Pith. sign in

REVIEW 3 major objections 3 minor 89 references

For any affine diffusion path, an intermediate state plus its velocity already determines the clean endpoint, so generation can stop early and decode instead of integrating to the end.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 04:25 UTC pith:U2AOOUPI

load-bearing objection A sound but overstated formalization of an early-exit trick: the algebra is right, the 'curvature-independence' claim does not survive contact with learned model error. the 3 major comments →

arxiv 2607.06114 v3 pith:U2AOOUPI submitted 2026-07-07 cs.LG cs.AI

x-Prediction Is All You Need:Training-Free Accelerated Generation via Endpoint Decodability

Xin Peng , Ang Gao This is my paper
classification cs.LG cs.AI MSC 68T07
keywords endpoint decodabilitytruncated jump samplingdiffusion modelsflow matchingaffine probability pathstraining-free accelerationearly-exit samplingminimum mean square error estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that accelerating diffusion and flow-matching generation does not require retraining, distillation, or trajectory redesign. Its central move is to formalize endpoint decodability: for any affine probability path, an intermediate state and its velocity uniquely determine the clean sample through a closed-form decoder, and under the standard ℓ2 objective this decoder is the Bayes-optimal estimator of the clean sample. The authors then introduce Truncated Jump Sampling (TJS), which stops the ODE early and returns the decoded estimate, and report 20–70% reductions in neural function evaluations with near-matched quality across six model families. The error decomposition being independent of trajectory curvature is what makes early exit principled rather than heuristic.

Core claim

For any non-degenerate affine probability path xt = αt x0 + σt ε with nonzero path determinant Δt = α̇t σt − αt σ̇t, the intermediate state and path velocity recover the clean endpoint through the closed-form decoder x0 = (σt ut − σ̇t xt)/Δt. At Bayes optimality under the usual ℓ2 flow-matching objective, this algebraic decoder coincides with the minimum-MSE estimator E[x0 | xt]. Consequently, stopping the ODE at an early time t* and returning the decoded estimate—Truncated Jump Sampling—has error E[||et*||²] + U(t*), where U(t*) is the irreducible endpoint uncertainty, and neither term depends on trajectory curvature. The paper verifies this across six model families, reporting 20–70% NFE r

What carries the argument

The central object is endpoint decodability: a property of affine probability paths that holds exactly when the path determinant Δt = α̇t σt − αt σ̇t is nonzero. The carrying identity is the Cramer's-rule decoder x0 = (σt ut − σ̇t xt)/Δt, which at Bayes optimality equals E[x0 | xt]. The error analysis then splits TJS error into model error plus irreducible uncertainty U(t*) = E[Tr Var(x0 | xt*)], with no curvature term; this decomposition is what turns early exit into a principled inference strategy rather than an ad hoc heuristic.

Load-bearing premise

The practical claims rest on the premise that a real trained model's velocity error stays small enough at the early-exit time that the decoded endpoint remains semantically and aesthetically acceptable; the theory bounds this model error only abstractly, and the paper itself flags slow U(t) decay for complex data as a primary limitation.

