{"id":"8bfe34c3-6562-4fb1-a10f-e86fdefd32ee","arxiv_id":"2502.01201","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A DreamBooth-style diffusion model personalizes each query image into a normal counterpart, and a triplet CLIP comparison combines query-versus-personalized, query-versus-normal-samples, and query-versus-text scores for few-shot anomaly detection.","lead":"This paper proposes a few-shot anomaly detection method that uses a personalized diffusion model to transform each test image into its imagined normal version, then compares the two with a CLIP-based triplet scoring. It reports accuracy gains over prior zero/few-shot AD methods across 11 industrial, medical, and semantic datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The prompt-selection objective in Eq. 2 minimizes SSIM against the query xq rather than against any normal reference, so the chosen prompt can preserve anomalies; the one-to-normal mechanism is not established as written.","rationale":"The paper's core idea is genuinely novel: instead of comparing a query to few-shot normal images directly, it generates a personalized normal counterpart and compares against it. The experimental sweep is broad, and the reported gains are plausible enough to merit further investigation. However, the central mechanism depends entirely on the one-to-normal transformation being faithful: anomalies must be removed while normal structure is preserved. The reader identified this transformation as the weakest assumption, and I agree that this is the load-bearing point. My specific concern sharpens that assumption: Eq. (2) does not actually implement the claimed selection of a normal-state prompt. The objective as written compares the generated image to the query itself, which is anomalous, so it can select a prompt that preserves the defect. This is an internal inconsistency with the prose, not merely a disagreement with external consensus. It is also testable: replacing the objective with a normal-reference-based criterion and measuring anomaly removal would settle whether the flaw is a harmless typo or a real defect in the method. I do not recommend rejection because the fix is straightforward and the empirical framework is otherwise sound; however, the paper as written does not substantiate its central mechanism, so conditional acceptance with a mandated correction and experiment is appropriate. Supporting issues such as missing code and fixed hyperparameters without a validation protocol reinforce the conditional verdict but are secondary to the Eq. (2) problem.","tokens_in":15955,"tokens_out":5184,"duration_ms":50038,"concrete_test":"On MVTec-AD in the 8-shot setting, run the described pipeline and for each anomalous query compute three personalized images: (a) using the prompt selected by Eq. (2) as written, (b) using a fixed state-level normal prompt such as '[object] without flaw', and (c) using a corrected objective that selects the prompt whose output maximizes SSIM against the few-shot normal reference images (or minimizes CLIP feature distance to the normal memory bank). Then measure anomaly removal by computing pixel-level AUROC of the difference map |xq − xhat| against ground-truth anomaly masks, and also report the image-level AUROC of the full triplet score (Eq. 7) for each variant. If variant (a) is not better than (b) and is worse than (c), the Eq.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3.3 claims the method selects the prompt whose generated image 'most closely resembles the normal state,' but Eq. (2) defines cq = argmin_{ci} L(xq, Dθ(xq,t, ci, t)) with L stated as SSIM. This is doubly problematic. First, if L is SSIM as written, argmin selects the least similar reconstruction, opposite to the stated goal. Second, if L is meant as a dissimilarity such as 1−SSIM, the objective still compares only to the query xq, not to any normal reference image or to the learned normal manifold. For an anomalous query, the prompt that maximizes similarity to xq is the one that best preserves the anomaly's appearance, not the one that maps the image onto Pnormal. The prompt set mixes state-level normal prompts such as 'without flaw' with physical-level prompts such as 'dark photo,' and Eq. (2) provides no way to prefer the normal-state prompt. The low noise setting t=0.3 makes the failure more likely: at this noise level the reconstruction is strongly conditioned by the input, so SSIM to xq is large for essentially every prompt and the selection is nearly arbitrary. If the personalized image retains the anomaly, the SP branch in Eq. (4) assigns low anomaly scores to defective queries; if it erases normal structure, normal queries receive high scores. Either case removes the claimed advantage over direct feature comparison. The paper's own ablation in Table 5 shows prompt choice affects results, so the flaw is not vacuous. Thus the central one-to-normal premise rests on an internally