{"id":"d8964a04-6410-4585-bd9b-09b02bf92afe","arxiv_id":"2412.17162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A handbook that unifies notation and practical guidance for diffusion, score-based, consistency, and rectified-flow models, including distillation and reward fine-tuning, but with some mathematical errors.","lead":"This paper is a practical handbook that standardizes notation and implementation guidance across major diffusion model families, from DDPM to rectified flow and TrigFlow. Practitioners may read it to implement and compare methods, though several equations contain errors that require cross-checking against original papers.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 39's velocity mapping only holds under a trigonometric time schedule, so the diffusion-to-rectified-flow unification in §2.4.5 is not valid for the general VP schedules used elsewhere in the handbook.","rationale":"The reader's weakest assumption identifies the most load-bearing issue. Section 2.6.3 is not an isolated aside: §2.4.5 builds the diffusion-to-rectified-flow velocity map on it, and §2.4.4 Approach 2 uses that map to define a distillation loss. If Eq. (39)'s derivative is schedule-specific, the handbook's advertised unified perspective is incomplete for the most common DDPM schedules. The DDPO variance error in §3.3.2 is also real, but it is localized to one reward-fine-tuning subroutine; the velocity mapping affects the core bridging claim. No other issue overturns the practical value of the handbook's summaries and derivations, so the conditional verdict remains appropriate rather than escalating to rejection.","tokens_in":33823,"tokens_out":8233,"duration_ms":72286,"concrete_test":"Independently differentiate b(t) in Eq. (39) for a standard VP schedule, e.g. ᾱ_t=exp(−∫_0^t β(s)ds) with β(s)=0.1+19.9s. At t=0.5, compute b'(t) by the handbook's formula √(1−ᾱ_t)−√ᾱ_t and by the true formula ᾱ_t'(√(1−ᾱ_t)−√ᾱ_t)/(2√(ᾱ_t(1−ᾱ_t))). If the two values differ (they do: approximately 0.68 versus −1.0), then Eq. (39) is schedule-dependent and the §2.4.5 velocity-mapping loss must be restricted to the TrigFlow angular-time case or re-derived for arbitrary VP schedules.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.6.3 differentiates b(t)=√ᾱ_t+√(1−ᾱ_t) and states b'=√(1−ᾱ_t)−√ᾱ_t. The true derivative is b'=ᾱ_t'(√(1−ᾱ_t)−√ᾱ_t)/(2√(ᾱ_t(1−ᾱ_t))). The displayed equality holds only when ᾱ_t'=2√(ᾱ_t(1−ᾱ_t)), i.e. when ᾱ_t=sin²φ_t with φ_t'=1 (the TrigFlow angular-time parameterization). For a standard VP schedule ᾱ_t=exp(−∫β), ᾱ_t'=−β_t ᾱ_t, so the printed b' is not correct in general. Because this derivative is the basis of Eq. (39) and of the velocity-mapping loss in §2.4.5, the claimed unification between diffusion-model velocities and rectified-flow velocities is not established for the DDPM schedules the handbook otherwise adopts. The downstream consequence is that §2.4.4 Approach 2, which enforces a rectified-flow loss on a mapped velocity, uses a mapping that does not follow from the stated definitions for general schedules.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a practical handbook on diffusion models that aims to unify notation and presentation across diffusion probabilistic models, score-based generative models, consistency models, rectified flow, flow matching, and TrigFlow. It reviews the fundamentals of forward and reverse processes, training objectives, inference algorithms, prediction parameterizations, and post-training techniques including distillation and reward-based fine-tuning. The handbook provides explicit algorithms and pseudocode throughout and attempts to connect the formulations of different methods to facilitate implementation and comparison.","tokens_in":34100,"tokens_out":11664,"duration_ms":85362,"significance":"If the technical claims are corrected, this handbook would be a valuable practical reference for practitioners, offering a single notation system, explicit algorithms, and coverage of recent methods such as LCM, DMD, and DDPO. Its strengths include a clear notation table, detailed derivations of standard DDPM/DDIM/score-matching results, and an extensive set