{"id":"cc0b185d-a885-4744-88ce-5dac123bf30d","arxiv_id":"2505.11936","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A replay-based diffusion training scheme with three consistency losses reduces generative forgetting on several 32x32 image benchmarks, though the supporting upper bound is not rigorously established.","lead":"This paper introduces CCD, a diffusion training scheme that adds three consistency losses to a rehearsal buffer to reduce generative forgetting across sequential tasks. It reports FID gains over several replay baselines on 32x32 image benchmarks and claims a theoretical upper bound that justifies the losses.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof relies on the false posterior-independence assumption pt(x0|xt,y)=pt(x0|y) at Eq. 32; without it the claimed bound and the three-loss decomposition are not established, so the theory remains conditional on empirical validation.","rationale":"I agree with the reader's weakest_assumption: Eq. 32 is the exact point where the proof's central identity breaks. A direct re-derivation or a simple Gaussian counterexample would settle it. The rest of the paper (CDG pipeline, empirical comparisons, ablations) can be evaluated independently of Theorem 1; the existing CONDITIONAL verdict already captures the needed repairs (fix or reframe the theory, add missing baselines/code/error bars). I do not see an additional load-bearing concern that would force REJECT, because the method is plausible and the empirical section gives some support. The main risk is overclaiming theoretical grounding, not the experimental protocol per se.","tokens_in":24987,"tokens_out":5190,"duration_ms":52566,"concrete_test":"Use a tractable 1D model where p(x0|y) and q(x0|y) are known (e.g., two Gaussians with different means). For the paper's ᾱt schedule, compute exactly the left side, ∇xt log E_{p(x0|xt,y)}[q(x0|y)/p(x0|y)], and the right side, -∇xt DKL(p(x0|xt,y)∥q(x0|xt,y)), at several t (e.g., 50, 200, 500, 800) and for y values from both tasks. If the two sides differ by more than a small relative tolerance, Eq. 32 fails. Then re-derive Eq. 45 without the posterior-independence assumption and check whether any nontrivial inequality of the form LUB ≤ κLIKC + λLUKC + ηLPKC holds with finite constants; if not, Theorem 1 should be reported as a heuristic objective, not a bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Appendix Eq. 32, where the proof replaces the posterior pt(x0|xt,y) with the label-conditional marginal pt(x0|y) (and likewise for q). This equality is false for diffusion processes: xt is defined from x0 by xt = ᾱt x0 + β̄t ε, so p(x0|xt,y) ∝ p(y|x0)p(x0) N(xt; ᾱt x0, β̄t²I), which depends on xt for every finite β̄t. The assumption is not a harmless simplification; under it ∇xt log E_{p(x0|xt,y)}[qt(x0|y)/pt(x0|y)] is exactly zero, so it cannot generate the non-trivial LUKC and LPKC terms. Eq. 32 is the bridge between the cross-task score difference and the three losses, so Theorem 1's claim that minimizing these losses bounds inter-task discrepancy is not derived. The existential constants {κ,λ,η} are also never specified or shown finite, making Eq. 1 tautological as written. The empirical results may still be useful, but the 'theoretically grounded' claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses generative catastrophic forgetting in continual learning of diffusion models. It first proposes a standardized pipeline, CDG, for evaluating continual diffusion generation, then introduces a method called CCD that combines three consistency losses: inter-task knowledge consistency (LIKC), unconditional knowledge consistency (LUKC), and prior knowledge consistency (LPKC). The central claim is Theorem 1, which states that there exist positive constants kappa, lambda, eta such that an inter-task discrepancy LUB is uniformly bounded by kappa LIKC + lambda LUKC + eta LPKC; the authors use this theorem as the theoretical foundation for CCD. Experiments on MNIST, OxfordPets, CIFAR-100, Flowers102, and CUB-200 report improvements in MF and IMF metrics over several continual learning baselines, with ablations and qualitative visualizations.","tokens_in":25347,"tokens_out":5493,"duration_ms":51883,"significance":"If Theorem 1 were valid, CCD would be the first principled, theoretically grounded method for continual diffusion generation, and the CDG pipeline could serve as a useful standardized benchmark. The empirical study is extensive: five datasets, multiple buffer sizes, several baselines, ablations, hyperparameter sensitivity, and qualitative results. The paper also provides detailed hyperparameter configurations and an explicit description of the benchmark pipeline, which are strengths. However, the theoretical contribution is the centerpiece of the submission, and it is not sound as written; the proof relies on a false posterior-independence assumption and several unquantified simplifications. The empirical results may be useful independently, but they do not compensate for the failure of the main theoretical claim.","major_comments":[{"comment":"The proof of Theorem 1 rests on the assumption p_t(x0|xt,y)=p_t(x0|y) and