER-Test, e.g., PSNR and LPIPS computed inside m_e or inside m_e \\ M_o, using the existing per-sample object-effect masks. Without such an evaluation, the central claim of 'comparable or better visual removal quality' for object-effect removal is not directly supported for the effect region.","section":"Experiment (Evaluation protocol) and Appendix A (Quality metric definitions)"},{"comment":"The LSF design and its gate regularization L_alpha (Eq. 18) explicitly encourage copying the input in regions classified as safe background, i.e., alpha_x near zero outside the dilated effect mask m_e^+. If m_e is imperfect or if visual effects extend beyond it, the learned gate can preserve residual effects in the output while actually improving metrics such as PSNR-BG by copying the original background. This is not an internal inconsistency, but it is a deployment risk directly tied to the quality of m_e annotations, and the paper itself notes in Section H that obtaining object-effect masks for new domains is expensive. The proposed m_e-based metrics would directly quantify whether TurboClear removes effects inside the annotated effect region and whether residuals leak outside it; without them, the model's behavior on real images where m_e is unavailable or inaccurate remains untested.","section":"Method (Learnable Spatial Fusion) and Limitations (Section H)"}],"minor_comments":[{"comment":"The headline speedup factors (40.04x vs. ObjectClear, 665x vs. OmniPaint) are based on theoretical denoising FLOPs that explicitly exclude condition encoders, VAE encoding/decoding, and the LSF head. This is transparently stated, but readers may conflate 'denoising FLOPs' with total inference cost; the latency numbers in Table 4 are more comprehensive and already support the speedup claim, so I suggest adding an end-to-end FLOPs figure or clearly labeling the headline numbers as backbone-only.","section":"Appendix A (Denoising FLOPs measurement)"},{"comment":"The fixed one-step DDIM timestep 399, the fixed prompt 'remove the instance of object', and object-token index 5 are important implementation choices that are reported only in the appendix; mentioning them in the main Experiments section would make the method easier to reproduce.","section":"Appendix B (One-step warm-up and initialization)"},{"comment":"In the comparison against OmniPaint, TurboClear is worse on DISTS-Local and MUSIQ (Table 5, with CIs in Table 6 that exclude zero for DISTS-L). The abstract's phrase 'comparable or better visual removal quality' is broadly supported by the other metrics, but a sentence acknowledging this trade-off would improve precision.","section":"Table 5 and Table 6"}],"recommendation":"major_revision","confidential_remarks":"The evaluation gap is fixable with annotations that the authors already have on OBER-Test, and the paper's engineering is otherwise solid. I recommend major revision rather than rejection because nothing in the present evidence suggests the trained model actually fails on effect regions; the issue is that the current quantitative evaluation does not directly measure the effect component of the central claim. Adding a small set of m_e-based metrics would resolve the concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid take: this is a genuinely useful one-step distillation paper, not a fake speedup. The new machinery (RDM, LSF) is incremental — region-masked DMD and learned dual-stream blending — but the combination lands in a previously empty regime: one-step object-effect removal with an SDXL inpainting teacher. The empirical package is stronger than most papers in this area: main tables, ablations that separate distillation from fusion, paired bootstrap CIs, a double-blind user study, and code. The Appendix E argument that attention alone cannot recover the optimal fusion gate is a real, parameter-free piece of analysis. The reported efficiency gain is credible because latency is reported alongside theoretical FLOPs, and the latency speedup (55x vs ObjectClear) is actually larger than the FLOPs number quoted in the abstract.\n\nThe main soft spot is the one the stress-test flags. Every fidelity metric that uses a mask uses the object mask M_o, not the object-effect mask m_e. So the 'effect' part of object-effect removal is never directly measured. That matters more than usual here because RDM, the paired losses, and LSF supervision all depend on m_e. The paper does report PSNR-BG, which includes the effect region outside M_o, so a residual shadow would hurt that number; and the user study asks about 'object remnants' holistically. But PSNR-BG over the whole complement dilutes small residuals, and the user study is perceptual rather than region-specific. This is an internal evaluation gap, not a sign the method is wrong. It can be settled in a revision by computing PSNR and LPIPS inside m_e (or m_e \\ M_o) on OBER-Test, where these masks already exist. I'd ask for that.