What would settle it

Measure the squared error of the endpoint decoder against true held-out samples for a pretrained flow-matching model at every early-exit time t*, and compare it with the error of the full ODE; if the decoded error at the recommended t* is not close to the full-ODE error, the practical claim fails. A more targeted check is to run TJS at a fixed early-exit fraction on a model with a curved noise schedule and confirm that FID or human preference does not degrade more than the theory's curvature-independent bound allows.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any pretrained diffusion or flow-matching checkpoint with a standard affine schedule encodes an endpoint predictor, so TJS works without retraining, distillation, or architecture changes.
  • TJS reduces NFE by 20–70% with near-matched quality across six model families, including on an already-distilled model where it adds further savings.
  • All four standard prediction parameterizations—x0, velocity, noise, and score—are equivalent at optimality, so existing checkpoints can be used directly with the appropriate decoder.
  • Because TJS error is curvature-independent, straightening the trajectory is sufficient but not necessary for accurate endpoint prediction, directly challenging the premise of trajectory-straightening acceleration methods.
  • Quality improves strictly monotonically with integration depth, and the stopping point can be reasoned about through the decay of U(t), which is tied to the information gain rate with respect to signal-to-noise ratio.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If endpoint decodability holds as broadly as claimed, it provides a universal no-training acceleration layer that can be layered on top of any future affine-path generative model without touching weights.
  • The curvature-independence result suggests that trajectory straightening and endpoint decoding address different bottlenecks; distillation compresses the informative trajectory while TJS trims the redundant tail, so the two savings could multiply rather than compete.
  • The information-theoretic characterization of U(t) points toward an adaptive stopping rule: estimate the remaining endpoint uncertainty per sample or per prompt and exit when the marginal gain drops below a threshold, rather than using a fixed exit fraction.
  • The non-affine counterexample suggests that nonlinear paths, while not globally endpoint-decodable, may still admit local endpoint decoding via iterative inversion—a testable extension for interpolants beyond the affine family.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes Truncated Jump Sampling (TJS), an inference-only acceleration scheme for diffusion and flow-matching models. For affine probability paths x_t = α_t x_0 + σ_t ε, the paper observes that (x_t, u_t) uniquely determines x_0 whenever the path determinant Δ_t = α̇_t σ_t − α_t σ̇_t is nonzero, with the closed-form decoder x_0 = (σ_t u_t − σ̇_t x_t)/Δ_t (Theorem 5). Under the standard ℓ2 flow-matching objective, plugging the Bayes-optimal velocity into this decoder yields the MMSE estimator E[x_0 | x_t] (Theorem 6). TJS stops the ODE at an early time t* and returns this decoded estimate, avoiding the final integration segment. The paper derives an error decomposition MSE = E||e_t*||² + U(t*) (Theorem 9), claims that TJS error is curvature-independent, and presents experiments across SDXL, SD3.5M, Z-Image-Turbo, ImageNet-256, CIFAR-10, and MNIST reporting 20–70% NFE reductions with near-matched quality, along with ablations over samplers, schedules, step counts, and CFG scales.

Significance. The core algebraic observation is elegant and broadly applicable: every affine-path generative model implicitly contains an endpoint predictor, and TJS is a simple, training-free way to exploit it. The MMSE-optimality argument and the I-MMSE connection are clean, and the experimental breadth — six model families, multiple solvers, schedules, and CFG scales — is a genuine strength. If the practical claims survive scrutiny, TJS would be a valuable complement to distillation and solver-based acceleration. However, the advertised 'curvature-independent error' and 'straightness is not necessary' claims are established only for the U(t*) term of the decomposition, not for the model-error term E||e_t*||², which is the load-bearing gap for learned models. This overclaim is acknowledged in the supplementary material itself and needs to be addressed before the central challenge to Rectified Flow/Consistency Models can be sustained.

