{"id":"741cc36d-177d-4a64-ac12-ada050a922ef","arxiv_id":"2411.14863","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using prompt optimization and an SNR-matching change of variables, pretrained Stable Diffusion can approximate the source, target, and noise predictors of a Schrödinger bridge ODE, enabling 8-step unpaired image translation.","lead":"The paper proposes a fast method for unpaired image-to-image translation by approximating a Schrödinger bridge ODE with components of a pretrained Stable Diffusion model. It reports competitive translations in as few as 8 neural function evaluations, much faster than prior diffusion-based translation methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SNR matching alone does not make a P0-trained VP denoiser compute E[x0|xt] under the LSB forward model, so Proposition 1's premise is unmet and the ODE's marginals are not guaranteed.","rationale":"The reader's weakest_assumption identifies the same gap: matching SNR does not align the posterior distributions because the clean signal in xt is a convex combination of source and target rather than a sample from P0. My stress-test confirms that this is the load-bearing point and makes it quantitative: for a Gaussian example with independent coupling, the exact SB posterior mean and the VP Tweedie estimate differ for all interior t. The paper's own Appendix G admits that the global OT coupling P^tau_01 is not constructed, which is exactly the premise needed for Proposition 1. The additional discrepancy in Algorithm 1 line 8 (inverse of Eq. (20)) reinforces that the written method is not internally consistent. Because the empirical results may still hold despite the theoretical gap, this does not change the reader's CONDITIONAL verdict; it strengthens the condition that the posterior-matching assumption must be either derived or validated empirically.","tokens_in":15058,"tokens_out":12513,"duration_ms":132589,"concrete_test":"Run a controlled low-dimensional replication. Set P0=N(0,I) and P1=N(0,Sigma1) with Sigma1 != I, choose the independent coupling Gamma01 = P0 x P1, and sample xt = (1-t)x0 + t x1 + sigma_t epsilon. Train a VP denoiser on P0 only; also compute the exact posterior means E[x0|xt], E[x1|xt], and E[epsilon|xt] analytically. For t in {0.1,...,0.9}, compare (i) the exact predictors, (ii) the SNR-matched VP estimates of Eq. (20), and (iii) the VP estimates with Algorithm 1's inverse scaling. Integrate Eq. (15) with each predictor set and measure the Wasserstein distance between the t=1 marginal and P1. If the VP-based final marginal deviates materially from P1, or the exact and VP posterior means differ beyond numerical tolerance, the load-bearing assumption fails; if they match, SNR matching is sufficient despite the derivation gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assumption is in Section 4.2, Eq. (20): a VP denoiser trained on P0 is claimed to provide xhat0(xt) and epshat(xt) after SNR matching. For this to be valid, the rescaled LSB state y_s must have the same conditional posterior as the VP training input. It does not. Under the LSB forward model (Eq. 13), y_s = sqrt(alpha_s) xt = sqrt(alpha_s)[(1-t)x0 + t x1] + sqrt(1-alpha_s) epsilon, whereas the VP training input is y_s = sqrt(alpha_s) x0 + sqrt(1-alpha_s) epsilon with x0 ~ P0. The extra target term t sqrt(alpha_s) x1 is structured signal from P1, not Gaussian noise, and it is not part of the VP training distribution. Hence xhat0^VP(y_s) is not the conditional expectation E[x0|xt] required in Eq. (14), and Proposition 1's premise that the predictors are posterior means under one coupling P_t|01 with marginals P0 and P1 is not satisfied. For a zero-mean isotropic Gaussian example with independent coupling, the true posterior is E[x0|xt] = (1-t)/((1-t)^2+t^2+sigma_t^2) xt, while the SNR-matched VP Tweedie estimate is 1/(sigma_t^2+1) xt; these differ for all t in (0,1). The paper concedes in Appendix G that it has not constructed the global OT map P^tau_01, so no single coupling underlies the three predictors. The LSB ODE is therefore an ad hoc flow whose final marginal is not guaranteed to be P1. Separately, Algorithm 1 line 8 writes y_s = x_t / sqrt(alpha_s), which is the inverse of Eq. (20) and would scale noise up rather than down; this is at best a typo and at worst a second internal inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Latent Schrödinger Bridge (LSB), a method for fast unpaired image-to-image translation that leverages a pre-trained