{"id":"8e809807-1e69-4f9e-9271-734d5ed70de9","arxiv_id":"2505.07481","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Centroids of many diffusion latents degenerate because small common channel-wise mean biases are amplified by roughly sqrt(N); subtracting each channel's mean before interpolation and adding it back after largely fixes the problem.","lead":"This paper identifies why blending many diffusion-model latents, for example to morph images or create training data, produces broken images: tiny per-channel biases get amplified as the number of blended inputs grows. It proposes a simple channel-wise mean adjustment that removes most of the degeneration and also improves quality for small input sets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Spatially structured or per-example bias, left open by the paper's own future-work list, would survive channel-mean subtraction and still be amplified by the Eq. (9) mechanism, so the remedy is only established for per-channel-constant bias.","rationale":"The paper's contribution rests on Eq. (9): a small common bias d is amplified by sqrt(N) under fFIX/fNIN, and the remedy removes d by channel-wise mean subtraction. The reader's conditional verdict already hinges on whether d is well approximated by a per-channel constant. My reading sharpens this: Section 4.3 itself identifies non-zero terminal SNR as a bias source with z_T = sqrt(alpha_T) z_0 + sqrt(1-alpha_T) epsilon, in which the deterministic component is image-specific and spatially structured, not a shared constant-per-channel vector. A bias of this form would survive channel-mean subtraction and be amplified in exactly the same way; the paper lists low-pass-filtered d as future work, which is an explicit admission that the chosen estimator may miss such structure. The experiments do not rule this out because Figure 4 tracks only the channel-wise mean of the subtracted component and Section 6 reports no residual statistics. I do not see an internal inconsistency: Eq. (9) is a reasonable approximation under the stated b^2 << 1/N condition, and the qualitative image results for ImageNet classes support that the per-channel-constant component is substantial in those settings. The concern is therefore about scope, not soundness, and it matches the reader's weakest_assumption; the conditional verdict should stand. Independent checks that would strengthen the paper include releasing code/data and adding error bars, but the decisive test is the residual-structure measurement described above.","tokens_in":10540,"tokens_out":10661,"duration_ms":102553,"concrete_test":"Compute the fFIX/fNIN centroid of the residuals e_n = z_n - mean_ch(z_n) for real inverted latents at N = 2, 4, 8, 16, 32, 48, 64, 96, and compare the centroid's norm and low-pass-filtered energy against the N(0, I) null distribution for the same N and dimensionality. If the residual centroid's low-frequency energy grows with N or exceeds the null distribution's upper tail, a spatially structured bias is still being amplified, confirming that the channel-mean-only remedy is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim, Eq. (9), derives sqrt(N) amplification for a common deterministic bias d that is constant over spatial positions. The proposed channel-wise mean adjustment removes only the per-channel spatial mean of each z_n before residual normalization. Section 4.3 identifies a concrete bias source with a different structure: with non-zero terminal SNR, z_T = sqrt(alpha_T) z_0 + sqrt(1-alpha_T) epsilon, where z_0 is image-specific; the deterministic component therefore varies across examples and, like any image latent, has spatial structure. The paper's own Section 5 lists low-pass-filtered d as an open option, acknowledging that a constant-per-channel d need not capture the true bias. If a non-constant bias component survives channel-mean subtraction, it enters the residual centroid and is amplified by the same sqrt(N) normalization that Eq. (9) describes. The experiments in Section 6 do not measure this residual: Figure 4 only tracks the channel-wise mean of the subtracted component, not the spatial structure or growth of the residual after chm. Consequently, the empirical support for the remedy is conditional on the bias being dominated by per-channel constants; the paper does not establish this for general inputs or for the terminal-SNR mechanism it invokes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies degeneracies in latent-space interpolation for diffusion models when computing centroids of N inverted latents. It observes that even with norm normalization (fFIX or fNIN), the generated images degenerate as N grows, and it attributes this to a small deterministic bias d in the latents that is amplified by roughly sqrt(N) during normalization. The central theoretical statement is Eq. (9), derived in Section 4.2 for a toy model z_n = d + e_n with i.i.d. Gaussian e_n. The paper then proposes a channel-wise mean adjustment (chm): subtract each channel's mean from each latent, interpolate the residuals with norm normalization, and add back the interpolated means. Experiments on ImageNet classes with Stable Diffusion 1.5 and 3.5 measure FID and CLIP distance for N up to 64 and show that chm improves quality relative to the baselines. The authors also discuss the non-zero terminal SNR issue and alternative inversion procedures, and