{"id":"0fe22e31-50ff-425e-962e-150e55a6f9d7","arxiv_id":"2412.05293","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FodFoM generates fake OOD images from ID semantics via shifted CLIP text embeddings and blurred backgrounds, and uses them to train classifiers that set new SOTA OOD detection scores on CIFAR10/100 and ImageNet100.","lead":"This paper trains image classifiers to detect out-of-distribution inputs using two types of fake outlier images: one synthesized by combining BLIP-2 captions, CLIP text embeddings, and Stable Diffusion, and one created by blurring foreground objects detected by GroundingDINO. It reports improved OOD detection scores over existing methods on CIFAR and ImageNet benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No quantitative verification that the Eq. 3 shifted-text images are OOD; if they sit on the ID manifold, the central mechanism is unsupported and the reported gains may come from extra training data rather than from genuinely OOD synthetic examples.","rationale":"The reader's weakest assumption is exactly the same soft spot: Eq. 3's geometric shift in CLIP text space is assumed to transfer to semantically OOD images, but no direct verification is provided. I agree this is the most load-bearing concern because it sits between the method's novel contribution (the fake-OOD text embedding construction) and the reported gains. If the generated images are not actually OOD, the method's stated mechanism is unsupported, even though the empirical recipe might still work accidentally through extra-data regularization. This is addressable and not a fatal flaw, so the conditional verdict stands; no change needed. The paper does deserve credit for ablations, clean-benchmark evaluation, and comparisons against Dream-OOD, which reduce but do not remove the need for direct data-level validation.","tokens_in":25494,"tokens_out":6708,"duration_ms":69426,"concrete_test":"Train the baseline model (e.g., ReAct on CIFAR10) and use its Energy score to score four sets: ID test images, real OOD test images, FodFoM's SD-OOD training images, and SD images generated with gamma=0 (unshifted text embeddings). Plot the score distributions. If the gamma=0 and gamma>0 distributions overlap and neither is clearly more OOD-like than ID, the Eq. 3 construction is not producing the claimed 'genuinely outside' images. As a second check, rerun the Table 4 ablation replacing SD-OOD with gamma=0 images; if FPR95/AUROC are unchanged, the outward shift is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanism (Eq. 3, Sec. 2.2.2) assumes that adding a small outward step gamma to a CLIP text embedding at the periphery of an ID class cluster yields a condition that makes Stable Diffusion generate images that are semantically similar to but genuinely outside the ID class. This assumption is never tested directly. The paper reports downstream OOD accuracy improvements and a few cherry-picked visual examples (Fig. 5), but no quantitative analysis of the generated images themselves: no off-the-shelf OOD score, no distance-to-ID measurement, no comparison to unshifted (gamma=0) generations. Because gamma is tiny (1e-5 to 1.5e-4) and chosen separately for CIFAR and ImageNet100, the shifted text embeddings are numerically very close to true ID embeddings; whether that translates into semantically OOD images is exactly what needs evidence. If the SD-OOD images are effectively on the ID manifold or are just generic synthetic images, then the improvement in Table 4 is not evidence for the proposed outward-shift mechanism; it could be produced by any extra training data added to the (C+1)-th class. The claim that fake OOD images help classifiers learn better decision boundaries therefore rests on an unverified property of the generated data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FodFoM, a training-based out-of-distribution (OOD) detection framework that generates two types of fake OOD images using foundation models. The first type is produced by BLIP-2 image captioning, CLIP text-embedding manipulation, and Stable Diffusion generation: captions of in-distribution (ID) images are embedded with CLIP, peripheral text embeddings are shifted outward from the class-mean embedding by a step gamma (Eq. 3), and the resulting fake text embeddings condition Stable Diffusion to synthesize images that are semantically close to but (by assumption) outside the ID classes. The second type is constructed by using GroundingDINO to detect foreground objects and blurring the detected regions to create pure-background images. The classifier is trained with an extra class for the fake OOD images using a cross-entropy loss and a supervised contrastive loss. Experiments on CIFAR10, CIFAR100, and ImageNet100 with ResNet18/34/50/101 report state-of-the-art FPR95/AUROC on most benchmarks and per-OOD-dataset comparisons, along with ablations, sensitivity studies for some hyperparameters, and a comparison with Dream-OOD in the supplementary material. The central claim is that classifiers trained with these fake OOD images learn better decision boundaries for OOD detection.","tokens_in":25767,"tokens_out":8267,"duration_ms":77591,"significance":"If the reported results are reproducible, the paper is a valuable practical contribution: it shows that off-the-shelf foundation models can be combined to generate challenging OOD training data without GAN training instability or restrictive feature-space assumptions. The evaluation is