{"id":"68a0b195-98ca-46b2-850a-dc2c737604fa","arxiv_id":"2507.09168","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SSD is a diffusion score-distillation loss for text-guided 2D and 3D editing that combines a CFG cross-prompt term, a null-text cross-trajectory regularizer, and a prompt-enhancement term to stabilize edits.","lead":"Stable Score Distillation (SSD) introduces a new score-distillation loss for text-guided image and 3D editing that replaces the source branch of DDS with a null-text regularization branch and adds a prompt-enhancement term. The paper claims faster, more stable edits that preserve original content while better matching target prompts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq.5's cross-trajectory term is a fixed θ-independent offset rather than a demonstrated structure regularizer; without a quantitative ablation isolating it, the source-preservation/stability claim is unsupported.","rationale":"The reader's weakest assumption identifies exactly the load-bearing risk: the null-text branch in Eq.5 is claimed to be a no-edit reference that preserves source structure, but the paper provides no derivation and only qualitative ablations. My reading adds a sharper technical point: because ẑ_t is the fixed source latent, ε_φ(ẑ_t, ∅) is independent of the current optimization variable, so the term is a constant offset in the gradient. That makes the 'regularizer' interpretation fragile: it can change the equilibrium but does not, by itself, couple the edit trajectory to the source structure. The quantitative Table 2 also shows the full method underperforms DDS+CDS on structure metrics, so the strongest advertised benefit is not evidenced. This does not prove the method is wrong; the idea may still work for other reasons, but the central claim needs the proposed ablation to distinguish a genuine stabilizer from a bias offset. The correct verdict remains CONDITIONAL, hence no change to the reader's judgment.","tokens_in":12360,"tokens_out":5762,"duration_ms":72182,"concrete_test":"Run a controlled PIE-Bench ablation with identical seeds: (a) full SSD (Eq.9); (b) SSD with the null term removed (Eq.4 gradient); (c) SSD with ε_φ(ẑ_t, ∅) replaced by DDS's source-prompt term ε_φ(ẑ_t, ŷ); and (d) SSD with the null term replaced by ε_φ(z_t, ∅) on the current latent. Report DINO/LPIPS/MSE/CLIP plus per-iteration gradient norms. If (a) does not significantly improve structure metrics over (b), or if (c) matches (a), the null-text branch is not a content-preserving regularizer; if (d) matches (a), the effect is independent of the source latent and the 'no-edit reference' interpretation fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq.5 defines L_ssd = ε_φ(z_t, ŷ) + s(ε_φ(z_t, y) − ε_φ(z_t, ŷ)) − ε_φ(ẑ_t, ∅). In the SDS update (Eq.1), ẑ_t is the noised source latent, fixed with respect to the optimized parameters θ; hence ε_φ(ẑ_t, ∅) contributes a θ-independent offset to the gradient. A constant offset can shift the fixed point of the optimization, but it cannot by itself impose structural consistency or reduce variance of the z_t-dependent part. The paper's only support is the qualitative ablation in Sec.5.3 (Figs.1 and 5), with hand-selected weights and no quantitative metrics. Moreover, in the quantitative PIE-Bench evaluation (Tab.2), full SSD has worse structure preservation than DDS+CDS (DINO 28.13 vs 7.15, LPIPS 82.43 vs 33.14, MSE 86.64 vs 25.29), directly undercutting the 'preserves source structure' wording of the central claim. Whether the null-text term behaves as a no-edit reference therefore remains an assumption, and Eq.9's gains could come from an arbitrary offset rather than a stabilizer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Stable Score Distillation (SSD), a score distillation loss for text-guided editing of both 2D images and 3D scenes (NeRF/3DGS). The loss replaces the source branch of DDS with a CFG-based cross-prompt term, adds a null-text term evaluated on the source latent (called a cross-trajectory term), and supplements these with a prompt-enhancement term and an ID regularization. The authors claim that this design improves stability, prompt alignment, source-content preservation, and convergence speed, and they report experiments on 3D scene editing (IN2N, LLFF, Mip-NeRF360) and 2D editing (PIE-Bench) against DDS, CSD, and other baselines. The paper includes qualitative comparisons, quantitative tables, a user study, and a discussion connecting the method to InstructPix2Pix.","tokens_in":12678,"tokens_out":5306,"duration_ms":57909,"significance":"If the central claims are correct, SSD would provide a simpler and more stable alternative to DDS/CSD for text-guided editing, with the additional conceptual contribution of linking score distillation to InstructPix2Pix. The paper is clearly written in structure, releases code, and identifies a concrete design space. However, the load-bearing theoretical justification for the null-text term is questionable from a gradient perspective, and the quantitative evidence is not fully consistent with the stated claims, particularly regarding structure preservation in Table 2. The absence of quantitative ablations and error bars further limits the strength of the empirical