they explicitly leave spatially structured bias models (e.g., low-pass-filtered d) to future work.","tokens_in":10767,"tokens_out":4471,"duration_ms":44995,"significance":"If the diagnosis and remedy hold, the work is practically valuable for image morphing and data augmentation, where multi-input latent interpolation is increasingly common. The theoretical derivation in Section 4.2 is clean, parameter-free, and consistent with the measured channel-wise mean growth in Figure 4. The paper is also honest about the limitations of FID and about the restricted bias model it considers. The main strengths are the simplicity of the proposed fix and the breadth of the empirical evaluation across two model families and several inversion procedures. However, the significance of the claimed remedy is conditional on the assumption that the bias is dominated by per-channel constants, which the paper acknowledges but does not directly verify in the experiments.","major_comments":[{"comment":"The derivation in Eq. (9) models the latents as z_n = d + e_n with a common deterministic bias d, and the proposed channel-wise mean adjustment in Section 5 removes only the per-channel spatial mean of d. If d also has a spatially varying component, or if the class-consistent part of the bias varies across examples as suggested by the terminal-SNR term sqrt(alpha_T) z_0 in Section 4.3, then that component survives the chm subtraction and is still amplified by the same sqrt(N) mechanism. The paper itself lists low-pass-filtered d as a possible decomposition in Section 5, indicating that a per-channel constant d is not guaranteed to capture the true bias. The experiments in Section 6 do not measure the residual bias after chm (e.g., the norm of the spatially varying part of the average latent), so the empirical support for the remedy is only established for biases that are approximately per-channel constant. To justify the broad claim that the method 'significantly reduces the degeneration effect', the authors should either directly measure the residual non-constant component or explicitly restrict the claim to the tested setting.","section":"Section 4.2 and Section 5"},{"comment":"The primary quantitative evidence consists of FID and CLIP distance curves without error bars or per-class spread, despite being averaged over only 10 randomly drawn ImageNet classes. At moderate N, the differences between the chm methods and the baselines, and between fix/chm and nin/chm, may be within the natural variation across classes. Reporting only point estimates makes it difficult to assess whether the reported improvements are statistically significant, especially since the paper claims a 'significant' reduction in degeneration. Please report standard errors, confidence intervals, or per-class curves, and state in the caption or text that the curves are averages over 10 classes.","section":"Section 6, Figure 7"}],"minor_comments":[{"comment":"The notation 'Pwn = 1' should be typeset with a summation symbol and limits; currently it appears as an ambiguous 'P' rather than a sum.","section":"Section 3.2"},{"comment":"The expression for fFIX(Z,{1,0,0,...}) has a missing delimiter; it should read sqrt(L)/||z1|| * z1 to avoid ambiguity.","section":"Section 3.3, Eq. (2)"},{"comment":"The sentence 'we can conclude that inverted latents do not in general follow the statistics of random samples drawn from a normal distribution' is a strong conclusion drawn from a few qualitative examples; a more cautious phrasing would be appropriate.","section":"Section 4.1"},{"comment":"The caption says 'centroids of N noisy diffusion model latents' but does not specify that these latents come from inverting real ImageNet images (as described in Appendix A); please clarify.","section":"Section 6, Figure 7 caption"},{"comment":"The qualitative offset experiment in Figure 6 applies a constant offset only to the top quarter of one channel, which is a useful illustration but not a quantitative measurement of the spatial structure of the bias; a brief note acknowledging this limitation would strengthen the presentation.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely a good fit for a computer vision venue, and the theoretical derivation is a genuine contribution. The main concern is that the proposed remedy is not shown to work for spatially structured or per-example-varying bias, which the paper itself raises as future work. The experimental section would also benefit from error bars or a statistical test. These issues are addressable with additional analysis rather than requiring a fundamentally new approach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper finds something real: when you average many inverted diffusion latents and renormalize, small per-channel biases get amplified by roughly sqrt(N), and the output degenerates. The derivation of Eq. (9) is simple, parameter-free, and consistent with the measured channel-mean growth in Figure 4. That is the genuinely new part, and it is solid. I also appreciate that the authors test the effect across two model families and several inversion schemes, including null-text inversion, which makes the phenomenon look robust rather than a quirk of one scheduler.