{"id":"22a6b624-6228-4b41-9527-5a365e2bde79","arxiv_id":"2505.12427","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DragLoRA improves drag-based image editing by online-optimizing LoRA adapters with a dual drag and score-distillation loss, adaptive mode switching, and input feature adaptation.","lead":"This paper presents DragLoRA, a method for drag-based image editing that fine-tunes lightweight LoRA adapters inside a diffusion model while the edit runs, instead of optimizing the image latent directly. It reports better point-movement accuracy and lower runtime than existing drag-editing methods on the DragBench benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reversed DDS loss (Eq. 8) is claimed to regularize LoRA toward the original model, but the gradient direction is the opposite of a consistency norm and is asserted without derivation; since DOO is the largest ablation contributor (MD 48.55→36.88), the central mechanism is unsupported.","rationale":"The paper's strongest claim is that optimizing LoRA online with a 'dual-objective' drag+DDS loss yields state-of-the-art precision and efficiency. The DDS term is not incidental: it is the only mechanism preventing the LoRA from drifting off the image manifold, and the only component whose theoretical justification is absent. The appendix explicitly acknowledges the sign is reversed relative to DDS's original use but asserts, rather than shows, that this converts an editing gradient into a regularizer. In a standard consistency regularizer, the gradient direction should reduce the squared difference between the two models; here the direction is opposite and the Jacobian is truncated to the z0 path. The effect of such a term is not obvious, and it could just as plausibly act as an amplifier of drift that happens to trade off against the drag loss. The ablation (Table 3) shows DOO is the largest contributor to the headline MD, so any uncertainty about DOO directly affects the central claim. A code check plus a sign-flipped or true-consistency ablation would settle whether the mechanism is as described. The reader's weakest-assumption (ASS thresholds) is a robustness issue about the efficiency-precision tradeoff, but it is secondary: even if thresholds are suboptimal, the method would still work with more steps; if DDS is mis-specified, the method's stability itself is unexplained.","tokens_in":15241,"tokens_out":21473,"duration_ms":204890,"concrete_test":"In the released code, verify the implemented DDS term matches Eq. (8). Then, on 20 DragBench images, replace Eq. (8) with the exact gradient of E[||ϵ_drag − ϵ_ori||^2] (full backprop through Δθ, no z0-only truncation), keeping all hyperparameters fixed. If MD or 1-LPIPS changes by more than the reported margin, the reversed-DDS term is not acting as a consistency regularizer; if performance is unchanged, the sign is not the operative component.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is the dual-objective optimization (DOO) with a delta denoising score (DDS) term. Eq. (8) defines ∇Δθ L_DDS = (ϵ_ori − ϵ_drag) ∂z0/∂Δθ, and Appendix B.1 states this is 'intentionally reversed' relative to the standard DDS editing gradient. The paper claims this term regularizes the online LoRA to stay close to the pretrained model. However, the gradient of a true consistency loss E[||ϵ_drag − ϵ_ori||^2] w.r.t. Δθ requires the Jacobian of ϵ_drag through the LoRA weights; by contrast, Eq. (8) only backpropagates through the predicted clean latent z0 and uses the opposite sign. No derivation or geometric argument is given for why a reversed DDS term behaves as a repulsive regularizer rather than a force that amplifies the divergence. This is not merely a formal issue: Table 3 shows DOO alone improves Mean Distance from 48.55 to 36.88 (≈12 points), the largest single gain among the modules. If the term is not in fact pulling the model toward the original, the reason DragLoRA avoids drift is unexplained, and the claimed advantage over latent-based optimization could be an artifact of the extra re-noising steps or the larger parameter budget (3.19M vs 0.07M). Since the code is public, this is checkable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DragLoRA, a drag-based image editing method for Stable Diffusion in which a LoRA adapter is optimized online during editing, replacing the conventional optimization of the DDIM-inverted latent. The method combines a drag loss with a delta denoising score (DDS) regularization, adapts the input latent through a denoise-renoise cycle (ILFA), and switches adaptively between LoRA optimization and input-adaptation-only modes based on point-tracking quality. An efficient point-tracking strategy is also introduced. On DragBench, the method reports the best Mean Distance (23.77) among compared methods, a runtime of 29.84 seconds, and favorable drag-back LPIPS (1.33), with ablations attributing the gains to the proposed modules.","tokens_in":15704,"tokens_out":18396,"duration_ms":164656,"significance":"The central idea of optimizing a low-rank adapter online rather than the latent feature is timely and plausible; it offers an appealing efficiency-precision trade-off for interactive drag editing. The paper's strengths include a broad comparison with recent drag-editing methods, a clean ablation sequence (Table 3), an independent drag-back evaluation, and publicly released code. If the results are robust, the contribution is a solid incremental advance over DragDiffusion, with notably faster runtime than GoodDrag at comparable precision. The main caveats are that the gains over the best baseline are modest (0.49 MD), the reported metrics lack error bars, and two load-bearing design choices (the DDS sign and the ASS switching signal) require more rigorous justification.","major_comments":[{"comment":"The DDS regularization is the largest single contributor in Table 3 (MD improves from 48.55 to 36.88), yet the paper provides no derivation showing that the reversed gradient (ϵ_ori − ϵ_drag) ∂ẑ0/∂Δθ in Eq. (8) pulls the LoRA toward the original model. Appendix B.1 only states that the sign is intentionally reversed relative to the original DDS editing loss. Because the gradient is taken only through ẑ0, the effective sign depends on the chain rule through the DDIM estimate ẑ0 = (z_t − √(1−α_t) ε_drag(z_t,t))/√(α_t); a proper derivation would make clear whether this is a consistency regularizer or an anti-regularizer. Without this derivation, the mechanism behind the largest ablation gain is unsupported. Please provide the derivation and state explicitly which terms are detached (stop-gradient) in the implementation.","section":"Section 3.2, Eq. (8), Appendix B.1"},{"comment":"The adaptive switching scheme (ASS) uses the point-tracking quality metric minD (Eq. 7) to decide when to skip LoRA optimization, but the paper's own ablation undermines this signal. The version without ASS and EPT (the +ILFA row in Table 3) achieves a lower minD per Appendix A and Figure 7, yet has worse MD (26.99) than the full DragLoRA (23.77). This contradicts the statement in Section 3.1 that lower minD indicates 'the success of the previous optimization for motion supervision.' Please clarify what minD actually measures in the ASS context, report the correlation between minD and final MD on DragBench, and provide a sensitivity analysis for the hand-set thresholds (d1=1, d2=1.3, l1=1, l2=1.4, k_ini=10). As it stands, the 1.68-point MD gain attributed to ASS is not robustly established.","section":"Section 3.4, Eq. (7), Table 3, Appendix A"},{"comment":"The comparison with DragDiffusion is not iso-parameter: Table 1 lists 3.19M optimized parameters for DragLoRA versus 0.07M for DragDiffusion, while Section 3.2 claims 'the model size unchanged compared to (Shi et al., 2024b).' Please clarify this discrepancy and discuss whether the improved accuracy over DragDiffusion could be partly due to the larger LoRA capacity rather than the algorithmic contributions. An iso-capacity baseline (e.g., a DragDiffusion variant with a rank-16 reconstruction LoRA) would strengthen the claim that the improvements come from the proposed optimization scheme rather than from parameter count.","section":"Section 3.2, Section 4.1, Table 1"}],"minor_comments":[{"comment":"The CLIP column values (9.47, 9.8, etc.) exceed the usual range for CLIP similarity; if these are scaled by 10, add a '(×10)' label as is done for LPIPS in Table 2.","section":"Table 3"},{"comment":"The variable n (the temporal target) is used in the if condition before it is first assigned in the else branch; please initialize or reorder the pseudocode for clarity.","section":"Algorithm 1"},{"comment":"The statement that DragLoRA 'outperforms existing approaches in both perceptual quality and the accuracy of drag editing' is too strong given that GoodDrag reports a better m-MD (21.86 vs 22.70) in Table 1 and a better CLIP score in Table 2; please qualify.","section":"Section 4.2"},{"comment":"The simplification of the denoise-renoise update is not immediately transparent; please define α_t and \\bar{α}_t