q_t(x0|xt,y)=q_t(x0|y). This is false for diffusion posteriors: xt = alpha_bar_t x0 + beta_bar_t epsilon makes x0 and xt dependent, and p(x0|xt,y) is proportional to p(y|x0)p(x0)N(xt; alpha_bar_t x0, beta_bar_t^2 I), which depends on xt for every finite beta_bar_t. Under this assumption the expectation in Eq. (31) becomes independent of xt, so its xt-gradient is exactly zero and cannot generate the nonzero LUKC and LPKC terms claimed in Eqs. (37) and (44). Eq. (32) is the bridge between the cross-task score difference and the three losses, so Theorem 1 is not established.","section":"Appendix, Eq. (32)"},{"comment":"As stated, Theorem 1 is tautological: LUB is defined as the right-hand side kappa LIKC + lambda LUKC + eta LPKC, so 'there exist positive constants' is true for any choice once the three losses are finite. No independent notion of inter-task discrepancy is bounded, no constants are specified, and no rate or uniformity in t is given. The theorem therefore provides no quantitative bound on generative forgetting, contrary to the abstract's claim of a theoretical foundation.","section":"Theorem 1 / Eq. (1)"},{"comment":"The approximation gradient of mu minus gradient of nu approximately delta is introduced with delta called a 'small perturbation determined based on the shared knowledge,' but delta is never defined and the approximation is not derived from the preceding assumptions. Since this step is what converts the KL gradient into the time-weighted mean difference in Eq. (37), the LUKC loss is not a justified consequence of the analysis.","section":"Appendix Eqs. (36)-(37)"},{"comment":"The term ||beta_t epsilon^p_theta + epsilon||^2 is 'non-optimizable' and is declared a learnable scaling factor, but it depends on t and on the frozen previous-task model and therefore cannot be absorbed into a constant factor without changing the objective. The proportionality in Eq. (44) silently drops this term, so the LPKC objective is not equivalent to the prior-knowledge KL gradient derived above.","section":"Appendix Eqs. (43)-(44)"},{"comment":"The 'Simplifiable Aspect' term gradient of log[q_t(y|xt)/p_t(y|xt)] is omitted on the informal ground that xt and y become independent under noise; no bound on this term is supplied. Because this term appears inside the claimed bound, dropping it without a quantitative estimate leaves Eq. (45) an inequality only under an additional unverified assumption.","section":"Appendix Eq. (34)"}],"minor_comments":[{"comment":"Eq. (2) in Theorem 1 defines LIKC as the vector difference epsilon^q_theta minus epsilon^p_theta, while the implemented loss in Eq. (7) is an expectation of a Bregman divergence; the theorem should use a norm or divergence to make LIKC a scalar loss.","section":"Theorem 1, Eq. (2)"},{"comment":"The paper alternates between CDG and CCD; the relationship between the 'Continual Diffusion Generation pipeline' and the 'Continual Consistency Diffusion' method should be stated once in a notation table.","section":"Introduction"},{"comment":"In Table 1 the 'Weighted Average' is said to be a normalized mean of MF and IMF using Non-CL scores, but the Non-CL row reports values 1.00 for both metrics while the other rows report values around 4-14; the normalization rule is not described precisely enough to reproduce.","section":"Table 1"},{"comment":"Figure 2 reports hyperparameter sensitivity for kappa, lambda, eta but does not state which metric is on the y-axis, whether the other two hyperparameters are held at their optimal values, or how many seeds were averaged.","section":"Ablation Studies, Figure 2"},{"comment":"There are several cross-reference errors: the visualization paragraph refers to 'Figure. 3' when the displayed images are in Figure 4, and the ablation subsection also refers to 'Figure. 3' for both the sensitivity and per-task ablation panels.","section":"Visualization and Ablation Studies"},{"comment":"The LPIPS results in Table 4 show essentially no improvement over ER, and the text attributes this to the UNet architecture; this should be discussed in the main paper rather than only in the appendix, since it tempers the claim of consistent generative-quality gains.","section":"Perceptual Metrics, Table 4"}],"recommendation":"reject","confidential_remarks":"The manuscript contains a large experimental study and a clearly described benchmark pipeline, which are valuable. However, the paper's central advertised contribution, the theoretical bound Theorem 1, is not mathematically grounded: the key equality in Eq. (32) is false, and the proof relies on several unquantified simplifications. In my view this is not a local fix; a resubmission without the theorem, or with a genuinely derived bound, would be a different paper. I recommend rejection of the present version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The useful part is the empirical package: the three-loss CCD stack (IKC/UKC/PKC) is new relative to the continual-diffusion literature, and the CDG pipeline with MF/IMF metrics gives the subfield a common protocol it has been missing. The ablations and visualizations are informative, and the authors are candid about where CCD fails (CIFAR100-10T, LPIPS, HDB on fine-grained). That honesty counts.