major comments (3)
  1. The decomposition MSE_TJS = E||e_t*||² + U(t*) is correct, but the main text's conclusion that 'neither error term depends on trajectory curvature' does not follow. U(t*) is a functional of the forward schedule and data distribution, but E||e_t*||² is the gap between the learned and Bayes-optimal velocity for a finite-capacity network; that gap generally depends on the complexity of the target regression function, which is controlled by α̇_t, σ̇_t and their higher derivatives. Supplementary §A.4 concedes exactly this ('learning a velocity field on a highly curved path may be harder'), which is in tension with the main-text claim. Proposition 12's construction only controls the algebraic conditioning Δ_t; it does not show that a practically trained model has small e_t on such high-curvature paths. All experiments use standard schedules with modest curvature. Because 'straightness is suffi
  2. The perturbed schedule σ_t^(ω) = 1 − t + ω^(−1) cos(ω t (1−t)) gives σ_1^(ω) = 1/ω, not the σ_1 = 0 required by Definition 1. The sentence 'boundary conditions are preserved up to O(1/ω)' does not make this an admissible affine probability path for finite ω. The construction should be repaired with an additional boundary-correcting term, or Proposition 12 should be restated for paths with approximate boundary conditions. As written, the construction does not prove the claim for paths satisfying Definition 1, and this proposition is the paper's key evidence for the 'curvature is unnecessary' thesis.
  3. The comparison is not well defined as stated. With h = 1/N and t* = N h, one gets t* = 1, which is not an early exit; if a different h is intended for the TJS leg, it is not specified. The proof sketch also invokes an Euler error lower bound with prefactor h²/2, but standard global error bounds for Euler are upper bounds; a lower bound requires additional assumptions not stated. The '+2ε' term in the final inequality is unexplained. Since Theorem 11 is advertised as the quantitative criterion for when TJS wins over continued integration, the NFE accounting and the derivation need to be corrected, or the theorem should be removed from the main text.
minor comments (3)
  1. The setup says 'CIFAR-10/MNIST: 30-step ODE,' but Table 1 lists CIFAR-10 full ODE as 40 steps and the text refers to a 'full 40-step ODE' for CIFAR-10. Please harmonize the step counts and the labels in Table 1 (e.g., the 'TJS-0.833/25/33' row mixes different γ values across datasets).
  2. The claim that U(t) is monotonically non-increasing by the data processing inequality requires the family {x_t} to be totally ordered by degradation, i.e., SNR(t) = α_t/σ_t to be non-decreasing. This is true for all standard schedules used in the experiments, but it is not true for an arbitrary C¹ affine path satisfying only Δ_t ≠ 0. Please state the monotone-SNR condition explicitly.
  3. There are several unresolved cross-references and broken section markers ('see §' in the Introduction and the 'Relationship to DDIM' section), and the text uses nonstandard placeholder-like glyphs in some figure captions. These should be cleaned up before publication.

Circularity Check

0 steps flagged

No significant circularity: the core derivation is a self-contained algebraic identity and no fitted parameter is relabeled as a prediction.

full rationale

The central derivation chain is not circular. Theorem 5 is Cramer's rule applied to the defining linear system (xt,ut) = M_t (x0,ε); it assumes only the paper's own Definition 1, and the decoder formula is an algebraic consequence. Theorem 6 follows by linearity of conditional expectation and the standard ℓ2 optimality of conditional means; no assumption equivalent to the conclusion is imported. Theorem 9 is an orthogonal decomposition (E||et*||² + U(t*)) with U defined as the MMSE; its content is definitional but it is not used to fit any parameter. The experiments compare early-exit quality against the full ODE without tuning constants of the theory; the monotonic improvement claimed is a direct consequence of U(t) being non-increasing and is confirmed, not used to infer the theory. The paper contains no self-citations that carry a uniqueness or ansatz argument; prior work (DDIM, ODE-Jump) is cited contextually and explicitly distinguished. The principal weakness is not circularity: §A.4 of the supplement concedes 'learning a velocity field on a highly curved path may be harder,' which undercuts the main text's stronger claim that TJS error is curvature-independent for learned models; that is a correctness/evidence gap, not a reduction of the prediction to its inputs. The TJS–Euler comparison (Theorem 11) uses an assumed bound on E||et||² and standard Euler error analysis; the bound is an assumption, not a fitted value. Therefore the paper's derivation does not reduce to its own assumptions.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The theory rests on a minimal set of standard affine-path assumptions. The main free parameters are the practical operating points: the early-exit fraction gamma and the quality-retention thresholds. The real unmeasured quantity is the model estimation error et, which is assumed small. The invented entities are re-namings of standard quantities (conditional mean, conditional covariance), not new physical postulates. The central derivation itself is parameter-free and does not fit constants.