in the appendix and verify the algebra, or add a derivation step.","section":"Appendix B.2, Eq. (9)"},{"comment":"Specify the number of images averaged for the minD/dT curves and report the standard deviation or shaded intervals.","section":"Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The DDS sign concern raised by the skeptical reviewer is in principle resolvable: with the chain rule through ẑ0, Eq. (8) has the same sign as the gradient of a squared error between ε_drag and ε_ori. However, the paper does not provide this derivation, and the current 'intentionally reversed' wording is actively confusing. The minD/ASS inconsistency is the more serious issue and should be addressed with data. The paper's public code is a strong point and makes these questions empirically checkable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick verdict: DragLoRA is a genuinely useful engineering result. The idea of optimizing LoRA weights online, instead of the inversion latent, is the right kind of twist, and the ablations show each component pulls Mean Distance down from 48.55 to 23.77. The drag-back test gives independent support, and the code is public. I think the paper is worth a serious referee.\n\nThe stress-test note on the reversed DDS loss does not land. The sign in Eq. (8) is exactly what you want if the goal is to keep the LoRA-augmented model consistent with the original: the DDS editing gradient is (epsilon_edit - epsilon_ori), so reversing it gives (epsilon_ori - epsilon_drag), which is an SDS-style update for minimizing the discrepancy between the two noise predictions. The paper calls this 'intentionally reversed' and that is a reasonable empirical choice. The missing Jacobian through epsilon_drag is the standard SDS approximation, not an error. It would be nice if they said so explicitly, but the mechanism is not unsupported.\n\nThe real soft spots are the usual ones for this kind of paper. The ASS thresholds (d1, d2, l1, l2, k_ini) are hand-tuned on DragBench, and there are no error bars or multiple seeds, so the reported 23.77 could be optimistic. The efficiency comparison is slightly confounded: DragLoRA has 3.19M parameters versus 0.07M for the baselines, and the method assumes a reconstruction LoRA has already been trained (they exclude that 48s from the time, which is fair). The m-MD metric is a reasonable addition, but it is derived from the same DIFT features, so it does not fully solve the evaluation-noise problem.\n\nOne thing I appreciated: Appendix A explicitly says ASS and EPT are not designed to minimize minD and that the ablated version achieves lower minD. That is honest and shows they thought about the metric's pathology.\n\nWho would get value: anyone working on interactive drag editing or on per-image adapter optimization. The method is simple to understand, the code is public, and the empirical story is clear. If this crossed my desk for peer review, I would send it out. It does not need another round of methodological proof; it needs a replication with error bars and a sensitivity analysis of the thresholds.\n\nBest,\n[your name]","headline":"Online LoRA optimization for drag editing is a real advance, and the reversed DDS loss is a reasonable SDS-style regularizer, not a load-bearing flaw; the main caveats are hand-tuned thresholds and missing error bars.","tokens_in":16162,"tokens_out":5384,"would_cite":true,"duration_ms":51099,"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":"By optimizing a LoRA adapter online instead of the DDIM-inverted latent, DragLoRA reports the best drag-editing precision on DragBench and a faster runtime.","keywords":["drag-based image editing","LoRA","diffusion models","online optimization","delta denoising score","DDIM inversion","point tracking","image editing"],"falsifier":"Run DragLoRA on a set of images with occlusions or textureless regions where handle points are known to drift, and compare the algorithm's minD-based switch decisions against ground-truth tracking obtained by manual annotation or optical flow. If ILFA-only mode is activated frequently while the tracked handle is actually far from its true location, the switching signal is unreliable and the reported precision-runtime tradeoff would not generalize beyond the benchmark.","tokens_in":15078,"feed_emoji":"🎯","tokens_out":5332,"duration_ms":50965,"temperature":0.7,"pith_summary":"The paper claims that drag-based image editing in diffusion models should be done by optimizing a per-image LoRA adapter online rather than directly optimizing the