{"id":"c490823f-8b60-4ab9-b0a8-1b4fcad8d278","arxiv_id":"2412.02693","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A diffusion-based method generates visual anagrams by treating each viewpoint as a task and adding anti-segregation, noise-balancing, and variance-rectification steps.","lead":"This paper improves a diffusion-model trick for creating visual anagrams, images that look like one object when flipped or rotated and another object when viewed normally. It treats each viewpoint as a task in a multi-task learning setup and adds three fixes to keep the concepts blended and balanced.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unverified Eq. (7) Gaussian-noise assumption undermines the NVR scale factor derivation in Sec. 3.4.","rationale":"After reading the paper in good faith, I agree with the reader that the weakest assumption is the Gaussian noise premise in Sec. 3.4. The NVR contribution is presented as a principled derivation, yet Eq. (7) is not an established property of diffusion model predictions; it is a convenient assumption. If it fails, the computed scale factor does not achieve the stated goal of unit variance. The paper's own Fig. 10 verifies correlation but not marginal variance, so the assumption is genuinely untested. This matters because NVR shows an isolated gain in Table 2 (Aavg 0.2717 to 0.2784) and is part of the full method; if the derivation is invalid, the component's functioning is unexplained and the central claim about 'statistical properties may not be preserved' loses its foundation. I also note secondary concerns about missing error bars and test-set hyperparameter tuning, but those are empirical robustness issues rather than a specific flaw in the argument. The proposed concrete test directly targets the assumption by measuring the actual distribution of predicted noise and checking whether the rectification would change with correct variances. If the test passes, the concern lands as minor and the verdict stands. If it fails, the authors would need to either revise the NVR derivation or provide empirical evidence that the unit-variance scaling still improves results for the right reason. Therefore I keep the reader's CONDITIONAL verdict.","tokens_in":14443,"tokens_out":12278,"duration_ms":124508,"concrete_test":"Instrument the released code to record predicted noise vectors epsilon_i^t for both views on the 2-view CIFAR-10 benchmark across all timesteps and several seeds. First, compute the per-element variance of each view's predicted noise; check whether it is within 10% of 1.0 at every timestep. Second, split the spatial map into patches and estimate rho_ij per patch via Eq. (12); verify the patch-wise estimates are tightly clustered (e.g., coefficient of variation < 0.2). Third, recompute the scale factor c using the measured per-view variances in Eq. (10) and re-run the Table 2 ablation. If the NVR-only row's Aavg changes by more than ~0.007 or the full model's Aavg shifts by more than ~0.005, the rectification is miscalibrated and the theoretical claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Noise Variance Rectification (NVR) preserves noise statistics rests on Eq. (7): for all i, epsilon_i^t ~ N(0,I). In practice, the diffusion model's predicted noise is a deterministic function of x_t, t, and prompt; it is a conditional expectation estimate, not a sample from a standard Gaussian. The paper provides no empirical check of the marginal variance or distribution of the predicted noise. If the variance is not 1, the scale factor c computed by Eq. (10) using the assumed unit variances will not set the combined noise variance to 1, and the rectification can systematically distort the denoising trajectory. Additionally, Eq. (12) estimates the covariance as a single spatial average, assuming the cross-view correlation is constant across all C x H x W elements; this is also unverified. Since NVR is one of the three core techniques and its isolated ablation (Table 2) shows a non-trivial gain, a miscalibrated scale factor would undermine the claimed mechanism, even if the empirical benefit could survive for other reasons.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses two failure modes in diffusion-based visual anagram generation: concept segregation (different concepts are generated independently rather than sharing a single image) and concept domination (one concept overwhelms the others). It casts anagram generation as a multi-task learning problem and contributes three inference-time techniques: Anti-Segregation Optimization (ASO), which encourages overlap between cross-attention maps of different views via a loss term; Noise Vector Balancing (NVB), which reweights each view's predicted noise by a CLIP-based task completion score; and Noise Variance Rectification (NVR), which scales the combined noise to restore unit variance under an assumed Gaussian noise model. Experiments on 2-view and 3-view CIFAR-10 prompts, plus free-form examples, compare against the previous baseline (Geng et al., 2024) and report consistent improvements in worst alignment, concealment, and average alignment, with an ablation study attributing gains to each component.","tokens_in":14653,"tokens_out":3766,"duration_ms":37421,"significance":"If the reported improvements are robust, the paper offers a practical, training-free way to improve visual anagram quality using only off-the-shelf diffusion and CLIP models, which is useful for a growing community working on diffusion-based illusions. The multi-task learning analogy is clearly presented, and the ablations show that each proposed component contributes positively to at least one metric. The paper also ships code, which