\n\nOther issues are minor. The headline FLOPs exclude encoders, VAE, and the fusion head; the paper is transparent about this and latency includes them, so it's fine. The method depends on accurate m_e and a fixed prompt with a hardcoded object-token index; Appendix H discloses both. The FlashClear overlap is cited but not explicitly disclosed as the authors' own prior work; I'd want that stated.\n\nBottom line: the central efficiency claim is solid, and the quality claim is well supported by everything except the one missing region-specific metric. The paper deserves a serious referee. Send it out; ask for an effect-mask evaluation and a conflict-of-interest note.","headline":"Solid one-step object-effect removal with real speedups; the evaluation mostly supports it, but the central claim would be stronger with a metric computed inside the effect mask.","tokens_in":19088,"tokens_out":2868,"would_cite":true,"duration_ms":25423,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TurboClear claims that object-effect removal can be compressed into a single denoising step, matching multi-step quality at up to 665x lower computational cost.","keywords":["object-effect removal","one-step distillation","region-calibrated distribution matching","learnable spatial fusion","SDXL inpainting","diffusion model acceleration","spatial asymmetry","image editing"],"falsifier":"Take a set of paired images where a visible shadow or reflection extends beyond the annotated object mask, train or fine-tune TurboClear with an object-effect mask deliberately eroded to exclude that effect, and check whether the output still contains the residual effect; if it does, the central removal claim fails in exactly the regime the mask is supposed to cover.","tokens_in":18007,"feed_emoji":"⚡","tokens_out":5770,"duration_ms":46924,"temperature":0.7,"pith_summary":"This paper tries to establish that object-effect removal—erasing a target object along with its shadows, reflections, and occlusion traces—can be compressed into a single denoising step without losing the task's asymmetric behavior, where affected pixels regenerate while unaffected pixels stay untouched. The authors argue that generic one-step distillation methods fail because their objectives are global and treat the whole image uniformly, so they design a region-calibrated objective that applies distribution-matching pressure only inside the object-effect mask. They also introduce a learned spatial fusion module that blends the generated removal stream with the original input image using per-pixel gates learned from local evidence. If the claim holds, real-time object removal becomes practical on consumer hardware while preserving removal quality.","feed_headline":"One-step eraser removes objects and effects at 665x lower cost","feed_subtitle":"TurboClear distills multi-step diffusion removal into a single pass while preserving background fidelity.","key_machinery":"The load-bearing object is the object-effect mask $m_e$, used only in training, which defines where removal happens. RDM is a masked and normalized distribution-matching gradient $g_{\\text{RDM}} = m_e^\\ell \\odot g_{\\text{DMD}} / (s+\\varepsilon)$ that aligns affected regions with the teacher's generative distribution while keeping background anchored via paired reconstruction losses. LSF is a lightweight convolutional head that predicts residual corrections to a logit-transformed attention prior, producing per-pixel gates $\\alpha_z$ and $\\alpha_x$ that blend the one-step latent prediction $\\hat{z}_0$ with the input latent $z_y$, then blend the decoded result with the original image; the LSF training loss uses $m_e$ supervision and gate regularization to force high generation weight inside the effect region and low weight in safe background. Together these carry the asymmetry argument.","core_discovery":"The central claim is that TurboClear, an SDXL-based one-step student distilled from the multi-step model ObjectClear, removes both the masked object and its associated visual effects in a single UNet evaluation while keeping background content unchanged, and does so at roughly 27–40x lower denoising computation than ObjectClear and 628–665x lower than OmniPaint, with comparable or better PSNR, LPIPS, and local metrics on OBER-Test and RORD-Val. The paper argues that the two ingredients—Region-Calibrated Distribution Matching (RDM), which masks the distribution-matching gradient to the effect region, and Learnable Spatial Fusion (LSF), which learns pixel- and latent-level gates between the generated stream and the input stream—are what preserve the edit-and-preserve asymmetry, and ablation comparisons against LCM, DMD2, RAD, and attention-guided fusion support that division of labor.","pith_inferences":["The same region-calibrated distillation and learned two-stream gating could transfer to other asymmetric editing tasks, such as object insertion, style transfer confined to a subject, or text-guided local editing, where one region must change while the rest must stay identical; the paper does not test this.","A weak-supervision variant that discovers the effect mask automatically, for example from shadow and reflection