free parameters (3)
  • early-exit fraction gamma (or k*) = per-model values: 0.6-0.83 for class-conditional, k*=18-24 for SDXL/SD3.5M, k*=2-4 for Z-Image-Turbo
    The early-exit time t* is the central knob of TJS. The paper evaluates many k* values and then reports the ones that achieve 90/95/99% quality retention. This is a hand-selected operating point per model/metric, used to quantify the headline NFE savings.
  • metric-specific quality retention thresholds (90/95/99%) = 90%, 95%, 99%
    The claim of 'near-matched quality' and the NFE-saving percentages depend on arbitrarily chosen retention thresholds. The same data can support '20-70% NFE saving' or '0% saving at 99% ImageReward retention for CIFAR-10' depending on the threshold and metric.
  • epsilon bound on model error E[||et||^2] <= eps (Theorem 11) = not quantified
    The TJS-vs-Euler comparison theorem requires a uniform bound on model estimation error that is never measured or estimated. The paper acknowledges this is a simplification but the theorem's quantitative utility depends on this unknown quantity.
axioms (5)
  • domain assumption The learned velocity field v_theta approximates the Bayes-optimal velocity E[ut|xt] well enough at early t* that E[||et*||^2] is small.
    Theorem 6 proves optimality only at the Bayes-optimal predictor. The practical success of TJS depends on real trained models being close to this optimum at intermediate times. The paper notes 'model estimation error' as a term but never bounds it empirically; Theorem 11 assumes E[||et||^2] <= eps uniformly without measuring eps.
  • domain assumption The MMSE endpoint estimate E[x0|xt] is a sufficient proxy for user-perceived image quality on the metrics used (FID, PickScore, ImageReward, etc.).
    The theory is ℓ2-optimality; the paper explicitly disclaims perceptual metrics in Theorem 6 discussion. Yet the experimental claims of 'near-matched quality' are made on FID and human-preference metrics, which are not necessarily aligned with MMSE. The gap between ℓ2-optimal endpoint estimates and perceptual quality is unexamined.
  • domain assumption Text-to-image CFG-modified velocity can be substituted into the affine-path decoder without invalidating the identity x0 = (sigma_t v - sigma_dot_t x)/Delta_t.
    The endpoint decoder is derived for the true conditional velocity. With classifier-free guidance the effective velocity is a linear combination of conditional/unconditional velocities. The paper claims CFG composes orthogonally and uses a 'CFG-aware endpoint estimate', but the exact formula is not derived in the main text and the supplementary is not fully included.
  • standard math The ODE solver, scheduler, and model families used (SDXL, SD3.5M, Z-Image-Turbo, DiT, U-Net) all operate on affine probability paths with Delta_t != 0.
    This is verified in the supplementary for VP, VE/EDM, and linear FM. It is a standard structural fact about these models. Not a weakness.
  • standard math The data-processing inequality monotonicity of U(t) applies for general p_data.
    The paper uses it to justify monotonic quality improvement. It is mathematically sound under the affine-path/Gaussian-noise model.
invented entities (2)
  • endpoint decodability no independent evidence
    purpose: A formal property of affine probability paths: (xt, ut) determines x0 when Delta_t != 0.
    This is a new name for a known algebraic fact. It has no falsifiable handle outside the paper beyond the experimental demonstrations. It is not a physical entity.
  • irreducible endpoint uncertainty U(t) no independent evidence
    purpose: A quantity measuring the minimum achievable MSE for predicting x0 from xt; used to decompose TJS error and predict monotonic improvement.
    This is the trace of the conditional covariance Var(x0|xt), a standard estimation-theoretic quantity. It is not independently falsifiable; it is defined in terms of the data distribution. The paper connects it to I-MMSE, which is a known identity.

pith-pipeline@v1.3.0-alltime-deepseek · 46829 in / 10431 out tokens · 75477 ms · 2026-08-04T04:25:16.793786+00:00 · methodology