of pseudocode algorithms. However, the claimed unified perspective relies on a velocity mapping between diffusion and rectified-flow trajectories that is established only under a trigonometric time schedule; as written, the proof in §2.6.3 is not valid for the general variance-preserving schedules otherwise used in the handbook. This undermines the generality of the unification as presented and requires either a corrected derivation or a restricted claim.","major_comments":[{"comment":"The proof of the velocity mapping between rectified-flow and diffusion trajectories assumes without stating a schedule-specific derivative. Differentiating b = sqrt(alpha_bar_t) + sqrt(1 - alpha_bar_t) gives b' = alpha_bar_t' (sqrt(1-alpha_bar_t) - sqrt(alpha_bar_t)) / (2 sqrt(alpha_bar_t(1-alpha_bar_t))), which equals sqrt(1-alpha_bar_t) - sqrt(alpha_bar_t) only when alpha_bar_t' = 2 sqrt(alpha_bar_t(1-alpha_bar_t)), i.e., for a trigonometric time parameterization such as TrigFlow. For standard VP schedules with alpha_bar_t = exp(-integral beta), the displayed derivative is incorrect. Since Eq. (39) and the equalities in §2.4.5 rely on this derivative, the claimed general mapping between diffusion velocities and rectified-flow velocities is not established for the VP schedules otherwise used in the handbook. The same issue also affects the definition of vx_t as dot x_t in the proof. Please either restrict the claim to TrigFlow or provide the corrected general mapping; this is load-bearing for Approach 2 in §2.4.4 and for the claimed unification.","section":"§2.6.3, Eq. (39); used in §2.4.5 and §2.4.4"},{"comment":"The DDPO variance formula in Eq. (50) is incorrect. The expression sqrt((1-alpha_bar_{t-1})/(1-alpha_bar_t)) * sqrt((1-alpha_bar_t)/alpha_bar_{t-1}) simplifies to sqrt((1-alpha_bar_{t-1})/alpha_bar_{t-1}), which is not the posterior standard deviation derived in Eq. (9) and used in Eqs. (12) and (49). For example, with alpha_bar_t=0.5 and alpha_bar_{t-1}=0.8, Eq. (50) gives sigma_t=0.5, while the posterior std from Eq. (9) is sqrt(0.15) approx 0.387. Since Eq. (51) uses sigma_t to compute log p_theta(x_{t-1}|x_t), the DDPO policy gradient estimate is affected. Please correct Eq. (50) to sqrt((1-alpha_bar_{t-1})/(1-alpha_bar_t)) * sqrt(1 - alpha_bar_t/alpha_bar_{t-1}) or the equivalent standard form.","section":"§3.3.1, Eq. (50)"}],"minor_comments":[{"comment":"The section title 'Foundamentals' should read 'Fundamentals'.","section":"§2.2.1, heading"},{"comment":"The sentence 'the formulations in many papers often differ from their their corresponding code implementations' contains a duplicated 'their'.","section":"Introduction"},{"comment":"The phrase 'due do lack of mean and standard deviation' should be 'due to the lack of mean and standard deviation'.","section":"§2.2.3 and §2.4.3"},{"comment":"In the discrete timestep inference formula, the notation 'with xtN ∼ N(0,I)' is inconsistent with the section's use of yt for the rectified-flow variable; this should likely refer to ytN or be clarified.","section":"§2.4.2"},{"comment":"The word 'upperscript' appears to be a typo for 'superscript'.","section":"§2.4.3"},{"comment":"The word 'acheiving' should be 'achieving'.","section":"§2.3.1"}],"recommendation":"major_revision","confidential_remarks":"This is a broad handbook-style manuscript rather than a novel research contribution. The two technical errors identified are correctable within the scope of the paper: the velocity-mapping issue requires either restricting the claim to TrigFlow or deriving the correct general mapping, and the DDPO variance formula appears to be a typo that can be fixed. The more consequential issue is the velocity mapping because it underpins the claimed unification. No concerns about citation practices or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a useful practical handbook, not a research contribution. The main value is the notation unification and the side-by-side derivations across DDPM, score matching, consistency models, rectified flow, and TrigFlow. Section 2.7 (the general velocity-score identity) is clean and correct, and the \"triangular\" parameterization diagrams are genuinely helpful for practitioners. The paper is honest about its scope: it disclaims comprehensiveness and theoretical depth, and it actually delivers a readable guide with code-aligned notation.