inconsistent objective, not merely on an empirical assumption about diffusion model behavior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a few-shot anomaly detection method built on a DreamBooth-style customized diffusion model. Given a small set of normal reference images, the method fine-tunes a diffusion model to model the normal distribution, then 'personalizes' each query image by reconstructing it at a fixed noise level under a text prompt selected from a set of normal-state prompts. The resulting personalized image is compared with the query, together with an anomaly-free sample memory bank and text-prompt scores, through a triplet contrastive inference scheme whose weighted combination gives the final anomaly score. The authors report experiments on 11 datasets in industrial, medical, and semantic domains, claim consistent AUROC/AUPRC improvements over several few-shot AD baselines, and show that the generated anomaly-free samples can improve other AD methods when used as augmentation.","tokens_in":16316,"tokens_out":5208,"duration_ms":50060,"significance":"The empirical scope is a genuine strength: 11 datasets across three domains, three few-shot settings, five runs, ablations of each inference branch and of the prompt design, plus transfer experiments with generated samples used as training data for PaDiM, PatchCore, WinCLIP, and InCTRL. If the one-to-normal premise were established, the method would be a valuable contribution to few-shot anomaly detection. However, as written, the central mechanism is not established: the prompt-selection objective in Eq. (2) is internally inconsistent, and the paper provides no quantitative verification that the personalized images actually remove anomalies while preserving normal structure. The empirical claims therefore do not yet validate the proposed mechanism, and the contribution is conditional on a corrected and verified formulation.","major_comments":[{"comment":"The prompt-selection objective is stated as cq = argmin_ci L(xq, Dθ(xq,t, ci, t)) with L identified as SSIM. An argmin over SSIM selects the reconstruction least similar to the query, which is the opposite of the 'most closely resembles the normal state' goal stated in the same paragraph. If L is instead intended as a dissimilarity such as 1−SSIM, the objective still only compares with xq, not with any normal reference image or with the learned normal manifold Pnormal. For an anomalous query, the prompt that maximizes similarity to xq is the one that best preserves the anomaly's appearance. Since Table 5 shows that the prompt selection changes results (e.g., MVTec-AD 95.6 to 96.2, OCT2017 98.7 to 99.3), the ambiguity is consequential. Please correct the sign and re-define the selection criterion relative to normal references or the normal-state distribution, and add evidence that the selected prompt does not preserve anomalies.","section":"Sec. 3.3, Eq. (2)"},{"comment":"The one-to-normal transformation uses t = 0.3 with a noise schedule in which the generated image is strongly conditioned by xq. At this noise level, the text prompt has limited influence and the reconstruction may retain the anomaly; conversely, if the schedule is more aggressive it may erase normal structure. No sensitivity analysis or quantitative verification that Dθ maps anomalous queries onto Pnormal is provided. Please report AUROC as a function of t and show, for example, that personalized images of anomalous queries are closer to normal reference features than the original queries are, or provide an alternative validation of the core premise.","section":"Sec. 3.3, Eq. (3) and t hyperparameter"},{"comment":"The final score combines SP, SN, and Stext with weights α = 1 and β = 0.5, and the personalization strength t = 0.3 is chosen globally. The manuscript does not describe a validation protocol (e.g., a held-out split) for these choices, and the same 11 datasets are used for both selection and reporting. Because the branch ablation in Table 4 shows that the relative contributions vary strongly by dataset (e.g., Stext alone gives 45.3 on OCT2017 but 96.3 on KSDD), the fixed weights may mask overfitting or underfitting to particular domains. Please report sensitivity to α, β, and t and clarify the selection protocol.","section":"Sec. 4.1 and Eq. (7)"}],"minor_comments":[{"comment":"Eq. (1) has an unmatched closing parenthesis in ||Dθ(xnormal,t, c) − x0||2, and the notation xnormal,t is not defined before it is used in Eq. (2); please define the forward noising process once and use it consistently.","section":"Sec. 3.2, Eq. (1)"},{"comment":"The sentence reporting 8-shot improvements lists six datasets (KSDD, ELPV, AFID, KSDD, VisA, MVTec) but gives five percentage values and duplicates KSDD; please reconcile the text with the values in Table 1.","section":"Sec. 4.2, paragraph after Table 1"},{"comment":"The AFID row in the 2-shot block is missing a closing parenthesis: '(78.3±1.7, 