Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

A proper distributional energy—the squared energy distance—can supervise a one-step generator through constant-size three-body interactions, reaching ImageNet-256 FID 2.23 (pixel) and 1.63 (latent).

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

2026-08-01 15:41 UTC pith:KE6ZIQ6C

load-bearing objection TBSM has a clean local-gradient identity and an honest appendix, but the headline ImageNet numbers come from the λ=0.9 surrogate, so the 'proper energy' framing overreaches the experiments. the 3 major comments →

arxiv 2607.18198 v1 pith:KE6ZIQ6C submitted 2026-07-20 cs.LG cs.CV

Three-Body Scattering for Generative Modeling

Peng Sun , Zhenglin Cheng , Deyuan Liu , Jun Xie , Xinyi Shang , Tao Lin This is my paper
classification cs.LG cs.CV
keywords generative modelingone-step generationenergy distancethree-body scatteringWasserstein gradient flowdistribution matchingfrozen-target regressionImageNet-256
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.

Three-Body Scattering Modeling (TBSM) tries to show that a proper distributional energy, the squared energy distance, can directly supervise a one-step generator: each generated sample is attracted toward one real sample and repelled from one independently generated sample, yielding a constant-size per-projectile loss. The paper proves that, conditioned on the projectile, the expected scattering vector equals the 2-Wasserstein gradient-flow velocity of the energy-distance functional, so frozen-target regression is locally equivalent to descending the proper objective. An online tracker learns the conditional expectation of the noisy scattering vector and can remove source-sampling variance at the exact endpoint, with exact error decompositions and convergence guarantees under explicit assumptions. Trained in frozen image-representation spaces, TBSM attains one-step ImageNet-256 FID 2.23 with a pixel-space transformer and 1.63 with a latent-space transformer, and maps its tuning parameters onto drift-like, diffusion-related, and GAN-like objectives. If the central claim holds, generation needs no adversarial critic, no denoising path, and no teacher—only a proper energy and per-sample scattering events.

Core claim

The central discovery is that the energy distance reduces to a constant-size per-sample interaction: for a projectile x_p, a real source x_r, and an independent generated source x_s, the bearing difference b = (x_r-x_p)/||x_r-x_p|| - (x_s-x_p)/||x_s-x_p|| has conditional expectation equal to the 2-Wasserstein gradient-flow velocity of 1/2 D_E^2(P_theta, Q). Frozen-target regression toward x_p + b therefore has the same gradient as the energy-distance objective at current parameters (Theorem 1). Tracked scattering estimates E[b | x_p] online, removing source-sampling noise at the exact endpoint with exact error decompositions. In frozen representation spaces this reaches one-step ImageNet-256

What carries the argument

The load-bearing object is the three-body scattering vector field v(x) = E_real[(x_r-x)/||x_r-x||] - E_generated[(x_s-x)/||x_s-x||], whose one-draw estimator b = b_r - b_s is a constant-size per-projectile interaction. Its expectation recovers the negative 2-Wasserstein gradient-flow velocity of the squared energy distance (1/2 D_E^2), and frozen-target regression toward x_p + b pulls this field back through the generator Jacobian, giving Theorem 1's local gradient equivalence. The companion machinery is the online scattering tracker, a network regressed to the instant vector and queried at the projectile (lambda=1) or along a fake-to-real corridor (lambda<1); it estimates the conditional ex

Load-bearing premise

The load-bearing premise is that training in frozen feature spaces with lambda=0.9 (as used in the best latent run) behaves like the theoretically proper lambda=1 energy-distance endpoint for projected distributions, and that the pretrained multi-step initialization is only a warm start—if the encoder drops pixel-critical information or the lambda=0.9 reweighting is what drives the FID gains, the experiments do not isolate the claimed proper-energy mechanism.