Stable Diffusion model. The authors decompose the SB probability flow ODE velocity into three terms—source predictor, target predictor, and noise predictor—and approximate each using a single text-conditioned VP diffusion denoiser, with domain-specific text embeddings obtained via textual inversion. They introduce an SNR-matching change of variables, a time-dependent noise predictor, CFG scaling, and a final denoising step. Experiments on Cat→Dog, Horse→Zebra, and Dog→Wild show improved FID/DINOv2 FD at 8 NFEs compared to DDIB and SDEdit, with additional ablations and examples across several domains.","tokens_in":15516,"tokens_out":3748,"duration_ms":37048,"significance":"If the theoretical grounding held, the paper would be a valuable practical contribution: it demonstrates that a single pre-trained diffusion model, without training an SB, can produce competitive unpaired translation in very few NFEs, and the component analysis (source/target/noise predictors) is intuitive. The empirical study is extensive and the code release is promised. However, the central theoretical claim—that VP diffusion denoisers compute the required SB posterior means after SNR matching—is not established and is in fact questionable. The paper itself concedes in Appendix G that no global OT map is constructed, so the proposed ODE is not a Schrödinger bridge in the strict sense. Thus the significance as a principled SB-based method is weakened; it may still be a useful empirical heuristic, but the current framing overstates the theoretical contribution.","major_comments":[{"comment":"The SNR-matching change of variables does not imply that a VP denoiser trained on P0 computes the SB conditional expectation E[x0|xt] required by Eq. (14). Under the LSB forward model, the rescaled input y_s = sqrt(alpha_s) x_t = sqrt(alpha_s)[(1-t)x0 + t x1] + sqrt(1-alpha_s) epsilon contains a structured signal from the target domain (the term t sqrt(alpha_s) x1), which is not present in the VP training distribution y_s = sqrt(alpha_s) x0 + sqrt(1-alpha_s) epsilon with x0 ~ P0. Matching only the scalar SNR does not align the conditional posteriors. For instance, with zero-mean isotropic Gaussian data and independent coupling, the true posterior E[x0|xt] differs from the Tweedie estimate produced by a VP denoiser for all t in (0,1). This undermines the premise of Proposition 1, which requires the predictors to be posterior means under a coupling with marginals P0 and P1. The Appendix G admission that the global OT map P^tau_01 is not constructed means the ODE's final marginal is not guaranteed to be P1. The authors should either provide a rigorous justification for why the VP denoiser approximation is valid (e.g., under additional assumptions or with a correction term), or reframe the method as an empirical approximation and validate the approximation directly (e.g., by measuring the discrepancy between the predictors and the true posterior means on synthetic data).","section":"Section 4.2, Eq. (20)"},{"comment":"The pseudo-code writes y_{s_i} <- x_{t_i} / sqrt(alpha_{s_i}), which is the inverse of the formula y_s = sqrt(alpha_s) x_t given in Eq. (20). Since alpha_{s_i} = 1/(sigma_{t_i}^2 + 1), the line as written scales the input by sqrt(sigma_{t_i}^2+1), increasing the noise component rather than decreasing it. This is not a harmless typo: it directly contradicts the derivation in Section 4.2 and, if actually executed, would likely degrade performance substantially. The authors must correct this line (it should be y_{s_i} <- sqrt(alpha_{s_i}) x_{t_i}) and clarify whether the reported experiments used the corrected formula.","section":"Algorithm 1, line 8"},{"comment":"The hyperparameters t0, tau, and the CFG scale omega are selected by grid search on the same tasks and metrics (FID) that are reported in the main results (Table 1 and Figure 3). This makes the reported improvements partly a product of tuning rather than a general property of the algorithm. To support the claim of broad applicability, the authors should either report results with a fixed, task-independent hyperparameter setting, or show sensitivity curves and report the performance on held-out tasks or with a validation split that is not used for tuning. Otherwise, the reader cannot assess how much of the advantage is due to the method versus the grid search.","section":"Section 5.1 and Appendix C.5"}],"minor_comments":[{"comment":"The limitations paragraph states that the method is not a true Schrödinger