supports reproducibility. However, the quantitative gains are small (on the order of 0.01 in cosine-similarity-based metrics), error bars are absent, and at least one hyperparameter (phi) appears to be tuned on the evaluation set, so the significance of the headline improvements is currently uncertain.","major_comments":[{"comment":"The NVR derivation is built on the assumption in Eq. (7) that each predicted noise vector epsilon_i^t is standard Gaussian. In practice, epsilon_theta(x_t, t, y_i) is a deterministic function of the current noisy image and the prompt, and its marginal distribution may differ from N(0, I). The paper never verifies this assumption, and Fig. 10 only reports cosine similarity and the correlation coefficient of Eq. (12), not marginal variances. If the variance of the predicted noises is not 1, the scalar c computed in Eq. (10) will not set the combined noise variance to 1, so the claimed statistical rectification mechanism is unsupported. Please add an empirical check of the marginal variance and distribution of predicted noises across views and timesteps, and verify that the rectified noise indeed has unit variance; if it does not, the explanation for the NVR improvement in Table 2 would need to be revised.","section":"Sec. 3.4, Eqs. (7)-(12)"},{"comment":"The target overlap hyperparameter phi is selected on the same 2-view CIFAR-10 evaluation set that is used for the headline comparisons: Fig. 9 reports that the best performance is achieved around phi=0.45, and the main tables presumably use this value. This gives the proposed method an advantage over the baseline, which has no equivalent evaluation-set tuning, and makes the reported gains optimistic. Please either fix phi before evaluation, tune it on a separate validation split, or report the main results across a range of phi values so readers can assess sensitivity.","section":"Sec. 5.2 and Fig. 9; Table 2"},{"comment":"All metrics are averages over 10 seeds per prompt pair, but no standard deviations, confidence intervals, or per-pair breakdowns are provided. The reported improvements are small: for example, Amin rises from 0.2583 to 0.2711 and Aavg from 0.2717 to 0.2816, while several ablation rows differ by less than 0.01 in C. Without error bars or a paired significance test over the 10 seeds, it is difficult to judge whether the claimed 'superior ability' and the component-wise contributions are statistically reliable. Please include error bars or a bootstrap test.","section":"Table 2 and Sec. 4.2"}],"minor_comments":[{"comment":"There is a typo in the third contribution: 'satet-of-the-art' should be 'state-of-the-art'.","section":"Sec. 1, Contributions"},{"comment":"The loss in Eq. (1) penalizes absolute deviation from the target overlap phi, so it does not monotonically push toward larger overlap for ratios above phi. The intuitive description in Sec. 5.2 ('a larger phi encourages more overlap') is acceptable, but the exact behavior near the target could be clarified in the text.","section":"Eq. (1) and Sec. 5.2"},{"comment":"The noise-aware CLIP image encoder is taken from GLIDE [39], but it is not stated whether this model has the same architecture and embedding space as the vanilla CLIP used for evaluation. Please specify the relation between the two CLIP models for reproducibility.","section":"Sec. 3.3"},{"comment":"Score Distillation Loss appears as '[?]' in the appendix text and figure caption; it should be a proper citation, e.g., to Poole et al. [43] which is already in the reference list.","section":"Appendix A and Fig. 11"},{"comment":"The formula for c in Eq. (10) is rendered with an unclear square-root symbol; please ensure the mathematical typesetting is unambiguous.","section":"Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental but potentially useful extension of Geng et al. (2024). The main technical concern is the unverified Gaussian-noise assumption behind NVR, which should be addressable with additional experiments. The lack of error bars and the evaluation-set tuning of phi are also common issues, but together with the small effect sizes they currently make the central quantitative claims look fragile. I would support publication if the authors can show that the improvements are statistically meaningful and that the NVR mechanism is actually doing what is claimed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a solid, honest incremental paper on a niche problem. It is the first to frame multi-view visual anagram generation as multi-task learning, and it backs the framing with three inference-time techniques that beat the Geng et al. baseline on every reported metric. The NVR derivation is mathematically clean under its assumptions, and the correlation analysis in Sec. 5.3 is a nice explanation of why the baseline still works. Worth a serious referee, though the evidence is not airtight.\n\nWhat's genuinely new: the multi-task framing itself, the anti-segregation loss that pushes cross-attention maps of different views to overlap, and the use of a noise-aware CLIP model to balance noise vectors. These are simple, sensible ideas, and the ablations show each component contributes a small but consistent gain. The code is released and the method runs on a single 3090 without fine-tuning, which is reproducible and good practice.\n\nThe main soft spot is the Gaussian assumption in Eq. (7). The paper assumes the diffusion model's predicted noise for each view is standard Gaussian. That is not generally true; the model outputs a conditional expectation, not a random draw from N(0,I). If the marginal variance of the predicted noise differs from 1, the scale factor in Eq. (10) is miscalibrated. The paper never checks this. The empirical correlation measurements in Sec. 5.3 suggest the effect may be small, but the claimed mechanism is not verified. Since NVR is one of the three pillars and its isolated ablation shows a nontrivial gain, this deserves an explicit test.