53.6±1.2' should be '(78.3±1.7, 53.6±1.2)'.","section":"Table 1"},{"comment":"The inference-time paragraph states that the proposed method is '+200-300ms per query image' higher than WinCLIP (389ms) and InCTRL (276ms), then says that the single-prompt configuration requires 326ms, which is lower than WinCLIP; please clarify which configuration the +200-300ms figure refers to.","section":"Sec. 4.3, Inference time"},{"comment":"The main text says three prompts were used for one-to-normal personalization, while the appendix describes experiments with 1, 3, 5, and 10 prompts and says all ten gave the best performance; please state the exact number and contents of prompts used in the reported main results.","section":"Appendix A.1"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline case between major revision and rejection. The empirical results are broad and the augmentation-transfer experiments are a real strength, but the central mechanism as written is internally inconsistent at a load-bearing point: Eq. (2) either has the wrong sign or compares against the wrong reference. The authors can likely address this by correcting the objective, grounding it in normal references, and adding verification that anomalies are removed; I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious referee, but the central mechanism has a flaw that needs fixing. The idea is genuinely new: combining DreamBooth-style personalization with few-shot anomaly detection, then scoring a query against its own personalized normal counterpart, an augmented normal pool, and text prompts. That integration is not in the cited WinCLIP/InCTRL line of work, and the evaluation is unusually broad—11 datasets across industrial, medical, and semantic domains, plus experiments showing the generated anomaly-free samples improve other AD methods. The augmentation transfer is a useful result on its own.\n\nThe stress-test concern lands. Eq. (2) says the prompt is chosen by argmin over SSIM between the query and the generated image. With SSIM, argmin selects the least similar image. If a dissimilarity like 1−SSIM was intended, the objective still compares only to the query, not to any normal reference or to the learned normal manifold. For an anomalous query, the prompt that preserves the anomaly will score well. At the stated t=0.3, the reconstruction is strongly conditioned by the input, so the personalized image likely keeps the defect, and the SP branch becomes uninformative or worse. The paper's own Table 5 shows prompt choice affects results, so this is not idle. The authors need to either correct the objective to use a normal-state reference or demonstrate empirically that the selected generated image actually removes anomalies while preserving normal structure. The figures claim this but are not compelling without a quantitative check.\n\nThe other soft spots are real but lesser. Hyperparameters t=0.3, α=1, β=0.5 are fixed across all datasets without any validation protocol; that is a tuning-fit concern, not fatal. No code is shipped. The writing has some ambiguities (e.g., the auxiliary-data sentence for MVTec), but nothing else load-bearing.\n\nThis paper is for researchers working on generative approaches to few-shot AD. It deserves peer review, not desk rejection, because the core idea is promising and the experimental sweep is substantial. My recommendation: send it out, but expect heavy revision. The authors should fix or re-justify the prompt selection, add a sensitivity analysis for t, and ideally release code. I would not cite it in its current form.","headline":"Promising idea with a load-bearing flaw in the prompt-selection objective (Eq. 2) that the authors must fix before the one-to-normal claim is credible.","tokens_in":16828,"tokens_out":2823,"would_cite":false,"duration_ms":28290,"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":"The paper's central claim is that few-shot anomaly detection can be sharpened by first generating a personalized normal version of each query image with a customized diffusion model and then scoring the query against that normal twin, a…","keywords":["few-shot anomaly detection","anomaly personalization","one-to-normal transformation","diffusion models","Dreambooth customization","triplet contrastive inference","CLIP","anomaly-free sample generation"],"falsifier":"Take a defect-free but visually atypical object (e.g., a normal image with unusual pose, lighting, or background) as query, run the one-to-normal personalization, and measure whether the pixel-level difference between query and personalized image is concentrated in genuinely anomalous regions; if normal regions shift as much as anomalous ones, or if a deliberately injected defect survives the transformation, the central premise is refuted. Concretely, on MVTec-AD with known ground-truth masks, compute the AUROC of a defect-localization score defined by the per-pixel difference between $x_q$ and $\\hat{x}_q$; if it is no better than chance on categories with high pose variation, the personalization is not selectively removing anomalies.","tokens_in":15777,"feed_emoji":"🔍","tokens_out":6937,"duration_ms":60859,"temperature":0.7,"pith_summary":"Few-shot anomaly detection usually compares a query image directly against a few stored normal reference images, a step that is easily confused by benign differences in pose, texture, or lighting. This paper proposes instead to first transform the query image itself into a personalized 'normal' version using a diffusion model fine-tuned on the few normal examples, so that anomalies are erased while normal structure is kept. The query is then scored in three complementary ways: against its personalized twin, against a pool of anomaly-free samples (real and generated), and against text prompts. The authors report that this triplet contrastive scheme outperforms existing few-shot anomaly detection methods across eleven datasets in industrial, medical, and semantic domains, with the largest gains in the 8-shot setting, and that the generated anomaly-free images can be dropped into other anomaly detectors to improve them.","feed_headline":"Turning each query into its own normal twin sharpens anomaly detection","feed_subtitle":"The method beats few-shot anomaly detection baselines on 11 industrial, medical, and semantic datasets","key_machinery":"The load-bearing mechanism is the one-to-normal transformation: a query image $x_q$ is partially noised to $x_{q,t}$ and denoised by the anomaly-free customized diffusion model $D_\\theta$ under the text prompt $c_q$ chosen to minimize SSIM loss against the query, producing a personalized image $\\hat{x}_q = D_\\theta(\\sqrt{\\alpha_t} x_q + \\sqrt{1-\\alpha_t}\\epsilon, c_q, t)$. This is the operation that maps the query onto the normal manifold, erasing anomalies while preserving identity. The second mechanism is the triplet contrastive anomaly inference, which pools three scores—$S_P$ from multi-level cosine dissimilarity between query and personalized image, $S_N$ from the nearest memory-bank feature among anomaly-free samples, and $S_{\\text{text}}$ from softmax over CLIP text features—into the final score. The customization itself rests on a Dreambooth fine-tune of Stable Diffusion with a class-specific prior-preservation loss so that the model generates normal-looking samples without language drift.","core_discovery":"The central claim is that precise and stable few-shot anomaly detection can be achieved by 'anomaly personalization': rather than matching the query to a few reference images, one matches it to its own one-to-normal transformation. An anomaly-free customized model—a diffusion model fine-tuned through Dreambooth on two to eight normal images per object, with data augmentation and prior-preservation loss—learns the normal distribution. At test time the query image is partially noised (t-step ratio 0.3) and denoised under a selected normal-state text prompt, yielding a personalized image that keeps normal regions and repairs defective ones. The final anomaly score combines three contrastive terms: multi-level feature similarity between query and personalized image ($S_P$), similarity between query and a memory bank of anomaly-free samples including generated ones ($S_N$), and CLIP text-prompt classification ($S_{\\text{text}}$), summed as $A_{\\text{score}} = S_P + \\alpha S_N + \\beta S_{\\text{text}}$. The paper reports that this beats previous few-shot AD methods on MVTec, VisA, KSDD, AFID, ELPV, OCT2017, BrainMRI, HeadCT, RESC, MNIST, and CIFAR-10, with AUROC gains up to about four points over the second-best method in 8-shot settings, and that generated normal images improve other detectors such as PatchCore, RegAD, WinCLIP, and InCTRL when added to their reference sets.","pith_inferences":["The one-to-normal transformation is essentially a per-query data augmentation; a testable extension is to use the personalized image as a 'normal twin' for anomaly localization via pixel-wise residual maps, which the paper only shows qualitatively.","Because the t-step ratio (0.3) is fixed across all datasets, one can infer that a per-domain or per-category t-step schedule could further improve results; the paper does not explore this.","The method's reliance on a few normal images per category means it cannot go zero-shot; an implicit consequence is that the same personalization machinery could be applied to a generic normal prior for unseen categories, but that would require the customized model to generalize across objects.","The triplet score weighting $\\alpha=1, \\beta=0.5$ is fixed; an inference is that the text branch matters less in medical domains where CLIP has weak domain knowledge, so adaptive weighting could shift the balance."],"forward_implications":["If the central claim holds, few-shot anomaly detection no longer needs to rely on direct query-to-reference