What would settle it

Train TBSM from random initialization (no pretrained multi-step checkpoint) at ImageNet scale at the exact lambda=1 endpoint and monitor the squared energy distance in the representation space; if the energy distance fails to decrease while one-step FID improves, or if competitive FID requires lambda<1 reweighting or a teacher-initialized warm start, then the paper's central claim that the proper energy-distance objective drives the results is falsified. A smaller-scale version: compare the parameter gradient of frozen-target three-body regression with the gradient of the squared energy distan

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

If this is right

  • One-step generators can be trained with per-sample supervision that needs only one real reference per condition, so the interface suits text-to-image data with one image per caption and avoids minibatch-wide pairwise fields.
  • At the exact lambda=1 endpoint, a sufficiently accurate tracker preserves the population energy-distance field and removes source-sampling variance; under explicit slope conditions the population flow drives the energy distance to zero.
  • The (rho, lambda) design map gives a common vocabulary for relating energy-distance motion to drift-like particle dynamics, diffusion-related denoising supervision, and GAN-like learned fields, with matched-update comparisons favoring tracked scattering at the energy-distance endpoint.
  • The method is reported to scale: one-step FID 2.23 (pixel) and 1.63 (latent) on ImageNet-256, a one-step conversion at ImageNet-512, and single-step text-to-image fine-tuning—all without classifier-free guidance, adversarial critics, or teacher queries.

Where Pith is reading between the lines

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

  • The representation-space framing suggests a testable division of labour: any frozen encoder defines a projected energy-distance objective, so choosing encoders is an explicit statement about which distributional projections matter; combining several encoders at lambda=1 descends the sum of proper projected discrepancies, which the paper notes does not guarantee full-image matching unless the encod
  • The single-real-source interface implies a concrete prediction about conditional data: on datasets with many images per caption, TBSM should suffer less from same-condition reference scarcity than methods needing same-condition minibatches, while on one-image-per-caption data the paper's own limitation predicts a need for augmentation or regularization to avoid collapsing to the single paired samp
  • The two-draw tracker diagnostic in the paper—paired Monte Carlo estimates of source variance and tracker error—could be turned into a practical scheduling rule for the mixing weight rho, because it estimates the MSE-optimal mixture without access to the true field; this is an engineering extension the paper does not claim.
  • If the lambda<1 tuning effect is genuinely CFG-like (the paper labels it an empirical analogy only), then the same reweighting trick may transfer to other energy-based one-step objectives, but nothing in the proof supports that transfer; it is an extrapolation, not a paper claim.

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 / 5 minor

Summary. The paper proposes Three-Body Scattering Modeling (TBSM), a one-step generative modeling method built on the squared energy distance. The generator is trained by regressing each generated sample toward a detached target obtained from a three-body scattering vector: attraction to one real sample minus repulsion from one independently generated sample. The central theoretical result (Theorem 1, Eq. 10) states that, at the current parameters, this frozen-target regression has the same parameter gradient as the squared energy distance, and the conditional expectation of the scattering vector is the 2-Wasserstein gradient-flow velocity of the energy-distance functional (Eqs. 4, 7). An online tracker approximates this conditional expectation, with exact bias-variance decompositions (Corollary 1, Theorem 2), population-flow convergence under a slope condition and relative tracker error (Theorem 3), and finite-generator stationarity (Proposition 2, Corollary 3). Experiments report one-step ImageNet-256 FID 2.23 with PixelDiT-XL and 1.63 with DiT-XL/2, plus a (ρ, λ) design map and a text-to-image demonstration. The paper is transparent that ImageNet runs initialize from pretrained multi-step checkpoints and that the best latent results use λ=0.9.

Significance. If the claim holds, TBSM is a meaningful contribution: it offers a non-adversarial, non-path-indexed, sample-level supervision signal for one-step generators, with a clean local gradient identity and a useful design map. The theoretical derivation of Theorem 1 is correct under the stated regularity assumptions, and the tracker decompositions are genuine identities rather than hidden circularities. The paper is unusually explicit about what is not proven, including the absence of general convergence guarantees and the restriction of properness to projected distributions. The empirical results are competitive but not state-of-the-art, and the paper does not claim numerical superiority over the strongest baselines. The main gap is that the headline large-scale results do not actually exercise the exact proper energy-distance endpoint: the best latent run uses λ=0.9, the full tracking endpoint (ρ=1, λ=1) is only run on a small JiT-B model, and all ImageNet runs start from pretrained multi-step checkpoints. These gaps are acknowledged in the limitations but need to be resolved or reflected in the abstract's claims.