0 comments
read the original abstract

Diffusion and flow matching models generate high-quality samples, but their ODE samplers often need tens to hundreds of neural function evaluations (NFEs). This remains a practical challenge for released checkpoints, since many accelerators require additional design choices and training cost through retraining, distillation, or trajectory redesign. We investigate a different route based on $x$-prediction. During sampling, standard affine probability paths already expose $x_0$ information: an intermediate state and its path velocity determine a principled estimate of the clean sample. We formalize this property as \textbf{endpoint decodability} and show that the decoder is the minimum-MSE estimator $\mathbb{E}[x_0\mid x_t]$ under the usual $\ell_2$ objective. This yields \textbf{Truncated Jump Sampling} (TJS): stop the ODE at an early-exit time $t^*$ and return the decoded $x_0$. TJS requires no retraining, distillation, or architecture change. Across SDXL, SD3.5M, Z-Image-Turbo, and three class-conditional benchmarks, it reduces NFEs by 20--70\% with near-matched quality. The analysis also shows why endpoint prediction can work without straightening the trajectory, providing inference acceleration without trajectory redesign.

Figures

Figures reproduced from arXiv: 2607.06114 by Ang Gao, Xin Peng.

Figure 1
Figure 1. Figure 1: Endpoint decodability in action. Top: xt decoded directly (noisy at early steps). Bottom: x0 via endpoint decoding (clean at any step). strictly beat coarse Euler (Theorem 11). All standard pa￾rameterizations are equivalent at optimality (see Supple￾mentary Material, §A). Critically, straight trajectories are sufficient but not necessary (Proposition 12), challenging the foundation of Rectified Flow and Co… view at source ↗
Figure 2
Figure 2. Figure 2: Visual x0 predictions for CIFAR-10 (left), MNIST (center), and ImageNet-256 (right). MNIST saturate at k ∗≈16 (43% NFE saving); CIFAR-10/ImageNet-256 at k ∗≈26 (33%). 0 20 40 NFE 0 25 50 75 100 125 150 175 200 FID ↓ TJS best: 12.09 | Full ODE: 13.30 CIFAR-10 0 10 20 30 NFE 0 25 50 75 100 125 150 175 200 FID ↓ TJS best: 1.65 | Full ODE: 3.20 MNIST 0 20 40 NFE 0 50 100 150 200 250 300 350 FID ↓ TJS best: 15.… view at source ↗
Figure 3
Figure 3. Figure 3: FID vs. NFE for TJS on MNIST (30-step) and CIFAR-10/ImageNet-256 (40-step, CFG=1.0). [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual x0 predictions for SDXL (left) and SD3.5M (right). Saturation at k ∗≈19 (∼33% NFE saving). Theorem 11 (TJS–Euler Comparison). Let the affine path have C 2 coefficients. Compare two strategies at the same NFE budget N + 1: (a) Coarse Euler from t=0 to t=1 with step h = 1/N; (b) TJS-N, integrating to t ∗ = Nh via Euler then applying endpoint decoding. Assume E[∥et∥ 2 ] ≤ ε uniformly, and define Cα,σ =… view at source ↗
Figure 5
Figure 5. Figure 5: Speed vs. quality trade-off. Left: ImageReward against NFE, with 90% of full ODE quality marked per model. Right: [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Four-panel detailed view of T2I quality metrics (PickScore, ImageReward, HPSv2, CLIP) for SDXL and SD3.5M [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comprehensive multi-benchmark TJS convergence analysis. Five metric panels, each with six curves (3 benchmarks [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sampler ablation (SDXL, k ∗=12, DrawBench). Five common ODE solvers (DDIM, DPM++, LMS, PNDM, UniPC) are evaluated with identical k ∗ and K. Bar chart displays PickScore, CLIP, and ImageReward for each solver. The near-identical bar heights across all five solvers confirm that endpoint decoding quality is sampler-agnostic: the endpoint predictor extracts the same information regardless of the specific integ… view at source ↗
Figure 9
Figure 9. Figure 9: Schedule ablation (SDXL, DrawBench). Four noise schedules (Beta, Exponential, Karras, Laplace) evaluated across [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Step-count ablation (SDXL, DrawBench). ImageReward evaluated at three fixed integration fractions ( [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: CFG scale ablation (SDXL, DrawBench). ImageReward evaluated at CFG scales [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: CFG scale ablation on ImageNet-256 (class-conditional generation). Full TJS FID sweep (40 steps) at three CFG [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visual x0 predictions for ImageNet-256 at CFG=1.25 (left) and CFG=1.5 (right). As k ∗ increases, image quality improves monotonically: global structure emerges first (k ∗=0–12), followed by texture detail (k ∗=18–26). The TJS-best (⋆) predictions at k ∗=26–32 are visually indistinguishable from or superior to the full 40-step ODE. Compare with [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comprehensive TJS convergence analysis for Z-Image-Turbo ( [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Visual x0 predictions for Z-Image-Turbo (K=10) at increasing k ∗ ∈ {0, 1, 2, 3, 4, 6, 8}, plus the full 10-step ODE as reference. Each row shows a different prompt. The visual progression confirms the quantitative findings: k ∗=0 (1 NFE) already produces semantically recognizable content; k ∗=2 (3 NFE) resolves fine details such as text rendering, facial features, and material textures; k ∗=3 (4 NFE) outp… view at source ↗
Figure 16
Figure 16. Figure 16: Direct visual comparison of xt (top row, the intermediate state at step k ∗ ) vs. x0 (bottom row, endpoint-decoded from the same xt) for Z-Image-Turbo (K=10). This figure provides the most direct illustration of Theorem 5: while xt remains corrupted by noise (top row, especially at early k ∗ ), the endpoint decoder recovers a clean, semantically coherent x0 (bottom row) from the exact same intermediate st… view at source ↗