feature matching; the query's own normal twin provides a finer-grained comparison.","The generated anomaly-free samples form a reusable augmentation: adding them to the reference pools of existing detectors improves their AUROC, with reported gains of up to about ten points for some baselines.","The method transfers across industrial, medical, and semantic domains, suggesting that the personalization approach is not tied to object categories seen during pretraining.","Because the customized model is frozen after Dreambooth fine-tuning and only the inference-time noise/denoise step is new, the pipeline can be grafted onto existing CLIP-based few-shot AD systems at a modest inference cost of a few hundred milliseconds per image."],"supporting_citations":[{"why":"Supplies the Dreambooth fine-tuning procedure that creates the anomaly-free customized model from a few normal images.","marker":"[31]"},{"why":"Provides the Stable Diffusion latent diffusion backbone used for generation and personalization.","marker":"[29]"},{"why":"Provides the CLIP image and text encoders used in the triplet contrastive anomaly inference.","marker":"[28]"},{"why":"Supplies the text prompt templates for normal and abnormal states and serves as a key few-shot vision-language baseline.","marker":"[16]"},{"why":"The strongest few-shot baseline and the source of dataset partitions for the eleven benchmark datasets.","marker":"[45]"},{"why":"Motivates the t-step noising/denoising schedule that edits the query image toward the normal manifold.","marker":"[26]"},{"why":"Justifies the choice of diffusion step $t$ for local sampling on the image manifold.","marker":"[25]"},{"why":"Provides the MVTec-AD industrial dataset used for evaluation and for auxiliary training data in the held-out protocol.","marker":"[2]"}],"fun_headline_variants":["Turning each query into its own normal twin sharpens anomaly detection","Personalized normal twin per query improves few-shot anomaly detection","Match query to its own normal image for precise anomaly detection","Anomaly personalization: transform query into its own normal","Triplet contrastive inference with personalized normals boosts detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method assumes that partially noising a query image at t-step 0.3 and denoising it with a diffusion model trained on two to eight normal images will erase the anomaly while preserving the object's normal appearance; if that transformation instead keeps the defect or destroys normal structure, the personalized comparison becomes uninformative or worse than direct reference matching.","fun_headline_variants_meta":{"raw":{"variants":["Turning each query into its own normal twin sharpens anomaly detection","Personalized normal twin per query improves few-shot anomaly detection","Match query to its own normal image for precise anomaly detection","Anomaly personalization: transform query into its own normal","Triplet contrastive inference with personalized normals boosts detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000939,"raw_usage":{"total_tokens":4082,"prompt_tokens":1078,"completion_tokens":3004,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":2921}},"tokens_in":694,"tokens_out":3004,"duration_ms":21425,"temperature":1.0,"reasoning_tokens":2921,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:13:29.116578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a defect-free but visually atypical object (e.g., a normal image with unusual pose, lighting, or background) as query, run the one-to-normal personalization, and measure whether the pixel-level difference between query and personalized image is concentrated in genuinely anomalous regions; if normal regions shift as much as anomalous ones, or if a deliberately injected defect survives the transformation, the central premise is refuted. Concretely, on MVTec-AD with known ground-truth masks, compute the AUROC of a defect-localization score defined by the per-pixel difference between $x_q$ and $\\hat{x}_q$; if it is no better than chance on categories with high pose variation, the personalization is not selectively removing anomalies.","supporting_citations":[{"cited_title":"Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation","cited_arxiv_id":null,"evidence_quote":"Supplies the Dreambooth fine-tuning procedure that creates the anomaly-free customized model from a few normal images."},{"cited_title":"Toward Generalist Anomaly Detection via In-context Residual Learning with Few-shot Sample Prompts","cited_arxiv_id":"2403.06495","evidence_quote":"The strongest few-shot baseline and the source of dataset partitions for the eleven benchmark datasets."},{"cited_title":"Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection","cited_arxiv_id":null,"evidence_quote":"Provides the MVTec-AD industrial dataset used for evaluation and for auxiliary training data in the held-out protocol."}],"review_version":1}