bridge because the global OT map is not constructed. This admission is in tension with the title and abstract, which present LSB as a Schrödinger bridge method. The authors should either soften the claim or rename the method to avoid overclaiming.","section":"Appendix G"},{"comment":"The derivation of Eq. (12) from Eq. (11) uses sigma_t = sqrt(t(1-t) tau), but the expression has a typo: the factor in the noise term should read (1/2-t) sqrt(tau) / sqrt(t(1-t)), not (1/2-t) sqrt(tau) / sqrt(t(1-t)) with the square root misplaced. Please correct the formula and re-check the subsequent equations.","section":"Section 4.1, Eq. (12)"},{"comment":"The text says 'we can improve the reflectance of text prompt'—likely intended as 'reflectance' or 'relevance'? Please clarify the wording.","section":"Section 4.3, 'Prompt optimization'"},{"comment":"Some references appear incomplete or have incorrect page numbers (e.g., [12] is cited as 'arXiv preprint' without a year, and [30] has an extra '2' at the end). Please ensure all references are accurately formatted.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has a core theoretical flaw in the justification of the SNR-matching step, but the empirical method appears to work. The authors should be given the opportunity to either fix the theory or reposition the contribution as an empirical heuristic with a clear analysis of when it is valid. The Algorithm 1 error is a concrete, fixable issue. The hyperparameter tuning concern is standard but should be addressed with additional experiments. If the theoretical gap cannot be closed, the paper may be better suited for a venue that accepts strong empirical results with a heuristic formulation, but as submitted, the overclaiming of a Schrödinger bridge foundation is problematic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper has two sides. The practical side is genuinely useful: an 8-NFE unpaired I2I method that reuses pretrained Stable Diffusion with textual inversion and an SNR-matching trick, and it beats SDEdit and DDIB on FID at low NFE. The ablations in Table 2 show each component matters, and the qualitative results look plausible. That part deserves a serious look.\n\nThe theoretical side is where it falls apart. The decomposition in Eq. (12) is a clean algebraic rearrangement of the standard flow-matching conditional velocity, and Proposition 1 is correct as a restatement of Tong et al.'s Thm 3.1 — nothing new there, but it's fine. The problem is the step from VP diffusion scores to the SB posterior means in Eq. (14). Matching SNR does not make a P0-trained VP denoiser compute E[x0|xt] under the LSB forward model. The clean signal in xt is (1-t)x0 + t x1, not just x0, so the extra target term is structured signal from P1, not Gaussian noise. The stress-test Gaussian example makes this precise: the true posterior and the SNR-matched Tweedie estimate differ for all t in (0,1). The paper itself concedes in Appendix G that it never constructs the global OT map P^τ_01, so no single coupling underlies the three predictors. That means Proposition 1's premise is unmet and the final marginal is not guaranteed to be P1. The method can still work as a heuristic — the results suggest it does — but the paper should say that honestly instead of claiming an SB approximation.\n\nThere's also an internal inconsistency: Eq. (20) gives y_s = sqrt(ᾱ_s) x_t, while Algorithm 1 line 8 uses y_s = x_t / sqrt(ᾱ_s), the inverse. One of them is wrong, and this needs fixing before anyone can reproduce the algorithm as written. Hyperparameters (τ, t0, ω) are grid-searched per task and per NFE in Appendix C.5; that's acceptable for an empirical paper, but it should be stated up front and accompanied by error bars. No code is released yet.\n\nMy take: this deserves peer review. The method is practical, the empirical claim is plausible, and the gaps are fixable with honest reframing and a corrected algorithm. As it stands I would not cite it for the theory, but I'd watch for the code release and the follow-up.","headline":"Fast unpaired I2I via pretrained SD is real; the SB theory is an overclaim, not a proof.","tokens_in":16003,"tokens_out":5128,"would_cite":false,"duration_ms":45903,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper derives a Schrodinger-bridge ODE whose velocity splits into source, target, and noise predictors, and shows that one pretrained Stable Diffusion model can supply all three for fast unpaired image-to-image translation.","keywords":["Schrodinger