DDIM-inverted latent feature. It argues that this increases the model's capacity to represent the requested deformation, and pairs the drag loss with a reversed delta denoising score loss that keeps the adapted model close to the pretrained one. An additional denoise-renoise step adapts the input latent to accumulated motion, and an adaptive scheme switches between full optimization and input-only adaptation based on point-tracking quality. On DragBench the method reports the lowest mean-distance error (23.77) among the compared drag methods and a runtime of 29.84 seconds for drag editing, with better drag-back fidelity than DragDiffusion and GoodDrag. If true, interactive point-based editing can be both more accurate and faster without offline training.","feed_headline":"Drag editing gets more precise by tuning LoRA adapters online","feed_subtitle":"Per-image LoRA tuning beats direct latent optimization on DragBench, cutting runtime to under 30 seconds.","key_machinery":"The load-bearing object is the per-image LoRA adapter, a low-rank parameter update added to the UNet's attention layers and initialized from a reconstruction LoRA; it is optimized online during dragging. The drag loss drives handle points toward targets, while the reversed DDS loss regularizes the adapter by penalizing divergence of its noise predictions from the pretrained model. Input Latent Feature Adaptation is a cyclic denoise-renoise update that propagates accumulated deformation into the input latent, making motion supervision stable. The Adaptive Switching Scheme, implemented in Algorithm 1, gates between costly LoRA optimization (DOO+ILFA) and cheap input-only adaptation (ILFA-only) using minD thresholds $d_1=1$, $d_2=1.3$ and distance thresholds $l_1=1$, $l_2=1.4$. Efficient Point Tracking restricts candidate handle locations to distance-closer or angle-closer regions, reducing the search space and avoiding error-prone reverse-direction points.","core_discovery":"The central claim is that replacing direct latent-feature optimization with online optimization of LoRA parameters—inserted into all attention layers of the UNet—gives drag editing finer control and better fidelity. DragLoRA detaches deformation control from the static inverted latent, so motion supervision acts on a more expressive parameter space. To keep the adapter from drifting, the paper minimizes the drag loss plus a mask loss plus a DDS loss whose gradient is $\\nabla_{\\Delta\\theta} L_{\\mathrm{DDS}} = (\\epsilon_{\\mathrm{ori}}-\\epsilon_{\\mathrm{drag}})\\,\\partial \\hat{z}_0/\\partial \\Delta\\theta$, reversing the original DDS direction and using it as a regularizer. Input Latent Feature Adaptation cycles the input latent through one DDIM denoise and one DDPM re-noise step inside the mask, feeding the accumulated deformation back into the input. An adaptive optimization scheme then switches between DOO+ILFA and ILFA-only based on the point-tracking confidence minD and the distance of the tracked point from its temporal target; this configuration reports the best mean distance of 23.77, a runtime of 29.84 seconds, and the lowest drag-back LPIPS of 1.33 among the compared methods on DragBench.","pith_inferences":["The hand-set switching thresholds in Algorithm 1 are a likely source of brittleness; an automatic calibration of $d_1$, $d_2$, $l_1$, and $l_2$ from tracking-confidence statistics could make the efficiency-precision tradeoff hold across datasets and handle-target configurations without manual re-tuning.","The reversed DDS regularizer is a transferable recipe for other single-image adapter tuning tasks, not just drag editing: any online adaptation of a diffusion model to one input could use the same noise-prediction discrepancy to stay close to the pretrained prior.","Because ILFA's update is essentially a SDS-style guidance on the input latent, testing it on other latent-optimization drag methods may reveal whether random-noise renoising, rather than exact inversion, is the key to stability in motion supervision.","The minD signal used for switching could also be exposed to the user as a per-point tracking confidence, enabling interactive systems to warn when a handle is stuck and to suggest where more careful optimization is needed."],"forward_implications":["Drag editing can become both more accurate and faster by using a per-image LoRA adapter that requires no offline training, since many steps skip backpropagation and use input adaptation only.","The reversed DDS loss offers a general way to regularize any per-image