support. The work is potentially useful, but the manuscript in its current form does not convincingly establish the claimed advantages.","major_comments":[{"comment":"The central novelty is the cross-trajectory term -epsilon_phi(z_hat_t, empty) in Eq. (5), but in the SDS update (Eq. (1)) the source latent z_hat_t is fixed with respect to the optimized parameters theta. Therefore this term contributes a theta-independent offset to the loss and zero to the gradient, so it cannot by itself enforce source-structure preservation or stabilize the optimization. In the decomposition of Eq. (6), the only theta-dependent part of the cross-trajectory term is w_t * epsilon_phi(z_t, y_hat), which merely reweights the source-prompt score rather than introducing a new constraint. The paper should provide a derivation showing how this term affects the optimization trajectory (for example through the fixed point) or revise the claim that it regularizes structure.","section":"Sec. 4.1, Eqs. (5)-(6)"},{"comment":"Table 2 directly contradicts the central claim of source-structure preservation. Full SSD has DINO distance 28.13 x 10^-3, LPIPS 82.43 x 10^-3, and MSE 86.64 x 10^-4, versus 7.15, 33.14, and 25.29 for DDS+CDS, respectively, i.e., roughly 4x, 2.5x, and 3.4x worse. The text in Sec. 5.2 describes this as 'a slightly higher structure distance,' which understates the magnitude. If the method's main selling point is preserving source structure, this quantitative result needs to be reconciled, for example by reporting structure metrics on the edited regions only or by demonstrating that the larger distance comes from intended edits rather than unintended changes.","section":"Sec. 5.2, Table 2"},{"comment":"The ablations for the cross-trajectory and prompt-enhancement terms are qualitative only in the main text, with hand-selected weights (w_t = 0, 1.0, 1.5, 2.0 and w_e = 0, 1.5, 5.5, 7.5) shown in Fig. 5. No quantitative metrics are reported, and the text refers to 'the Supp.' for quantitative evaluation, but no supplementary material is included in the arXiv submission. Without a quantitative ablation that isolates the contribution of the null-text term, and without error bars or significance tests, the claim that this term provides stability and structure preservation is not empirically supported.","section":"Sec. 5.3, Fig. 5"},{"comment":"The loss in Eq. (5) fixes the coefficient of the null-text term to 1, which corresponds to w_t = 1 in the decomposition of Eq. (6). However, Fig. 5 sweeps w_t over values 0, 1.5, and 2.0, so the empirical study is not evaluating the proposed loss in Eq. (5) but a family of losses with a free parameter. The relationship between the theoretical loss and the implemented loss should be clarified, and the final reported results should state which weight values were used.","section":"Sec. 4.1, Eq. (5) vs Fig. 5"}],"minor_comments":[{"comment":"The abstract contains grammatical issues: 'achieves' should be 'achieve' and 'introduces a constant term null-text branch' is awkward; please rephrase for clarity.","section":"Abstract"},{"comment":"Equation (4) uses 'Grad' as a variable without defining whether it is a loss or a gradient, and 'The scale factor s is equal to control weight' is vague; please specify the role of s and its allowed range.","section":"Sec. 4.1, Eq. (4)"},{"comment":"The notation in Eq. (8) is unclear: x_t and z_hat_t are not explicitly defined; please state that x_t is the noised rendered latent and z_hat_t is the noised source latent, and clarify how the time indices are aligned.","section":"Sec. 4.3, Eq. (8)"},{"comment":"There are typos in Sec. 4.4 ('ours edit gard provide new angle') and inconsistent naming of InstructPix2Pix (Instructpix2pix, InstructP2P); please unify the terminology and proofread.","section":"Sec. 4.4"},{"comment":"The user study in Table 1 reports a single percentage without details on the number of questions, the rating scale, or statistical significance; please provide confidence intervals or a significance test.","section":"Sec. 5.1, Table 1"},{"comment":"The horizontal axis of Fig. 5 is not labeled; it is unclear what quantity is being varied (e.g., iteration number or weight value). Please add axis labels and a legend.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's abstract and introduction claim stronger source-preservation results than the quantitative data support, and the theoretical justification for the key cross-trajectory term appears to be a gradient-independent offset. These issues can be addressed by reframing the contribution as a reweighting of the source-prompt score rather than a novel regularizer, and by supplying the missing quantitative ablations and error bars. I also note that the manuscript references a supplementary for quantitative ablation, but no supplementary file is present in the arXiv submission; the editor should ensure it is provided in any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the central mechanism — the null-text branch in Eq.5 — is a constant offset with zero gradient. Since z_hat_t is fixed, epsilon_phi(z_hat_t, null) does not depend on the optimized parameters, so it cannot stabilize, regularize, or preserve structure. The decomposition in Eq.6 is algebraically