broad, covering three ID benchmarks, multiple backbones, six OOD test sets for CIFAR, clean OOD benchmarks (NINCO, OpenImage-O), a Dream-OOD comparison, and generalization studies for the captioning model and similarity metric. The authors also release code, and the framework can be fused with several post-hoc scores, which strengthens the paper's applicability. The main weakness is that the core generative mechanism—the outward text-embedding shift—is not quantitatively verified at the image level, and the absence of error bars leaves some state-of-the-art claims under-supported.","major_comments":[{"comment":"The paper claims that the shifted text embeddings produce 'challenging fake OOD images,' but it never quantitatively verifies that the generated images are OOD relative to the ID classes. Figure 5 shows only a few qualitative examples, and Figure 3 reports cosine similarities in the text-embedding space, which is the construction space rather than the image space. I request that the authors measure, for example, the feature distance between generated images and ID images using a pre-trained encoder, or apply an off-the-shelf OOD detector (e.g., one of the baselines from Table 1) to the generated images and report the fraction classified as OOD. Without such evidence, the reported gains cannot be attributed to the outward-shift mechanism.","section":"2.2.2, Eq. (3)"},{"comment":"The step length gamma is set to three or five different values per benchmark, but there is no sensitivity analysis for gamma and no stated criterion for choosing these values. Since the outward shift is the core mechanism that moves text embeddings outside the ID cluster, the choice of gamma is load-bearing: too small a gamma keeps the embedding inside the cluster (so the generated images are still ID-like), and too large a gamma may produce unrelated images. Please provide a sensitivity study over gamma and describe how the values were selected (e.g., on a validation OOD set). If a validation set was used, this should be disclosed to rule out test-set overfitting.","section":"3.1 and Eq. (3)"},{"comment":"No error bars or multiple-seed results are reported anywhere. Some of the claimed improvements over the strongest baselines are small (e.g., Table 2, ImageNet100 ResNet50 FPR95 33.44 vs. DICE+ReAct 34.75; Table 3, OpenImage-O FPR95 48.68 vs. LINe 62.15 is larger). To support the 'state-of-the-art' claim, the authors should report mean and standard deviation over at least three independent training runs for the main benchmarks, and ideally a significance test for the key comparisons.","section":"Tables 1–3 and 5"},{"comment":"The ablation demonstrates that adding the SD-OOD images improves performance, but it does not isolate the effect of the outward shift itself. A control with gamma=0 (i.e., using the periphery text embeddings without any outward step) should be included; if the performance gain disappears or reverses with gamma=0, this would confirm the role of the shift, whereas if the gain persists, the improvement may be due simply to adding extra synthetic images to the (C+1)-th class. Similarly, a control with random text embeddings would clarify whether semantic proximity to the ID classes matters.","section":"3.3, Table 4"}],"minor_comments":[{"comment":"Typo: 'emebdding' should be 'embedding' in the description of the hyperparameter gamma.","section":"3.1"},{"comment":"Typo: 'Stabel Diffusion' should be 'Stable Diffusion'.","section":"3.3"},{"comment":"Typo: 'achived' should be 'achieved' in the conclusion.","section":"5"},{"comment":"Typo: 'Comparsion' should be 'Comparison'.","section":"Supplementary, Figure 6 caption"},{"comment":"Typo: 'the the class of fake OOD images' should be 'the class of fake OOD images'.","section":"2.4"},{"comment":"Grammar: 'All results supports' should be 'All results support'.","section":"3.2.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be the camera-ready version of an ACM MM 2024 paper. If this is a journal submission, the authors should clarify in a cover letter or footnote the additional contribution beyond the conference version. The per-benchmark hyperparameter selection (alpha, gamma) and the absence of a disclosed validation set for these choices raise a mild risk of test-set overfitting; the revision should address this explicitly. The missing image-level verification of the outward-shift mechanism is the main scientific concern and should be the focus of the requested revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a practical, well-ablated recipe for training OOD detectors with synthetic outliers, and it deserves a serious referee. The weakest point is exactly the one flagged in the stress-test note: the paper never checks whether the images generated from the Eq. 3 shifted text embeddings are actually OOD with respect to the ID classes. The downstream results are consistent with the mechanism, but also consistent with the alternative that any extra synthetic data in the C+1 class would help.\n\nWhat is new: combining BLIP-2 captions, CLIP text-space geometry, Stable Diffusion, and GroundingDINO foreground blurring in one pipeline. The outward-shift construction in CLIP text space (Eq. 3) is a simple, cheap trick that Dream-OOD, NPOS, and VOS do not use. The empirical package is broad: CIFAR10/100 with ResNet18/34, ImageNet100 with ResNet50/101, six OOD sets plus clean NINCO/OpenImage-O, ablations, fusion with post-hoc scores, a direct comparison with Dream-OOD, and robustness checks on similarity metric and captioning model. Code is available. That is genuine value.