Review history (3 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 12 linked inside Pith

  1. [1]

    Flow Matching for Generative Modeling , author=

  2. [2]

    Building Normalizing Flows with Stochastic Interpolants , author=

  3. [3]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=

  4. [4]

    Generative modeling by estimating gradients of the data distribution , author=

  5. [5]

    Score-Based Generative Modeling through Stochastic Differential Equations , author=

  6. [6]

    Denoising diffusion probabilistic models , author=

  7. [7]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015. 2015

    2015

  8. [8]

    Denoising Diffusion Implicit Models , author=

  9. [9]

    2021 , organization=

    Improved denoising diffusion probabilistic models , author=. 2021 , organization=

    2021

  10. [10]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-Free Diffusion Guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  11. [11]

    Diffusion models beat gans on image synthesis , author=

  12. [12]

    Progressive Distillation for Fast Sampling of Diffusion Models , author=

  13. [13]

    Masked Generative Distillation , author=

  14. [14]

    Consistency Models , author=

  15. [15]

    2023 , eprint=

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference , author=. 2023 , eprint=

    2023

  16. [16]

    Zero-shot text-to-image generation , author=

  17. [17]

    arXiv preprint arXiv:2204.06125 , year=

    Hierarchical text-conditional image generation with clip latents , author=. arXiv preprint arXiv:2204.06125 , year=

    Pith/arXiv arXiv

  18. [18]

    2022 , organization=

    GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models , author=. 2022 , organization=

    2022

  19. [19]

    Photorealistic text-to-image diffusion models with deep language understanding , author=

  20. [20]

    Transactions on Machine Learning Research , issn=

    Scaling Autoregressive Models for Content-Rich Text-to-Image Generation , author=. Transactions on Machine Learning Research , issn=. 2022 , url=

    2022

  21. [21]

    High-resolution image synthesis with latent diffusion models , author=

  22. [22]

    arXiv preprint arXiv:2307.01952 , year=

    Sdxl: Improving latent diffusion models for high-resolution image synthesis , author=. arXiv preprint arXiv:2307.01952 , year=