adapter tuning against the pretrained model, preventing unrestricted updates from degrading fidelity.","ILFA is portable and is shown to improve another drag method, DragNoise, suggesting that cyclic denoise-renoise input adaptation can benefit latent-optimization approaches beyond DragLoRA.","Restricting point tracking to distance-closer or angle-closer candidate regions, with a confidence check based on minD, reduces wasted computation and avoids misleading tracked points in ambiguous regions.","The reported MD of 23.77 and runtime of 29.84 seconds on DragBench position DragLoRA as the most accurate optimization-based drag method with a competitive runtime among the evaluated approaches."],"supporting_citations":[{"why":"Provides the base drag-editing pipeline, the DDIM-inversion latent at t=35, reconstruction LoRA initialization, and the DragBench benchmark used for comparison.","marker":"(Shi et al., 2024b)"},{"why":"Supplies the delta denoising score loss that DragLoRA reverses into a regularizer for the online LoRA optimization.","marker":"(Hertz et al., 2023)"},{"why":"Introduces the motion supervision and point tracking paradigm that DragLoRA builds on for drag-based editing.","marker":"(Pan et al., 2023)"},{"why":"Provides the Stable Diffusion latent diffusion base model in which the LoRA adapters are inserted.","marker":"(Rombach et al., 2022)"},{"why":"Defines the DDIM sampling and deterministic inversion equations used to obtain the editable latent and to perform the denoise step in ILFA.","marker":"(Song et al., 2020a)"},{"why":"Establishes the score distillation sampling formulation that ILFA is compared against and connected to in the appendix.","marker":"(Poole et al., 2022)"},{"why":"Provides the drag-back evaluation protocol and the linear point-search strategy that DragLoRA compares with in its efficient point tracking experiments.","marker":"(Ling et al., 2024)"},{"why":"Provides the distance-closer and angle-closer candidate regions used by DragLoRA's efficient point tracking.","marker":"(Jiang et al., 2024)"},{"why":"Supplies GoodDrag, the strongest optimization-based baseline that DragLoRA must beat on both editing accuracy and runtime.","marker":"(Zhang et al., 2024)"},{"why":"Supplies the DIFT features used to compute the mean-distance metric that measures drag editing accuracy.","marker":"(Tang et al., 2023)"}],"fun_headline_variants":["DragLoRA: Online LoRA tuning sharpens drag edits","Online LoRA optimization gives finer drag control","Adaptive LoRA scheduling speeds drag editing without loss","DragLoRA cuts drag edit time to under 30s","Replace latent optimization with online LoRA for drag edits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The adaptive switching scheme assumes that the point-tracking quality signal minD and the Euclidean distance to the temporal target reliably indicate whether motion supervision is still needed, and that the hand-set thresholds ($d_1=1$, $d_2=1.3$, $l_1=1$, $l_2=1.4$, $k_{\\mathrm{ini}}=10$) generalize across images, content types, and handle-target configurations.","fun_headline_variants_meta":{"raw":{"variants":["DragLoRA: Online LoRA tuning sharpens drag edits","Online LoRA optimization gives finer drag control","Adaptive LoRA scheduling speeds drag editing without loss","DragLoRA cuts drag edit time to under 30s","Replace latent optimization with online LoRA for drag edits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000698,"raw_usage":{"total_tokens":3195,"prompt_tokens":1025,"completion_tokens":2170,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2092}},"tokens_in":641,"tokens_out":2170,"duration_ms":15817,"temperature":1.0,"reasoning_tokens":2092,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:33:42.520310+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DragLoRA on a set of images with occlusions or textureless regions where handle points are known to drift, and compare the algorithm's minD-based switch decisions against ground-truth tracking obtained by manual annotation or optical flow. If ILFA-only mode is activated frequently while the tracked handle is actually far from its true location, the switching signal is unreliable and the reported precision-runtime tradeoff would not generalize beyond the benchmark.","supporting_citations":[{"cited_title":"Furthermore, we apply ILFA to DragNoise (Liu et al.,","cited_arxiv_id":null,"evidence_quote":"Supplies the delta denoising score loss that DragLoRA reverses into a regularizer for the online LoRA optimization."}],"review_version":1}