cues, would remove the annotation bottleneck the paper acknowledges in Section H; whether RDM degrades gracefully under noisy $m_e$ is an open testable question.","Because peak memory is still 7,962 MiB, one-step distillation alone does not solve edge deployment; combining RDM/LSF with quantization or structured pruning is a direct extension the paper leaves implicit.","RDM's normalization term makes the gradient scale comparable across different effect-region sizes, which implies the objective could transfer to datasets with widely varying object sizes without re-tuning; the paper does not ablate this."],"forward_implications":["One-step inference with no classifier-free guidance at test time brings per-image denoising FLOPs down to 1.59 T at 512x512 and 3.21 T at 960x540, which would allow object-effect removal in roughly 0.04 seconds on an A800 and open the door to real-time deployment.","RDM-type masking of the distribution-matching gradient is reported to beat generic one-step objectives (LCM, DMD2, DMD2+GAN, RAD) on both global and local fidelity, suggesting that region calibration, not just the distillation recipe, drives one-step removal quality.","LSF is shown to outperform attention-guided fusion under the same RDM student, with strict confinement of change to the effect region, so learned gating from local stream evidence is a better fusion rule than attention alone.","TurboClear generalizes from 512x512 training to 960x540 inputs, with quality metrics on RORD-Val remaining competitive, indicating the one-step student does not overfit the training resolution.","The paper's user study says TurboClear matches or beats its teacher ObjectClear in 69.0% of pairwise judgments, so the acceleration is not bought by a systematic perceptual regression."],"supporting_citations":[{"why":"Supplies the teacher model (ObjectClear) that the one-step student is initialized from and distilled against, the AGF attention-fusion baseline, and the OBER-Test benchmark with its object-effect masks.","marker":"(Zhao et al. 2026)"},{"why":"Defines the distribution-matching distillation formulation and the fake-score model that RDM modifies with region masking.","marker":"(Yin et al. 2024a)"},{"why":"OmniPaint, the Flux-based multi-step method used as the strongest quality and efficiency baseline in the 628-665x comparison.","marker":"(Yu et al. 2025)"},{"why":"FlashClear, the four-step distilled removal method that TurboClear must beat in the one-step regime.","marker":"(Tang et al. 2026)"},{"why":"SDXL, the architecture underlying the student, teacher, and VAE used throughout.","marker":"(Podell et al. 2024)"},{"why":"Provides RORD-Val, the 960x540 high-resolution evaluation split used to test generalization.","marker":"(Sagong et al. 2022)"},{"why":"Latent Consistency Model, a generic one-step distillation baseline shown in ablations to underperform RDM.","marker":"(Luo et al. 2023)"},{"why":"LPIPS, the perceptual metric used both as a training objective and as a global and local quality measure.","marker":"(Zhang et al. 2018)"}],"fun_headline_variants":["TurboClear: one-step object-effect removal with up to 665x speedup","One-step diffusion eraser cuts compute by up to 665x for object and effect removal","Region-calibrated fusion enables one-step erasure of objects and effects","TurboClear: single-step removal, 40x cheaper than ObjectClear, 665x vs OmniPaint","One-step object-effect removal via region-calibrated distillation, up to 665x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training pipeline assumes the object-effect mask $m_e$ is available and accurately covers the object plus every visual effect it causes; if that mask misses a shadow, a reflection, or an occlusion trace, the student is never taught to remove those pixels and may anchor them as background.","fun_headline_variants_meta":{"raw":{"variants":["TurboClear: one-step object-effect removal with up to 665x speedup","One-step diffusion eraser cuts compute by up to 665x for object and effect removal","Region-calibrated fusion enables one-step erasure of objects and effects","TurboClear: single-step removal, 40x cheaper than ObjectClear, 665x vs OmniPaint","One-step object-effect removal via region-calibrated distillation, up to 665x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001401,"raw_usage":{"total_tokens":5656,"prompt_tokens":932,"completion_tokens":4724,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":4608}},"tokens_in":548,"tokens_out":4724,"duration_ms":27526,"temperature":1.0,"reasoning_tokens":4608,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:09:52.633925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of paired images where a visible shadow or reflection extends beyond the annotated object mask, train or fine-tune TurboClear with an object-effect mask deliberately eroded to exclude that effect, and check whether the output still contains the residual effect; if it does, the central removal claim fails in exactly the regime the mask is supposed to cover.","supporting_citations":[],"review_version":2}