bridge","unpaired image-to-image translation","probability flow ODE","latent diffusion","Stable Diffusion","textual inversion","SNR matching","fast sampling"],"falsifier":"Construct a synthetic pair of distributions P0 and P1 where the posterior means E[x0|xt], E[x1|xt], and E[epsilon|xt] can be computed in closed form, then integrate the LSB ODE with the exact predictors and compare the terminal distribution with P1; a mismatch would invalidate the general-predictor claim. Separately, to test the SNR-matching premise, draw xt from the bridge construction with known x0, x1, and epsilon, and compare the VP denoiser output with the empirical conditional average over many samples: large systematic disagreement means the approximation fails.","tokens_in":14805,"feed_emoji":"🖼️","tokens_out":5637,"duration_ms":59035,"temperature":0.7,"pith_summary":"The paper tries to establish that unpaired image-to-image translation can be made fast by reusing a pretrained latent diffusion model as an approximate Schrodinger bridge sampler. It derives a decomposition of the Schrodinger-bridge probability-flow ODE velocity into three interpretable predictors, and then argues that all three can be obtained from a single Stable Diffusion model via prompt optimization and an SNR-matching change of variables. If the construction is right, an image can be translated between domains in as few as eight neural function evaluations, with quality competitive to diffusion baselines that need many more steps.","feed_headline":"Image translation in 8 diffusion steps","feed_subtitle":"A single Stable Diffusion model supplies source, target, and noise predictors for a Schrodinger-bridge ODE.","key_machinery":"The central object is the decomposed Schrodinger-bridge ODE velocity v(xt,t) = ((1/2-t) $\\sqrt$(tau) / $\\sqrt$(t(1-t))) epsilon_hat(xt) + xhat_1(xt) - xhat_0(xt), where each term is a conditional expectation: the source predictor xhat_0, the target predictor xhat_1, and the noise predictor epsilon_hat. Proposition 1 provides the load-bearing justification that any coupling between P0 and P1 yields a valid probability-flow ODE between the marginals. The SNR-matching map ys = xt / $\\sqrt$($sigma_t^{2}$ + 1) aligns the bridge state with what the VP diffusion denoiser saw during training, and prompt optimization via textual inversion turns one Stable Diffusion model into both the source and target predictors.","core_discovery":"The central claim is that the Schrodinger-bridge probability-flow ODE velocity has the form v(xt,t) = ((1/2-t) $\\sqrt$(tau) / $\\sqrt$(t(1-t))) epsilon_hat(xt) + xhat_1(xt) - xhat_0(xt), and that this ODE still transports samples between P0 and P1 even if the three predictors are not the true Schrodinger-bridge posteriors. The source predictor xhat_0 repels from the source domain, the target predictor xhat_1 attracts toward the target domain, and the noise predictor epsilon_hat adds or removes noise depending on whether t < 0.5 or t > 0.5. The paper asserts that a variance-preserving diffusion denoiser trained on source images, after rescaling the bridge state by ys = xt / $\\sqrt$($sigma_t^{2}$ + 1), provides valid estimates for x0 and epsilon, and the same trick with target-domain prompts provides x1. Prompt-optimized text embeddings let one Stable Diffusion model act as both domain predictors, and the result is claimed to be competitive unpaired translation with as few as 8 NFEs.","pith_inferences":["The three-predictor decomposition is not tied to Stable Diffusion specifically: any score-based or flow-based denoiser could in principle supply the predictors, so the architecture may extend to faster distilled models or other latent diffusion backbones.","The SNR-matching argument aligns only the marginal noise level, not the full posterior distribution, so the reliability of the method on new domains should be tested by comparing denoiser outputs with exact posterior means on controlled data.","The dependence on textual inversion implies that domains far outside the pretrained model's prior may require finetuning or more than 1k images, a regime the paper does not explore.","Combining the LSB ODE with fast ODE solvers or learned step-size schedules might push the NFE count below 8 without the quality loss the paper reports at 4 NFEs."],"forward_implications":["Unpaired image-to-image translation can run with as few as 8 neural function evaluations using a single pretrained diffusion model, avoiding the separate inversion stage