\n\nNow the soft spots, in proportion. Two mathematical errors are real and load-bearing. In §2.6.3, Eq. (39) differentiates b = sqrt(ᾱ_t) + sqrt(1−ᾱ_t) and treats b′ as sqrt(1−ᾱ_t) − sqrt(ᾱ_t). That is only true when ᾱ_t follows a trigonometric time parameterization like TrigFlow (φ_t′ = 1). For the general VP schedules the handbook otherwise adopts, b′ has an extra factor involving ᾱ_t′, so the velocity mapping to rectified flow does not follow. Since §2.4.5's Approach 2 and the paper's \"unified perspective\" claim depend on this mapping, this is not a minor typo. Second, in §3.3.2, Eq. (50) for the DDPM posterior noise σ_t is wrong: it simplifies to sqrt((1−ᾱ_{t−1})/ᾱ_{t−1}), while the actual posterior variance has an additional (1−α_t) factor in the numerator. That is exactly the kind of formula a practitioner would copy into code, so it matters for the paper's stated goal.\n\nBeyond those, there are typos (e.g., \"due do lack\" in Algorithms 2 and 7) and some loose derivations elsewhere, but I do not see broader circularity or systematic unsoundness. The standard DDPM, DDIM, and score-matching equations are correctly presented, and the notation-clearing is real. The citation pattern is fine, and the acknowledgment to blog sources is appropriate.\n\nWho should read it? Practitioners and graduate students who want one notation to map between these methods, and anyone planning code implementations. It is not for someone looking for new methods or theory. I would send it back for revision, not desk-reject it. The two errors are fixable, and the rest of the handbook is solid enough to merit referee time. If I were refereeing, I would ask the authors to fix Eq. (39) with the correct schedule-specific derivative, or explicitly restrict the velocity-mapping claim to TrigFlow-type schedules, and to correct Eq. (50). After that, I would be happy to see it used widely.","headline":"A genuinely useful notation-unifying handbook, but two real equation errors—the velocity mapping and the DDPO variance—mean the unified perspective overshoots for general schedules.","tokens_in":34515,"tokens_out":4033,"would_cite":false,"duration_ms":32659,"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":"This handbook claims a shared notation and explicit velocity mappings can close the paper-to-code gap across diffusion, consistency, and rectified-flow models.","keywords":["diffusion probabilistic models","score-based generative models","consistency models","rectified flow","flow matching","TrigFlow","velocity mapping","paper-to-code gap"],"falsifier":"Evaluate both sides of the claimed velocity mapping for a standard variance-preserving schedule (for example, a linear noise schedule over 1000 steps) at several timesteps on the same pair $(x_0,\\varepsilon)$: if the mapped diffusion velocity $v^y_t$ does not equal $(x_0-\\varepsilon)/(1+2\\sqrt{\\bar{\\alpha}_t}\\sqrt{1-\\bar{\\alpha}_t})$ for those timesteps, then the mapping's general validity is refuted.","tokens_in":33637,"feed_emoji":"🧩","tokens_out":9235,"duration_ms":75533,"temperature":0.7,"pith_summary":"The paper is a practical handbook whose central claim is that the many flavors of diffusion-based generation—denoising diffusion probabilistic models, score-based models, consistency models, rectified flow, flow matching, and TrigFlow—can be described in one shared notation that is deliberately aligned with actual code implementations. It argues that the notorious \"paper-to-code\" gap comes mostly from inconsistent notation and missing implementation details rather than from differences in the underlying mathematics, and that a unified presentation closes that gap. The handbook derives common training objectives, inference procedures, prediction parameterizations, distillation methods, and reward-based fine-tuning in that notation, and it supplies explicit transformations between $\\epsilon$-, $x$-, $v$-, and $f$-predictions as well as a velocity mapping between diffusion trajectories