\n\nOther concerns are moderate. Tables 1 and 2 report no error bars, and the gains are a few hundredths in CLIP similarity; without variance estimates, some of those differences could be noise. The target overlap phi is tuned on the same evaluation set, which is a mild circularity. The evaluation is limited to CIFAR-10 style prompts and flips/rotations, which is a narrow but honest scope. The authors state in their limitations that they cannot handle latent diffusion models or non-orthogonal transformations; I'd like to see that discussed in the main text, not buried.\n\nOverall, the central claim holds up: the method does produce better anagrams than the baseline, and the framing is useful. The weak spots are about verification, not the idea. I'd send it to review, with requests for error bars, a direct check of Eq. (7), and a validation split for phi. For a niche venue or workshop it's a clean accept; for a top conference, it needs those experiments.","headline":"Solid incremental paper on visual anagrams: new multi-task framing, clean math, but the NVR derivation rests on an unverified Gaussian-noise assumption and the empirical evidence lacks error bars.","tokens_in":15171,"tokens_out":3546,"would_cite":false,"duration_ms":36439,"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":"Casting visual anagram generation as multi-task learning, with anti-segregation attention overlap, completion-based noise reweighting, and variance rectification, produces anagrams that avoid concept segregation and domination.","keywords":["visual anagrams","diffusion models","multi-task learning","optical illusions","cross-attention maps","noise vector balancing","noise variance rectification","text-to-image generation"],"falsifier":"Record the per-element variances and pairwise correlations of the predicted noise vectors from the pixel-space diffusion model across prompts and timesteps; if the variances are not close to one or the correlation varies strongly across spatial locations, the single scale factor in the rectification equation cannot restore unit variance, and the rectification step rests on a false premise. A simpler check: replace the rectified noise with an empirical re-standardization per channel and see whether image quality changes.","tokens_in":14242,"feed_emoji":"🎭","tokens_out":6559,"duration_ms":62234,"temperature":0.7,"pith_summary":"This paper claims that visual anagram generation with diffusion models fails when noise predictions from different views are simply averaged, because the concepts split apart or one dominates. It reframes the problem as multi-task learning: each view is a task, and the denoising trajectory should be a compromise that serves all tasks at once. The authors introduce three interventions: encouraging cross-view overlap in attention maps, reweighting each view's noise by how close that view already is to completion, and rescaling the combined noise to restore the variance the diffusion model expects. On two-view and three-view benchmarks they report higher worst-view alignment, better concealment, and better average text-image alignment than the averaging baseline, without training or fine-tuning the diffusion model.","feed_headline":"Treating views as tasks fixes visual anagram illusions","feed_subtitle":"Noise reweighting and attention overlap stop one concept from hiding another—no training.","key_machinery":"Three mechanisms carry the argument. Anti-segregation optimization operates on the cross-attention maps of the subject tokens in each view and minimizes the difference between a target overlap ratio and the actual intersection-over-union of those maps, applying a one-step gradient modulation to the denoised image. Noise-vector balancing measures each view's task completion as the cosine similarity between a text embedding and a noise-aware image embedding of the currently denoised image, then reweights that view's predicted noise inversely to its completion with a time-dependent exponent. Noise-variance rectification treats the combined noise as a weighted sum of per-view noise vectors assumed to be standard normal with a constant per-element correlation, estimates that correlation from the dot product of the two noise maps, and rescales the combined noise by the single factor that restores unit variance.","core_discovery":"The central claim is that the two failure modes of prior visual anagram methods, concept segregation and concept domination, are symptoms of treating each view's generation as independent, and that a multi-task learning view of the denoising trajectory fixes both. The paper's contributions are the anti-segregation loss, which pushes the spatial attention maps of different prompts' subject tokens to overlap in the shared image; the noise-vector balancing rule, which down-weights views whose task-completion score is already high; and the noise-variance rectification, which multiplies the averaged noise by a scale factor derived from the estimated correlation between views so the combined noise keeps unit variance. In the paper's own terms, the result is a denoising trajectory that aligns well across tasks, generating images that read as one concept under one view and a different concept under the transformed view rather than two objects pasted into one frame.","pith_inferences":["The variance-rectification step is generic: any method that averages or interpolates noise predictions from multiple