\n\nThe problem is Theorem 1. The proof rests on Eq. 32 in the appendix, where the posterior p_t(x0|xt,y) is replaced by p_t(x0|y). That equality is false under the paper's own forward process: xt is a noisy version of x0, so the posterior depends on xt for every finite beta-bar_t. Moreover, under the assumed equality the gradient term is identically zero, so it cannot produce the LUKC and LPKC terms. The bound is then a tautological weighted sum with unspecified constants, an undefined delta at Eq. 36, and a dropped 'simplifiable aspect' plus a non-optimizable factor hand-waved into a learnable scaling at Eq. 44. This is not a harmless simplification; it is the bridge between the score difference and the three losses. Minimizing those losses may well help in practice, but the paper's 'theoretically grounded' claim is unsupported.\n\nEmpirically, the gains are real but modest and inconsistent. On CIFAR100-10T, CCD is basically at ER level; the strong wins are Flowers and CUB with large buffers. Weights kappa, lambda, eta are tuned per dataset in 1e-5 to 1e-7, which weakens any claim of a fixed principle. No code, no error bars, and the exclusion of DDGR/TD/AM/CUTER is argued rather than demonstrated, so 'SOTA' should be read as 'best among the baselines we chose.'\n\nWho is this for? Researchers actively working on continual diffusion will find the benchmark and the loss stack useful to try and compare against. Readers looking for a theoretical foundation should not cite Theorem 1 as it stands. I'd send it to peer review because the empirical contribution and benchmark merit referee time, but with a strong recommendation that the theory be reframed as a heuristic decomposition, the proof either corrected or removed, and code released. If that doesn't happen, the theorem is a liability, not a contribution.","headline":"Useful empirical package and benchmark, but Theorem 1's proof is invalid: the central theoretical claim should be reframed as a heuristic.","tokens_in":25794,"tokens_out":3630,"would_cite":false,"duration_ms":37934,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that generative catastrophic forgetting in diffusion models has a principled fix: bounding inter-task disagreement by three consistency losses, then minimizing them.","keywords":["continual diffusion generation","generative catastrophic forgetting","diffusion models","rehearsal buffer","score-based SDE","consistency regularization","lifelong generative modeling","knowledge retention"],"falsifier":"On a standard DDPM trained on MNIST, estimate $p_t(x_0|x_t,y)$ and $p_t(x_0|y)$ at intermediate noise levels such as $t\\approx 500$ and show that they differ materially; that difference invalidates the simplification in Equation 32 on which Theorem 1 rests.","tokens_in":24788,"feed_emoji":"🎨","tokens_out":6129,"duration_ms":60523,"temperature":0.7,"pith_summary":"This paper claims that generative catastrophic forgetting in diffusion models can be understood as a misalignment between task-specific reverse-time diffusion processes, and that it can be controlled by enforcing exactly three consistency constraints across tasks. The authors prove an upper bound stating that the discrepancy between two tasks' generative dynamics is bounded by a weighted sum of inter-task knowledge consistency, unconditional knowledge consistency, and prior knowledge consistency. They then build the Continual Consistency Diffusion (CCD) framework, which turns these three terms into training losses applied alongside a rehearsal buffer of real images. If the bound holds, continual diffusion training becomes a matter of minimizing these consistency objectives rather than borrowing heuristics from continual classification, and the result is a standardized way to evaluate and mitigate forgetting in lifelong generative modeling.","feed_headline":"Diffusion forgetting is bounded by three consistency losses","feed_subtitle":"The new CCD framework turns that bound into training losses and outperforms replay baselines on five benchmarks.","key_machinery":"The load-bearing object is the cross-task diffusion evolution bound of Theorem 1, derived by comparing two task-specific DDPM/SDE processes through a shared noisy state $x_t$. Its three terms are realized as losses: $L_{\\mathrm{IKC}}$ uses a Bregman divergence with a curvature (Fisher-information) preconditioner that keeps the student noise estimator close to the frozen teacher; $L_{\\mathrm{UKC}}$ aligns unconditional reverse-process means with a time-dependent weight $\\bar{\\alpha}_t^2/\\bar{\\beta}_t^2$ that emphasizes mid-diffusion phases; $L_{\\mathrm{PKC}}$ aligns label-conditioned priors on original samples through a task-adaptive label regressor. The paper also introduces the CDG pipeline, a standardized rehearsal-buffer protocol with two metrics, mean fidelity and incremental mean fidelity, that make continual diffusion evaluation systematic.","core_discovery":"The paper's central discovery is Theorem 1: for two tasks with diffusion processes equivalent when conditioned on clean samples and labels, there exist positive constants $\\kappa$, $\\lambda$, $\\eta$ such that the inter-task discrepancy is uniformly bounded by $$L_{\\mathrm{UB}} = \\kappa L_{\\mathrm{IKC}} + \\$\\lambda$ L_{\\mathrm{UKC}} + \\eta L_{\\mathrm{PKC}},$$ where $L_{\\mathrm{IKC}}$ is the difference between the tasks' noise estimators, $L_{\\mathrm{UKC}}$ penalizes differences between unconditional reverse-process means, and $L_{\\mathrm{PKC}}$ penalizes divergence between label-conditional priors. The paper argues that minimizing this bound aligns the reverse-time diffusion gradients between tasks, preserving shared generative knowledge in a continual sequence. This turns the empirical goal of mitigating generative forgetting into a concrete optimization objective and directly motivates the three hierarchical losses in the CCD framework.","pith_inferences":["A testable extension is to measure, on real diffusion trajectories, how strongly the simplification $p_t(x_0|x_t,y)=p_t(x_0|y)$ is violated at mid-diffusion noise levels; if it is violated exactly where the loss weights peak, the theorem's derivation would need a corrected posterior term.","The same three consistency axes could be ported to latent or text-to-image diffusion models by replacing the label regressor with a text-embedding similarity measure, so that prior consistency constrains semantic drift of the prompt representation rather than class labels.","The bound also suggests a principled buffer-selection rule: choose rehearsal samples that minimize the prior-consistency divergence term rather than relying on diversity heuristics, which could improve behavior on coarse-grained datasets where buffered-sample quality matters most."],"forward_implications":["In task sequences where tasks share visual structure, CCD should keep early-task FID close to the non-continual level, and its advantage should grow as the rehearsal buffer grows.","Parameter-regularization baselines and generative-replay approaches should underperform CCD because they do not align the score geometry captured by the three consistency terms.","CCD provides a theoretical justification for using direct sample replay rather than generative replay, since the consistency losses are defined on real buffered samples.","On tasks with minimal cross-task semantic overlap, the bound offers weak protection, so CCD's effectiveness should diminish on datasets like CIFAR100-10T, matching the paper's reported limitation."],"supporting_citations":[{"why":"Defines the DDPM forward noise schedule and training objective that the paper's base loss and reverse-mean formulas rely on.","marker":"(Ho, Jain, and Abbeel 2020)"},{"why":"Supplies the score-based SDE formulation whose reverse-time gradients the theorem is meant to align across tasks.","marker":"(Song et al. 2021)"},{"why":"Gives the reverse-time diffusion equation used to write the conditional score inside the SDE.","marker":"(Anderson 1982)"},{"why":"Establishes the score-matching equivalence between noise prediction and score estimation that connects the theoretical terms to trainable objectives.","marker":"(Vincent 2011)"},{"why":"Provides the gradient-alignment assumption used in Theorem 1 to treat task gradient directions as approximately aligned.","marker":"(Yu et al. 2020)"},{"why":"Motivates the teacher-student distillation strategy that $L_{\\mathrm{IKC}}$ instantiates in score space.","marker":"(Hinton 2015)"},{"why":"Defines experience replay, the buffer baseline and optimization backbone on which CCD is applied.","marker":"(Lopez-Paz and Ranzato 2017)"},{"why":"Supplies FID, the generative-quality metric behind the mean fidelity and incremental mean fidelity used in all comparisons.","marker":"(Heusel et al. 2017)"}],"fun_headline_variants":["Three losses cap diffusion forgetting","A triple bound for lifelong diffusion","Consistency rules curb generative drift","CCD: three losses, bounded memory","Diffusion forgetting: the theorem and the fix"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof requires that, given a clean image and its label, adding noise to that image makes the label provide no further information about the clean image: the posterior of the clean image given the noisy image and the label must equal the posterior given only the label.","fun_headline_variants_meta":{"raw":{"variants":["Three losses cap diffusion forgetting","A triple bound for lifelong diffusion","Consistency rules curb generative drift","CCD: three losses, bounded memory","Diffusion forgetting: the theorem and the fix"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000319,"raw_usage":{"total_tokens":1827,"prompt_tokens":997,"completion_tokens":830,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":770}},"tokens_in":613,"tokens_out":830,"duration_ms":8427,"temperature":1.0,"reasoning_tokens":770,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:45:11.959376+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a standard DDPM trained on MNIST, estimate $p_t(x_0|x_t,y)$ and $p_t(x_0|y)$ at intermediate noise levels such as $t\\approx 500$ and show that they differ materially; that difference invalidates the simplification in Equation 32 on which Theorem 1 rests.","supporting_citations":[],"review_version":1}