\n\nSoft spots, in proportion. First, no error bars anywhere; the SOTA claim rests on point estimates. Second, alpha and gamma are set per benchmark and gamma is never swept; the sensitivity study covers tau, lambda, alpha only. Third, and most important, the central assumption is under-verified. The paper shows a cosine-similarity plot of the constructed text embeddings (Fig. 3), which confirms the prompts sit outside the ID text cluster, but it never verifies the generated images are OOD. No off-the-shelf OOD score, no distance-to-ID in image or feature space, no comparison to unshifted (gamma=0) generations. Without a gamma=0 control, the ablation gains attributed to SD-OOD could come from adding any diverse synthetic data, not from the outward-shift mechanism. The paper should also address the possible overlap between Stable Diffusion's training data and the OOD test sets; that is an unexamined source of inflation.\n\nNone of this is fatal. The pipeline is reproducible, the ablations are internally consistent, and the gains are large across multiple backbones and benchmarks. The right outcome is a revision that adds error bars, a gamma sensitivity study, a gamma=0 control, and a direct OOD-ness check on the generated images.\n\nBottom line: send it to peer review. It would be a useful citation for anyone working on synthetic outlier training, but the mechanistic claim needs the missing controls before I would take the SOTA statement at face value.","headline":"Practical fake-OOD training pipeline with strong ablations, but the core outward-shift mechanism is never verified on the generated images.","tokens_in":26321,"tokens_out":2850,"would_cite":true,"duration_ms":27926,"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":"FodFoM shows that training image classifiers on fake outlier images synthesized by foundation models—semantically similar to but different from in-distribution images—produces stronger out-of-distribution detection, reaching…","keywords":["Out-of-Distribution Detection","Foundation Models","Fake OOD Image Generation","Stable Diffusion","CLIP Text Embeddings","BLIP-2","GroundingDINO","Supervised Contrastive Learning"],"falsifier":"Feed the Stable Diffusion generated fake OOD images into a strong OOD detector trained only on ID data (e.g., ViM or Energy with ReAct on the same backbone). If most fake images are scored as firmly ID (high energy score) or as arbitrary noise (extreme low density), the construction failed to produce near-boundary outliers. A second check is a sweep of the step size gamma: if performance does not peak in the middle and degrade at very small and very large gamma, the boundary-nudging story is not what drives the gain.","tokens_in":25300,"feed_emoji":"🖼️","tokens_out":6889,"duration_ms":57477,"temperature":0.7,"pith_summary":"The paper tries to establish that a classifier's decision boundary between in-distribution and out-of-distribution images can be sharpened by training on synthetic 'fake outlier' images that are deliberately close to the in-distribution classes. It builds these images by chaining BLIP-2 captions, CLIP text embeddings, and Stable Diffusion to synthesize images that differ slightly from each ID class, plus GroundingDINO-based blurring to turn ID images into background-only versions. When these fake outliers are added to training as an extra class, with supervised contrastive loss, the model becomes less overconfident and detects real OOD images better. The paper reports new state-of-the-art results on CIFAR10, CIFAR100 and ImageNet100 benchmarks, and shows the gains hold across backbones and combine with post-hoc OOD scores.","feed_headline":"Foundation-model fake outliers plus blurred backgrounds tighten OOD decision boundaries","feed_subtitle":"Near-miss synthetic outliers plus blurred-ID backgrounds give tighter decision boundaries.","key_machinery":"The central object is the fake OOD text embedding construction of Eq. (3), which takes a CLIP text embedding at the periphery of an ID class cluster and moves a small step gamma outward along the unit vector from the class mean; the resulting embedding conditions Stable Diffusion to generate an image that stays visually close to the ID class while falling outside it. A second mechanism uses GroundingDINO to detect foreground ID objects and blurs those regions to produce background-only fake OOD images. Together these define an extra (C+1)-th class during training, separated from ID classes by cross-entropy loss and supervised contrastive loss, and at inference only the original C logits feed the Energy-based OOD score.","core_discovery":"The central claim is that classifiers trained with the constructed fake OOD images learn better decision boundaries for effective OOD detection. Concretely, on CIFAR10 with ResNet18 the average FPR95 drops to 8.43% and AUROC rises to 98.33%, versus 14.64% and 97.10% for the strongest baseline CIDER; on ImageNet100 with ResNet50 the average FPR95 is 33.44% and AUROC 93.79%, versus 36.46% and 92.11% for DICE. The benefit is attributed to the fact that the fake outliers are semantically similar to ID images (via the CLIP text-space perturbation) and share background context with ID (via GroundingDINO blurring), so the model is forced to separate ID from near-miss OOD cases.","pith_inferences":["One implicit implication is that the CLIP text-embedding perturbation of