and rectified-flow trajectories. If the unified perspective is correct, a practitioner who follows the handbook can implement any of these methods, convert one into another, and compare them fairly without re-deriving formulas from each original paper.","feed_headline":"One notation unifies diffusion, flow, and consistency models","feed_subtitle":"A velocity mapping and shared prediction formulas let practitioners switch methods without reimplementing each paper.","key_machinery":"The carrying object is the linear interpolant $x_t = \\alpha_t x_0 + \\beta_t x_1$ with differentiable coefficients, together with the derived velocity–score identity $$v(x_t,t) = \\frac{\\dot{\\$\\alpha$}_t}{\\alpha_t} x_t + \\beta_t\\!\\left(\\frac{\\dot{\\$\\alpha$}_t \\beta_t}{\\alpha_t} - \\dot{\\$\\beta$}_t\\right) s(x_t,t),$$ where $s$ is the score function and dots are time derivatives. This identity supplies the \"unified formulation\" that the handbook places at the center of the family. The other load-bearing piece is the velocity mapping between the diffusion variable $x_t = \\sqrt{\\bar{\\alpha}_t}x_0 + \\sqrt{1-\\bar{\\alpha}_t}\\varepsilon$ and the rectified-flow variable $y_t = x_t / (\\sqrt{\\bar{\\alpha}_t} + \\sqrt{1-\\bar{\\alpha}_t})$, written as $$v^y_t = \\frac{v^x_t b - x_t(\\sqrt{1-\\bar{\\$\\alpha$}_t} - \\sqrt{\\bar{\\$\\alpha$}_t})}{$b^{2}$},\\quad b = \\sqrt{\\bar{\\$\\alpha$}_t} + \\sqrt{1-\\bar{\\$\\alpha$}_t},$$ which converts a diffusion trajectory into a rectified-flow trajectory and back. Together these identities let every prediction head and every training or inference scheme be expressed in the same variables, and they are what make the claimed paper-to-code bridge concrete.","core_discovery":"The central claim is that a single formulation $x_t = \\alpha_t x_0 + \\beta_t x_1$, with time-dependent coefficients $\\alpha_t, \\beta_t$ and a velocity–score relationship, covers the diffusion-family models treated in the handbook, and that the apparent differences among them are notation and scheduling choices rather than fundamental divisions. Within this formulation, the three common prediction heads ($\\epsilon$-prediction, $x$-prediction, $v$-prediction) are linked by \"triangular\" identities, and a consistency model's $f$-prediction can be reparameterized through a diffusion model's noise prediction. A further derived velocity mapping connects the velocity along a diffusion trajectory to the velocity along a rectified-flow trajectory, which the handbook uses to justify applying rectified-flow training objectives to pre-trained diffusion models and to enable straight-line few-step sampling. The handbook also presents TrigFlow as the formulation that unifies continuous-time consistency models with rectified flow through an angular parameterization. The paper's discovery, on its own terms, is that these methods form one family whose members are interconvertible in code if the notation is fixed first.","pith_inferences":["A plausible extension the handbook leaves implicit is that adopting this notation as a shared convention would let downstream benchmarks report the exact schedule, prediction head, and sampler configuration as part of the result, making reproduction failures traceable to parameter choices instead of notation.","Because the velocity mapping's derivation uses a trigonometric schedule, a natural test is whether the same mapping holds, with a correction term, for variance-preserving schedules; if it does not, the unification would need to be reformulated as schedule-dependent rather than universal.","If the handbook's code-aligned notation is implemented as a single reusable code path, the time cost of switching between DDPM-style sampling, consistency-model sampling, and rectified-flow sampling should drop to changing a few schedule and prediction parameters; this is directly measurable and would confirm or undermine the paper-to-code claim."],"forward_implications":["A practitioner can convert a model trained with $\\epsilon$-prediction into $x$- or $v$-prediction using the triangular identities, without retraining the network.","The velocity mapping lets a pre-trained diffusion