prompts, not just anagram generation, could inherit unit-variance corrections, and its benefit could be tested on ordinary multi-concept composition.","The target overlap ratio is a hyperparameter tuned per dataset; one could make it prompt-adaptive, raising it for concepts that naturally share silhouette structure and lowering it for unrelated pairs, which the paper does not explore.","The correlation coefficients the paper measures are close to one across timesteps, which suggests the main practical role of rectification is protective rather than large-magnitude; ablations show it mostly lifts average alignment, so its effect may be on image quality rather than on concept blending.","If the predicted noise from a diffusion model is not actually standard normal, the single scalar correction will not restore exact statistics; checking the marginal distribution of the predicted noise would resolve whether the derivation's premise holds."],"forward_implications":["On the paper's two-view CIFAR-10-style benchmark, the full method raises worst-view alignment from 0.2583 to 0.2711 and concealment from 0.6744 to 0.6913 relative to the averaging baseline.","The same gains hold on three-view prompts, indicating that the balancing and rectification machinery scales beyond pairs of views.","Each of the three components contributes independently, and the best results come from their combination, so the multi-task framing is doing real work rather than one trick carrying the method.","Because the pipeline needs no training or fine-tuning, the method can be plugged into any pixel-space diffusion model that exposes cross-attention maps and per-view noise predictions."],"supporting_citations":[{"why":"The baseline visual-anagram method that averages noise predictions; the paper compares against it and builds on its pipeline.","marker":"[19]"},{"why":"Establishes the denoising assumption that predicted noise is standard Gaussian, which the variance-rectification derivation relies on.","marker":"[26]"},{"why":"Supplies the multi-task-learning principle of reweighting tasks by training progress, which noise-vector balancing adapts.","marker":"[11]"},{"why":"Releases the noise-aware image-text model used to measure task completion scores.","marker":"[39]"},{"why":"Provides the dataset setup and evaluation metrics (worst alignment and concealment) that the paper follows.","marker":"[18]"},{"why":"The pixel-space diffusion model used as the backbone, chosen because layout is formed before fine details.","marker":"[1]"}],"fun_headline_variants":["Multi-task view unifies diffusion anagram generation","Noise reweighting and attention overlap fix anagram illusions","Solving concept segregation and domination in diffusion anagrams","Diffusion anagrams via multi-task denoising alignment","Attention overlap plus adaptive noise create true anagrams"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The rectification step assumes that the diffusion model's predicted noise for every view is standard Gaussian and that the correlation between views is the same for every pixel, so a single multiplicative constant can restore the correct statistics; if the predicted noise is not standard Gaussian, the correction can itself distort the denoising trajectory.","fun_headline_variants_meta":{"raw":{"variants":["Multi-task view unifies diffusion anagram generation","Noise reweighting and attention overlap fix anagram illusions","Solving concept segregation and domination in diffusion anagrams","Diffusion anagrams via multi-task denoising alignment","Attention overlap plus adaptive noise create true anagrams"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000533,"raw_usage":{"total_tokens":2556,"prompt_tokens":926,"completion_tokens":1630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":1554}},"tokens_in":542,"tokens_out":1630,"duration_ms":12589,"temperature":1.0,"reasoning_tokens":1554,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:09:38.399201+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the per-element variances and pairwise correlations of the predicted noise vectors from the pixel-space diffusion model across prompts and timesteps; if the variances are not close to one or the correlation varies strongly across spatial locations, the single scale factor in the rectification equation cannot restore unit variance, and the rectification step rests on a false premise. A simpler check: replace the rectified noise with an empirical re-standardization per channel and see whether image quality changes.","supporting_citations":[{"cited_title":"Visual ana- grams: Generating multi-view optical illusions with diffu- sion models","cited_arxiv_id":null,"evidence_quote":"The baseline visual-anagram method that averages noise predictions; the paper compares against it and builds on its pipeline."},{"cited_title":"Denoising diffu- sion probabilistic models, 2020","cited_arxiv_id":null,"evidence_quote":"Establishes the denoising assumption that predicted noise is standard Gaussian, which the variance-rectification derivation relies on."},{"cited_title":"Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-task-learning principle of reweighting tasks by training progress, which noise-vector balancing adapts."},{"cited_title":"Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022","cited_arxiv_id":null,"evidence_quote":"Releases the noise-aware image-text model used to measure task completion scores."},{"cited_title":"https://stability.ai/news/deepfloyd-if-text-to-image- model","cited_arxiv_id":null,"evidence_quote":"The pixel-space diffusion model used as the backbone, chosen because layout is formed before fine details."}],"review_version":1}