\n\nThe proposed remedy, subtract per-channel means from each latent, interpolate the residuals with fFIX or fNIN, then add back blended means, is sensible and cheap. It resembles AdaIN, and the paper acknowledges that in the related work. The theoretical motivation makes the heuristic less ad hoc than in previous morphing papers, so credit is due there.\n\nWhere are the soft spots? The experimental evaluation would need error bars or at least multiple seeds to be fully convincing; the FID and CLIP curves in Figure 7 are presented without any spread. The closely related AdaIN baseline is mentioned but not directly compared, which leaves a bit of ambiguity about whether the specific decomposition matters or any instance normalization would do. And the stress-test concern stands: the d + e model assumes the bias is roughly constant per channel across spatial positions. If the bias comes from the terminal-SNR z0 component, it is image-specific and spatially structured, so channel-mean subtraction leaves a residual that still gets amplified by the Eq. (9) mechanism. The paper itself lists low-pass-filtered d as future work, so the authors are aware of the limitation, but the experiments do not measure how much spatial structure survives in the residual after chm. That is a genuine gap, though not a fatal one: the qualitative results show a large improvement, and the analysis correctly identifies the dominant bias.\n\nI also see no circularity problem. The derivation starts from an explicit additive-bias model without fitting constants, and the single self-cited baseline, fNIN, is only used for comparison, not to support the main claim.\n\nWho is this for? Anyone working on diffusion-based morphing, latent-space data augmentation, or centroid-based mixing with more than a handful of inputs. It is a subfield contribution with clear practical value, not a paradigm shift.\n\nI would send it to peer review. The theoretical core deserves a serious referee, and the revisions needed are straightforward: add error bars, include an AdaIN-style baseline, and discuss or measure the residual spatial bias.","headline":"A clean diagnosis of a real failure mode in many-input latent interpolation, with a simple fix that mostly works; the main soft spots are in the evaluation, not the theory.","tokens_in":11285,"tokens_out":1718,"would_cite":true,"duration_ms":17173,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A small bias in diffusion latents is amplified by sqrt(N) and makes centroid interpolation degenerate; the paper proposes a channel-wise mean adjustment that fixes it.","keywords":["diffusion models","latent interpolation","DDIM inversion","bias amplification","normalization","FID","CLIP distance","image morphing"],"falsifier":"Take latents with exactly zero channel-wise bias by drawing pure independent Gaussian noise, then compute normalized centroids for $N=2,8,\\dots,96$ and generate images from them: if those images still degenerate, Eq. (9) is not the cause. In the opposite direction, add a known constant $d$ to each channel before normalization and check whether the image degeneration scales with $d\\sqrt{N}$.","tokens_in":10309,"feed_emoji":"🎨","tokens_out":11003,"duration_ms":92011,"temperature":0.7,"pith_summary":"Diffusion models are increasingly used to merge several images by interpolating between their inverted latent codes, but the paper argues that these operations quietly break down as the number of inputs grows. The cause is not simply the low norm of an averaged latent, the usual explanation, but a small per-channel bias that is common to the latents and gets amplified by roughly $\\sqrt{N}$ when the centroid is normalized. The paper derives this amplification in a simplified noise-plus-bias model and verifies it experimentally on two Stable Diffusion versions. It then proposes a simple remedy: subtract each channel's mean from every latent, interpolate the residuals using a norm-adjusted blending method, and add the blended means back. Measured by FID and CLIP embedding distance, the fix reduces degeneration and improves quality even where the baseline images still look acceptable.","feed_headline":"Latent bias, not low norm, breaks diffusion interpolation","feed_subtitle":"Channel-wise mean adjustment removes the amplified bias and lifts FID and CLIP scores even for small N.","key_machinery":"The load-bearing object is the amplification identity $z'\\approx\\sqrt{N}\\,d + e'$ for the normalized centroid of $N$ latents, derived in Section 4.2. It works through the sharply concentrated norm of an average of independent Gaussian variables, $\\lVert \\frac{1}{N}\\sum_n e_n\\rVert\\approx\\sqrt{L/N}$, which makes the normalization factor $\\alpha$ grow like $\\sqrt{N}$. That identity identifies channel-wise mean bias, not low norm alone, as the cause of degeneration, and it dictates the repair: remove the estimated per-channel means before norm-adjusting the residual, then reinsert the blended means.","core_discovery":"Concretely, the paper claims that if inverted latents obey $z_n = d + e_n$ with $d$ a small common deterministic bias and $e_n$ independent Gaussian noise, then the normalized centroid satisfies $z' \\approx \\sqrt{N}\\,d + e'$. The normalization factor is boosted because the average of the noise terms has norm of order $\\sqrt{L/N}$, so scaling the centroid back to the nominal norm $\\sqrt{L}$ multiplies everything by about $\\sqrt{N}$. The deterministic trace left in latents by imperfect schedulers and denoisers is therefore no longer negligible: it dominates the