    Pith/arXiv arXiv

  23. [23]

    Scalable diffusion models with transformers , author=

  24. [24]

    Scaling rectified flow transformers for high-resolution image synthesis , author=

  25. [25]

    2026 , url=

    Improving Classifier-Free Guidance in Masked Diffusion: Low-Dim Theoretical Insights with High-Dim Impact , author=. 2026 , url=

    2026

  26. [26]

    Black-Forest-Labs. FLUX.1. 2024

    2024

  27. [27]

    arXiv preprint arXiv:2405.12230 , year=

    Qwen-Image Technical Report , author=. arXiv preprint arXiv:2405.12230 , year=

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2502.17332 , year=

    Wan: Open and Advanced Large-Scale Video Generative Models , author=. arXiv preprint arXiv:2502.17332 , year=

    Pith/arXiv arXiv

  29. [29]

    PriorGrad: Improving Conditional Denoising Diffusion Models with Data-Dependent Adaptive Prior , author=

  30. [30]

    2021 , url=

    Diffusion Schr\"odinger Bridge with Applications to Score-Based Generative Modeling , author=. 2021 , url=

    2021

  31. [31]

    2025 , eprint=

    CFG-Zero*: Improved Classifier-Free Guidance for Flow Matching Models , author=. 2025 , eprint=

    2025

  32. [32]

    2024 , eprint=

    CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models , author=. 2024 , eprint=

    2024

  33. [33]

    Rectified

    Shreshth Saini and Shashank Gupta and Alan Bovik , booktitle=NeurIPS, year=. Rectified

  34. [34]

    2026 , eprint=

    CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance , author=. 2026 , eprint=

    2026

  35. [35]

    2026 , eprint=

    C ^2 FG: Control Classifier-Free Guidance via Score Discrepancy Analysis , author=. 2026 , eprint=

    2026

  36. [36]

    2025 , eprint=

    Warm Starts Accelerate Conditional Diffusion , author=. 2025 , eprint=

    2025

  37. [37]

    Flowing from Words to Pixels: A Noise-Free Framework for Cross-Modality Evolution , author=

  38. [38]

    Chen Chen and Pengsheng Guo and Liangchen Song and Jiasen Lu and Rui Qian and Tsu-Jui Fu and Xinze Wang and Wei Liu and Yinfei Yang and Alex Schwing , booktitle=NeurIPS, year=

  39. [39]

    Neural ordinary differential equations , author=

  40. [40]

    What uncertainties do we need in bayesian deep learning for computer vision? , author=

  41. [41]

    and Bottou, L

    Lecun, Y. and Bottou, L. and Bengio, Y. and Haffner, P. , journal=. Gradient-based learning applied to document recognition , year=

  42. [42]

    Learning multiple layers of features from tiny images , year =

    Krizhevsky, Alex and Hinton, Geoffrey , address =. Learning multiple layers of features from tiny images , year =

  43. [43]

    ImageNet: A large-scale hierarchical image database , year=

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Kai Li and Li Fei-Fei , booktitle=. ImageNet: A large-scale hierarchical image database , year=

  44. [44]

    2017 , isbn =

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , title =. 2017 , isbn =

    2017

  45. [45]

    2016 , isbn =

    Salimans, Tim and Goodfellow, Ian and Zaremba, Wojciech and Cheung, Vicki and Radford, Alec and Chen, Xi , title =. 2016 , isbn =

    2016

  46. [46]

    2025 , eprint=

    SD3.5-Flash: Distribution-Guided Distillation of Generative Flows , author=. 2025 , eprint=

    2025

  47. [47]

    2016 , volume=

    Szegedy, Christian and Vanhoucke, Vincent and Ioffe, Sergey and Shlens, Jon and Wojna, Zbigniew , booktitle=CVPR, title=. 2016 , volume=

    2016

  48. [48]

    arXiv , primaryClass=

    Flow Matching in Latent Space , author=. arXiv , primaryClass=. 2023 , eprint=

    2023

  49. [49]