model be fine-tuned with a rectified-flow loss and then sampled along a straighter trajectory, reducing the number of sampling steps.","A consistency model can be initialized from a pre-trained diffusion model through the $f$-prediction reparameterization, so distillation starts from the teacher's weights rather than from scratch.","TrigFlow's angular parameterization gives one formulation in which the same model can be trained by the diffusion objective or by continuous-time consistency training.","Post-training techniques—progressive distillation, score distillation, distribution matching, adversarial distillation, consistency-based distillation, and reward-based fine-tuning—are all expressible in the same notation, which makes fair comparisons of their costs and gains possible."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model formulation, training loss, and discrete-time inference that the handbook standardizes.","marker":"[Ho et al., 2020]"},{"why":"Supplies the score-based SDE/ODE perspective and reverse-time sampling that the handbook uses as the common stochastic-process backbone.","marker":"[Song et al., 2021]"},{"why":"Supplies the consistency model, its self-consistency property, and the consistency distillation and training objectives.","marker":"[Song et al., 2023]"},{"why":"Supplies rectified flow, its linear interpolation, and the reflow/distillation procedure that the handbook integrates.","marker":"[Liu et al., 2022]"},{"why":"Supplies flow matching, which the handbook treats as a synonym or parallel formulation of rectified flow.","marker":"[Lipman et al., 2022]"},{"why":"Supplies TrigFlow's trigonometric interpolant and its unification of continuous-time consistency models with rectified flow.","marker":"[Lu and Song, 2024]"},{"why":"Supplies the LCM reparameterization that lets a consistency model's $f$-prediction be expressed through a diffusion model's noise prediction.","marker":"[Luo et al., 2023]"},{"why":"Supplies InstaFlow prediction, the rectified-flow velocity prediction that the handbook maps to diffusion velocity.","marker":"[Liu et al., 2023b]"},{"why":"Supplies the EDM parameterization and boundary-condition coefficients used by consistency models and TrigFlow.","marker":"[Karras et al., 2022]"}],"fun_headline_variants":["Unified notation makes diffusion, flow, and consistency interchangeable","One velocity mapping connects diffusion and rectified flow","Handbook bridges paper-to-code gap for diffusion models","Triangular identities link all diffusion prediction heads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The unified picture in Section 2.6.3 rests on the assumption that the velocity mapping between rectified-flow and diffusion trajectories holds for all the noise schedules the handbook uses, even though the derivation relies on a special trigonometric schedule; if the mapping fails for ordinary variance-preserving schedules, the unification is incomplete.","fun_headline_variants_meta":{"raw":{"variants":["Unified notation makes diffusion, flow, and consistency interchangeable","One velocity mapping connects diffusion and rectified flow","Handbook bridges paper-to-code gap for diffusion models","Triangular identities link all diffusion prediction heads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1247,"prompt_tokens":859,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":328}},"tokens_in":475,"tokens_out":388,"duration_ms":4065,"temperature":1.0,"reasoning_tokens":328,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:44:27.973826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate both sides of the claimed velocity mapping for a standard variance-preserving schedule (for example, a linear noise schedule over 1000 steps) at several timesteps on the same pair $(x_0,\\varepsilon)$: if the mapped diffusion velocity $v^y_t$ does not equal $(x_0-\\varepsilon)/(1+2\\sqrt{\\bar{\\alpha}_t}\\sqrt{1-\\bar{\\alpha}_t})$ for those timesteps, then the mapping's general validity is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the EDM parameterization and boundary-condition coefficients used by consistency models and TrigFlow."}],"review_version":1}