interpolated latent as $N$ grows, producing washed-out and degenerate images even though the norm is correct. The proposed fix decomposes each latent into a per-channel mean part $d$ and a residual $e$, interpolates the residuals with the norm-preserving schemes $f_{FIX}$ or $f_{NIN}$, and adds back linearly interpolated means; this preserves the input-reproduction property when $f_{NIN}$ is used.","pith_inferences":["If the same bias-plus-noise model describes any normalized averaging of features whose mean differs from zero, the $\\sqrt{N}$ amplification may appear in other generative latent spaces, not only diffusion models; the remedy of per-dimension mean subtraction would then generalize to GAN inversion or embedding-space centroiding.","A low-pass-filtered estimate of $d$, which the paper lists as future work, should handle biases that vary across spatial positions; if spatial variation is strong, the channel-wise fix alone would still show residual quality loss at large $N$.","The paper's observation that $f_{FIX}$ slightly beats $f_{NIN}$ suggests the correct normalization target for far-from-input centroids is closer to the nominal $\\sqrt{L}$ than to interpolated input norms; a closed-form norm schedule interpolating between the two regimes could combine both advantages without optimization."],"forward_implications":["Centroid-based augmentation and multi-image morphing pipelines should expect normalized averages of many inverted latents to drift away from valid image latents as $N$ grows, even when every individual latent looks normal.","Quality metrics such as FID and CLIP distance degrade with $N$ before degeneration is visually obvious, so latent interpolation methods should be evaluated at several $N$ values rather than by inspecting a few images.","Using channel-wise mean adjustment is claimed to improve FID and CLIP distance even at small $N$, so it can serve as a drop-in replacement for plain fixed normalization or normalization to interpolated norms.","The input-reproduction property survives the fix when residual interpolation uses $f_{NIN}$, because a one-hot weight vector returns the original latent exactly."],"supporting_citations":[{"why":"baseline that identified the low-norm failure and introduced norm-guided interpolation, which the paper extends and corrects.","marker":"[24]"},{"why":"documented non-zero terminal SNR in common schedulers, used as one identified source of the deterministic latent bias.","marker":"[13]"},{"why":"provides the DDIM inversion procedure that produces the latents whose centroids are studied.","marker":"[27]"},{"why":"the Stable Diffusion latent model family on which the main FID and CLIP experiments are run.","marker":"[21]"},{"why":"defines the diffusion-model framework that the interpolation operates on.","marker":"[9]"},{"why":"Stable Diffusion 3.5, the second model family used to show the degeneracy persists after the terminal-SNR flaw is fixed.","marker":"[5]"},{"why":"the rectified flow noise model behind Stable Diffusion 3.5, showing the effect is not tied to the original noise schedule.","marker":"[14]"},{"why":"null-text inversion, an improved inversion procedure tested to show the degeneracy is not specific to DDIM inversion.","marker":"[17]"}],"fun_headline_variants":["Latent bias, not norm, causes diffusion interpolation degeneration","Remove latent bias to fix diffusion image morphing","Normalize channel means to stop diffusion interpolation collapse","Diffusion interpolation degrades from latent bias, fix is simple"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the amplified term is a common per-channel bias shared across latents and constant across spatial positions; if the bias varies from image to image or across space, the channel-wise mean fix stops removing the part that normalization amplifies.","fun_headline_variants_meta":{"raw":{"variants":["Latent bias, not norm, causes diffusion interpolation degeneration","Remove latent bias to fix diffusion image morphing","Normalize channel means to stop diffusion interpolation collapse","Diffusion interpolation degrades from latent bias, fix is simple"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000588,"raw_usage":{"total_tokens":2735,"prompt_tokens":897,"completion_tokens":1838,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1774}},"tokens_in":513,"tokens_out":1838,"duration_ms":13093,"temperature":1.0,"reasoning_tokens":1774,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:16:24.991125+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take latents with exactly zero channel-wise bias by drawing pure independent Gaussian noise, then compute normalized centroids for $N=2,8,\\dots,96$ and generate images from them: if those images still degenerate, Eq. (9) is not the cause. In the opposite direction, add a known constant $d$ to each channel before normalization and check whether the image degeneration scales with $d\\sqrt{N}$.","supporting_citations":[{"cited_title":"Advances in Neural Information Processing Systems 36 (2024)","cited_arxiv_id":null,"evidence_quote":"baseline that identified the low-norm failure and introduced norm-guided interpolation, which the paper extends and corrects."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"null-text inversion, an improved inversion procedure tested to show the degeneracy is not specific to DDIM inversion."}],"review_version":1}