Eq. (3) is a transferable device: the same outward-step construction could be applied to any CLIP-based OOD detector without image generation, e.g., as negative prompts for zero-shot detection.","The background-blur component suggests a cheap, model-agnostic augmentation for any OOD method: any detector of foreground objects could replace GroundingDINO, making the recipe lighter.","Because the pipeline depends on the generative fidelity of Stable Diffusion and the alignment of CLIP text space, its quality will track improvements in those foundation models; the OOD gains are likely to grow as the generators improve.","A testable extension would be to use the same framework on fine-grained ID classes or medical images, where the 'semantically similar but different' construction may be more or less effective depending on how separable the classes are in CLIP space."],"forward_implications":["Classifiers trained with the generated fake OOD images can serve as stronger OOD detectors for the same backbone and ID dataset, improving average FPR95 and AUROC over existing training-based and post-hoc methods.","Existing post-hoc scores (MSP, Energy, ViM, ReAct) all improve when applied on the FodFoM-trained backbone, so the training recipe is score-agnostic.","The method generalizes to clean, challenging OOD benchmarks such as NINCO and OpenImage-O, where conventional evaluation sets are known to be contaminated.","The framework needs no external real OOD data and, per the comparison with Dream-OOD, reaches better average performance with fewer generated images.","The benefits persist across ResNet18/34/50/101 backbones, indicating the effect is not tied to one architecture."],"supporting_citations":[{"why":"BLIP-2, the image-to-text model that generates captions for ID images, anchoring the fake text embeddings in ID semantics.","marker":"[27]"},{"why":"CLIP, whose text encoder defines the semantic space in which fake OOD text embeddings are constructed via Eq. (3).","marker":"[41]"},{"why":"Stable Diffusion, which turns the fake OOD text embeddings into the first type of fake OOD images.","marker":"[44]"},{"why":"GroundingDINO, the open-set detector used to locate and then blur foreground ID objects to build background fake OOD images.","marker":"[30]"},{"why":"Outlier exposure, the training-with-auxiliary-OOD idea this paper extends to synthetically generated outliers.","marker":"[17]"},{"why":"Dream-OOD, the main diffusion-based fake-OOD competitor; the paper compares directly against it and uses fewer generated images.","marker":"[11]"},{"why":"CIDER, the strongest training-based baseline on CIFAR10 that FodFoM outperforms.","marker":"[35]"},{"why":"ReAct, the post-hoc scoring method used at inference on the FodFoM-trained backbone.","marker":"[51]"}],"fun_headline_variants":["FodFoM: Fake outliers from foundation models sharpen OOD detection","Near-miss synthetic outliers tighten OOD boundaries","Foundation models craft fake outliers for stronger OOD detection","Fake outlier images from BLIP-2, CLIP, and Stable Diffusion boost OOD","Blurred backgrounds plus fake outliers give state-of-the-art OOD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that taking a CLIP text embedding near the edge of an ID class cluster and nudging it outward a small step produces a text prompt whose Stable Diffusion image is both visually plausible and genuinely outside the ID class; the paper does not quantitatively verify the generated images with an independent OOD score, and the step size is tuned per benchmark without sensitivity analysis.","fun_headline_variants_meta":{"raw":{"variants":["FodFoM: Fake outliers from foundation models sharpen OOD detection","Near-miss synthetic outliers tighten OOD boundaries","Foundation models craft fake outliers for stronger OOD detection","Fake outlier images from BLIP-2, CLIP, and Stable Diffusion boost OOD","Blurred backgrounds plus fake outliers give state-of-the-art OOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000867,"raw_usage":{"total_tokens":3769,"prompt_tokens":972,"completion_tokens":2797,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2705}},"tokens_in":588,"tokens_out":2797,"duration_ms":19311,"temperature":1.0,"reasoning_tokens":2705,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:30:38.212352+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the Stable Diffusion generated fake OOD images into a strong OOD detector trained only on ID data (e.g., ViM or Energy with ReAct on the same backbone). If most fake images are scored as firmly ID (high energy score) or as arbitrary noise (extreme low density), the construction failed to produce near-boundary outliers. A second check is a sweep of the step size gamma: if performance does not peak in the middle and degrade at very small and very large gamma, the boundary-nudging story is not what drives the gain.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Outlier exposure, the training-with-auxiliary-OOD idea this paper extends to synthetically generated outliers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Dream-OOD, the main diffusion-based fake-OOD competitor; the paper compares directly against it and uses fewer generated images."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CIDER, the strongest training-based baseline on CIFAR10 that FodFoM outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ReAct, the post-hoc scoring method used at inference on the FodFoM-trained backbone."}],"review_version":1}