that slows down dual-diffusion baselines.","The source-repelling term addresses a failure mode of SDEdit, where incomplete translation leaves source structure intact because there is no force pushing away from the source domain.","Because any coupling between the source and target distributions suffices, expensive Schrodinger-bridge training is not needed to obtain a valid transport ODE.","Textual inversion with fewer than a thousand domain images is enough to create usable source and target predictors in Stable Diffusion.","The same ODE can translate in reverse by integrating from t = 1 to t = 0, giving bidirectional translation from one model and one set of optimized prompts."],"supporting_citations":[{"why":"Supplies the conditional vector field and the regression result that define the Schrodinger-bridge probability-flow ODE.","marker":"[32]"},{"why":"Provides the textual inversion technique used to optimize source and target text embeddings.","marker":"[5]"},{"why":"Defines the VP diffusion probability-flow ODE and score-based framework that the pretrained denoiser approximates.","marker":"[28]"},{"why":"Gives Tweedie's formula used to convert noise prediction into the x0 estimate.","marker":"[14]"},{"why":"Used in the proof of Proposition 2 to connect the tau = 0 case with the diffusion probability-flow ODE.","marker":"[18]"},{"why":"Serves as the dual-diffusion baseline requiring inversion, which the paper compares against.","marker":"[31]"},{"why":"Serves as the SDEdit baseline that lacks the source-repelling term.","marker":"[20]"}],"fun_headline_variants":["8-step unpaired image translation via latent Schrodinger bridge","Latent Schrodinger bridge cuts unpaired I2I to 8 NFEs","Fast unpaired I2I: latent Schrodinger bridge + Stable Diffusion","Schrodinger-bridge ODE enables 8-step unpaired translation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a denoiser trained on source images can, after the SNR-matching rescaling, return an accurate estimate of the original source image from the bridge state, and that the same holds for target images; the paper does not prove this.","fun_headline_variants_meta":{"raw":{"variants":["8-step unpaired image translation via latent Schrodinger bridge","Latent Schrodinger bridge cuts unpaired I2I to 8 NFEs","Fast unpaired I2I: latent Schrodinger bridge + Stable Diffusion","Schrodinger-bridge ODE enables 8-step unpaired translation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00046,"raw_usage":{"total_tokens":2312,"prompt_tokens":965,"completion_tokens":1347,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":1269}},"tokens_in":581,"tokens_out":1347,"duration_ms":10425,"temperature":1.0,"reasoning_tokens":1269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:47:15.918964+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic pair of distributions P0 and P1 where the posterior means E[x0|xt], E[x1|xt], and E[epsilon|xt] can be computed in closed form, then integrate the LSB ODE with the exact predictors and compare the terminal distribution with P1; a mismatch would invalidate the general-predictor claim. Separately, to test the SNR-matching premise, draw xt from the bridge construction with known x0, x1, and epsilon, and compare the VP denoiser output with the empirical conditional average over many samples: large systematic disagreement means the approximation fails.","supporting_citations":[{"cited_title":"Bermano, Gal Chechik, and Daniel Cohen-Or","cited_arxiv_id":null,"evidence_quote":"Provides the textual inversion technique used to optimize source and target text embeddings."},{"cited_title":"Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole","cited_arxiv_id":null,"evidence_quote":"Defines the VP diffusion probability-flow ODE and score-based framework that the pretrained denoiser approximates."},{"cited_title":"Noise2Score: Tweedie’s Approach to Self-Supervised Image Denoising without Clean Images","cited_arxiv_id":null,"evidence_quote":"Gives Tweedie's formula used to convert noise prediction into the x0 estimate."},{"cited_title":"Dual Diffusion Implicit Bridges for Image-to-Image Trans- lation","cited_arxiv_id":null,"evidence_quote":"Serves as the dual-diffusion baseline requiring inversion, which the paper compares against."},{"cited_title":"SDEdit: Guided Im- age Synthesis and Editing with Stochastic Differential Equa- tions","cited_arxiv_id":null,"evidence_quote":"Serves as the SDEdit baseline that lacks the source-repelling term."}],"review_version":1}