    2020 , eprint=

    An Ensemble of Simple Convolutional Neural Network Models for MNIST Digit Recognition , author=. 2020 , eprint=

    2020

  50. [50]

    Deep Residual Learning for Image Recognition , author=

  51. [51]

    generation: Taming optimization dilemma in latent diffusion models , author=

    Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models , author=

  52. [52]

    2026 IEEE International Conference on Robotics and Automation (ICRA) , year=

    Dense-Jump Flow Matching with Non-Uniform Time Scheduling for Robotic Policies: Mitigating Multi-Step Inference Degradation , author=. 2026 IEEE International Conference on Robotics and Automation (ICRA) , year=

    2026

  53. [53]

    Elucidating the Design Space of Diffusion-Based Generative Models , author=

  54. [54]

    InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation , author=

  55. [55]

    ICML 2023 Workshop on Structured Probabilistic Inference and Generative Modeling , year=

    BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping , author=. ICML 2023 Workshop on Structured Probabilistic Inference and Generative Modeling , year=

    2023

  56. [56]

    Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation , author=

  57. [57]

    arXiv preprint arXiv:2306.09341 , year=

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis , author=. arXiv preprint arXiv:2306.09341 , year=

    Pith/arXiv arXiv

  58. [58]

    ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation , author=

  59. [59]

    Journal of the American Statistical Association , volume =

    Bradley Efron , title =. Journal of the American Statistical Association , volume =. 2011 , publisher =

    2011

  60. [60]

    IEEE Transactions on Information Theory , volume=

    Mutual information and minimum mean-square error in Gaussian channels , author=. IEEE Transactions on Information Theory , volume=. 2005 , publisher=

    2005

  61. [61]

    Learning Transferable Visual Models From Natural Language Supervision , author=

  62. [62]

    DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps , author=

  63. [63]

    Machine Intelligence Research , year=

    DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models , author=. Machine Intelligence Research , year=

  64. [64]

    SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations , author=

  65. [65]

    Variational Diffusion Models , author=

  66. [66]

    Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models , author=

  67. [67]

    Fast Sampling of Diffusion Models via Operator Learning , author =

  68. [68]

    Pseudo Numerical Methods for Diffusion Models on Manifolds , author=

  69. [69]

    arXiv preprint arXiv:2402.19159 , year=

    Trajectory Consistency Distillation: Improved Latent Consistency Distillation by Semi-Linear Consistency Function with Trajectory Mapping , author=. arXiv preprint arXiv:2402.19159 , year=

    Pith/arXiv arXiv

  70. [70]

    Flow Matching for Conditional Generation with Optimal Transport , author=

  71. [71]

    Improved Techniques for Training GANs , author=

  72. [72]

    Flow Matching on Manifolds , author=

  73. [73]

    2024 , eprint=

    Improving and generalizing flow-based generative models with minibatch optimal transport , author=. 2024 , eprint=

    2024

  74. [74]

    PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores , author=

  75. [75]

    TADA: Improved Diffusion Sampling with Training-Free Augmented Dynamics , author=

  76. [76]

    arXiv preprint arXiv:2505.24210 , year=

    STORK: Faster Diffusion and Flow Matching Sampling by Resolving Both Stiffness and Structure-Dependence , author=. arXiv preprint arXiv:2505.24210 , year=

    arXiv

  77. [77]

    SADA: Stability-Guided Adaptive Diffusion Acceleration , author=

  78. [78]

    arXiv preprint arXiv:2502.11420 , year=

    TreeG: Tree Search-Based Path Steering Guidance for Diffusion and Flow Matching , author=. arXiv preprint arXiv:2502.11420 , year=

    Pith/arXiv arXiv

  79. [79]

    A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models , author=

  80. [80]

    DuoDiff: Accelerating Diffusion Models with a Dual-Backbone Approach , author=

Showing first 80 references.