correct, but the 'cross-trajectory' term reduces in gradient to a source-prompt SDS term. That is not new; it is a weighted combination of target and source SDS, similar to CSD with different coefficients. The paper's own Table 2 undercuts the structure-preservation claim: full SSD has worse DINO, LPIPS, and MSE than DDS+CDS, and the text dismisses this as a necessary trade-off for editability. That is legitimate but not the advertised claim. What is genuinely useful: the prompt-enhancement branch (Eq.7) and the ID regularization (Eq.8) are sensible, incremental additions. The paper also correctly notes that DDS can modify unedited regions. The writing is rough but the logic is otherwise clear. Soft spots: the stability claim rests entirely on a constant term. The ablation in Sec.5.3 is qualitative, with hand-swept weights and no held-out validation. There are no error bars or significance tests. The faster-convergence claim is supported only by a mention of decreasing timestep sampling, with no runtime comparison. The user study details are minimal. Who this is for: practitioners might try the final loss — a weighted target/source SDS plus CFG and ID regularization — but they would get the same from CSD with a source-prompt term. A referee should see this because the flaw is instructive and the empirical work might be salvageable, but as written the main claim is wrong. My recommendation: send to peer review only if the authors reframe the contribution; otherwise desk reject.","headline":"The paper's main novelty is a constant offset with zero gradient, so the cross-trajectory term does nothing to the optimization and the stability claim is unsupported.","tokens_in":671,"tokens_out":2102,"would_cite":false,"duration_ms":80623,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Stable Score Distillation adds a null-text branch to the editing score, producing text-guided 2D and 3D edits that preserve source structure and align more closely with the target prompt than DDS or CSD.","keywords":["text-guided image editing","3D scene editing","score distillation","classifier-free guidance","null-text branch","diffusion models","NeRF editing","3D Gaussian splatting"],"falsifier":"Sweep the weight of the null-text term $\\epsilon_\\phi(\\hat{z}_t, \\emptyset)$ in Eq. 5 while holding all other terms fixed, across a set of 2D and 3D editing tasks, and measure structure-preservation (DINO, LPIPS) and prompt-alignment (CLIP) metrics. If no consistent improvement or monotonic trend appears relative to weight zero, the term is not acting as a structure regularizer. Alternatively, set the term to a constant zero; if the edited results are unchanged, the paper's central novelty has no causal effect.","tokens_in":12141,"feed_emoji":"🎨","tokens_out":8637,"duration_ms":79489,"temperature":0.7,"pith_summary":"Text-guided editing of images and 3D scenes usually relies on score distillation losses, but existing formulations like $\\Delta$ Denoising Score (DDS) and Classifier Score Distillation (CSD) can blur unedited regions or deform structure. This paper introduces Stable Score Distillation (SSD), a loss that adds a null-text branch to the editing score. The loss combines a cross-prompt Classifier-Free Guidance (CFG) term with a cross-trajectory null-text term and a prompt-enhancement term. The authors report that SSD preserves the source structure and aligns edits with the target prompt more reliably than DDS and CSD on 2D and 3D editing benchmarks, while converging in fewer iterations.","feed_headline":"Null-text branch steadies text-guided 2D and 3D edits","feed_subtitle":"Replaces dual-branch pipelines with one classifier that preserves the original structure in images and scenes.","key_machinery":"The load-bearing object is the SSD score of Eq. 5, $\\mathcal{L}_{\\mathrm{ssd}} = \\epsilon_\\phi(z_t, \\hat{y}) + s(\\epsilon_\\phi(z_t, y) - \\epsilon_\\phi(z_t, \\hat{y})) - \\epsilon_\\phi(\\hat{z}_t, \\emptyset)$. The first two terms form the cross-prompt CFG direction from source prompt $\\hat{y}$ to target prompt $y$ on the current latent $z_t$; the third term, a null-text prediction on the source latent $\\hat{z}_t$, is the paper's main novelty and is decomposed into the cross-trajectory term. The full loss adds a prompt-enhancement branch $w_e(\\epsilon_\\phi(z_t, y) - \\epsilon_\\phi(z_t, \\emptyset))$ and an ID regularization term that uses a noisy source latent to avoid exploding gradients in 3D Gaussian splatting. The cross-trajectory term is said to measure the distance between the transitions of two latents, which stabilizes optimization by keeping the structure smooth.","core_discovery":"On its own terms, the paper establishes that a score distillation loss built from three terms—a cross-prompt CFG direction, a cross-trajectory null-text term, and a prompt-enhancement term—produces text-guided 2D and 3D edits that are more stable and more aligned with the target prompt than those from DDS and CSD. The decisive element is the cross-trajectory term $\\epsilon_\\phi(\\hat{z}_t, \\emptyset)$, which the paper interprets as a structure-preserving regularizer that keeps the optimization trajectory close to the source content. The resulting method also converges in about 3,000 iterations for NeRF and 1,500 for 3D Gaussian splatting, with no auxiliary source branch.","pith_inferences":["If the null-text term truly acts as a no-edit reference, the same regularization principle could be transferred to other SDS-based tasks such as text-to-3D generation, where source-structure preservation is less defined.","The paper's reading of InstructPix2Pix suggests that other one-step editing models may be decomposable into cross-prompt and cross-trajectory components, which could guide future loss designs.","Extending the paper's ablation to a full weight sweep would reveal whether the cross-trajectory term behaves monotonically as a structure regularizer, since only a few hand-selected weights are currently shown.","The method may struggle when the source image is far from the diffusion model's training distribution, because the null-text prediction then becomes unreliable as a 'no-edit' reference."],"forward_implications":["Pipelines based on DDS can drop the source branch, since the null-text term supplies the structure-preserving signal.","The method works directly on NeRF and 3D Gaussian splatting scenes, converging in about 3,000 and 1,500 iterations respectively.","Style edits that DDS-based methods often fail to apply become feasible with the prompt-enhancement branch.","The framework connects to InstructPix2Pix, showing that its one-step reverse sampling is, in effect, a simple version of the SSD loss with the cross-trajectory and cross-prompt terms.","The streamlined design requires no LoRA or fine-tuning and integrates with Stable Diffusion and InstructPix2Pix."],"supporting_citations":[{"why":"Defines Delta Denoising Score, the baseline whose dual-branch design SSD simplifies by removing the source branch.","marker":"[10]"},{"why":"Supplies the Classifier-Free Guidance equation that SSD reuses for the cross-prompt term.","marker":"[12]"},{"why":"Decomposes the CFG score and identifies the classifier as the prompt-direction driver, motivating the single-classifier design.","marker":"[19]"},{"why":"Introduces Score Distillation Sampling, the optimization framework that SSD modifies for editing.","marker":"[34]"},{"why":"Defines Classifier Score Distillation, the dual-classifier baseline that SSD extends with the null-text branch.","marker":"[49]"}],"fun_headline_variants":["SSD: null-text regularizer stabilizes image and 3D edits","Null-text branch replaces auxiliary pipelines for stable edits","Stable Score Distillation: one classifier for smooth 2D and 3D edits","SSD: faster convergence, stable edits with single CFG anchor","Null-text branch steadies text-guided 2D and 3D edits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that subtracting the diffusion model's null-text prediction on the source latent is a faithful, unbiased 'no-edit' reference; if that term instead injects an uncontrolled bias, the method collapses to Classifier Score Distillation plus an arbitrary offset.","fun_headline_variants_meta":{"raw":{"variants":["SSD: null-text regularizer stabilizes image and 3D edits","Null-text branch replaces auxiliary pipelines for stable edits","Stable Score Distillation: one classifier for smooth 2D and 3D edits","SSD: faster convergence, stable edits with single CFG anchor","Null-text branch steadies text-guided 2D and 3D edits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000881,"raw_usage":{"total_tokens":3776,"prompt_tokens":880,"completion_tokens":2896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":2800}},"tokens_in":496,"tokens_out":2896,"duration_ms":24253,"temperature":1.0,"reasoning_tokens":2800,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:02:28.633720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sweep the weight of the null-text term $\\epsilon_\\phi(\\hat{z}_t, \\emptyset)$ in Eq. 5 while holding all other terms fixed, across a set of 2D and 3D editing tasks, and measure structure-preservation (DINO, LPIPS) and prompt-alignment (CLIP) metrics. If no consistent improvement or monotonic trend appears relative to weight zero, the term is not acting as a structure regularizer. Alternatively, set the term to a constant zero; if the edited results are unchanged, the paper's central novelty has no causal effect.","supporting_citations":[{"cited_title":"Delta denoising score","cited_arxiv_id":null,"evidence_quote":"Defines Delta Denoising Score, the baseline whose dual-branch design SSD simplifies by removing the source branch."},{"cited_title":"Classifier-free diffusion guidance","cited_arxiv_id":null,"evidence_quote":"Supplies the Classifier-Free Guidance equation that SSD reuses for the cross-prompt term."},{"cited_title":"Noise-free score distillation","cited_arxiv_id":null,"evidence_quote":"Decomposes the CFG score and identifies the classifier as the prompt-direction driver, motivating the single-classifier design."},{"cited_title":"Barron, and Ben Mildenhall","cited_arxiv_id":null,"evidence_quote":"Introduces Score Distillation Sampling, the optimization framework that SSD modifies for editing."},{"cited_title":"Text-to-3d with classifier score distillation","cited_arxiv_id":null,"evidence_quote":"Defines Classifier Score Distillation, the dual-classifier baseline that SSD extends with the null-text branch."}],"review_version":1}