major comments (3)
  1. [§4.3, Tab. 3; §2.4, Eq. (11); App. A] The headline latent result (FID 1.63, DiT-XL/2) and the ImageNet-512 conversion use (ρ=0.9, λ=0.9), not the proper λ=1 endpoint. At λ<1 the objective is Fλ = F1 + ((1−λ)/2) E[||x−x'||], which is non-proper and shifts the optimum away from Q. The paper's own JiT-B numbers show λ=0.9 improves FID over λ=1 (2.92 vs 3.35), so the reported gains may be due to the dispersion penalty or the corridor-conditioned tracker rather than the energy-distance mechanism emphasized in the abstract. No large-scale run at (ρ=1, λ=1) is reported; the only full-endpoint run is JiT-B. A matched-budget λ=1 latent experiment, ideally also with ρ=1, is needed to support the claim that the proper energy-distance endpoint drives the results; otherwise the abstract should explicitly restrict the claim to the non-proper surrogate.
  2. [App. B, Generator initialization; §6, Limitation 3] All ImageNet-1K TBSM runs except the FD-loss refinement initialize the generator from a pretrained multi-step diffusion or flow-matching checkpoint. Thus NFE=1 demonstrates conversion of an already-trained multi-step generator, not that TBSM can train a one-step generator from random initialization. The paper acknowledges this in Limitation 3 and in App. B, but the abstract's 'route to high-dimensional one-step generation' and the comparison with from-scratch methods such as StyleGAN, Drift, and FD-loss are not controlled for this difference. A from-scratch result, even at reduced scale, or a controlled warm-start ablation with the same initialization for comparator methods, would substantially strengthen the claim that the proposed supervision itself supports one-step generation.
  3. [§2.6, Representation-space scattering; §6, Limitation 4] Even at λ=1, the theory guarantees properness only for the projected distributions f#Pθ and f#Q, not for the full image law. ImageNet scattering is computed in three frozen feature spaces (ResNet-18, SigLIP2-B, MAE-B) and summed; the text-to-image setup uses three further feature spaces. The paper acknowledges that full-image matching requires measure-determining representations, which are not established. The abstract and §5 nevertheless claim that 'a proper distributional energy can induce sample-level motion' without this representational scope. Please either add evidence on the informativeness of the representation fields (for example, a reconstruction or measure-determining test) or revise the abstract and conclusion to state explicitly that the proper energy operates on projected distributions.
minor comments (5)
  1. [Fig. 3] The caption and axis labels are difficult to parse. The corner FID/IS values should be explicitly associated with the (ρ, λ) settings and with the model and training budget used for the design-map comparison.
  2. [App. A] The t=0.5 stage of the high-resolution curriculum uses paired noisy inputs and therefore violates real-source–projectile independence. This is acknowledged as an empirical stabilization heuristic, but it should be flagged more prominently in the main text near the ImageNet-512 results so that readers do not take the conversion experiment as a test of the theoretical field.
  3. [References] There are formatting/encoding artifacts in a few references, e.g., 'W ondraˇcek' in the Schilling et al. entry. Please check all author names for similar issues.
  4. [Fig. 5] The FD-loss checkpoint case study reports that aggregate metrics worsen (FID 0.99→1.84) while the displayed grid appears cleaner. The text says this is illustrative of a metric–artifact mismatch; this should be stated as a qualitative observation on a single case, not as a general property of FD-loss.
  5. [Eq. (23)] The two-draw diagnostic gives unbiased estimates of S1 and Eφ, but the plug-in ratio ρ* is not unbiased and may require clipping. The text already notes this, but the distinction should be repeated in the main text if the diagnostic is used to set ρ.

Circularity Check

0 steps flagged

No circularity: the central gradient identity is proven from the energy-distance first variation, and the disclosed experimental gaps are external-validity caveats, not reductions to inputs.

full rationale

Walking the claimed derivation chain: Eq. (4) defines the scattering field as the difference of expected inter- and intra-source bearings; Eq. (7)'s statement that the triplet estimator has this as its conditional expectation follows directly from definition and source independence. The load-bearing result is Theorem 1 / Eq. (10), which proves that the frozen-target regression loss Lgen has, at current parameters, exactly the generator gradient of F = (1/2)D_E^2(Ptheta,Q). This is a genuine chain-rule identity (App. E.1, Eq. 19), not a fitted parameter renamed as a prediction. Corollary 1 / Theorem 2 are exact bias-variance decompositions around the conditional expectation; they are proven consequences of L2 projection and do not presuppose the conclusion. The tracker is trained on the same bearing vectors it smooths, but the paper claims only the conditional-expectation projection and the explicit condition that tracking helps when tracker excess risk is below source variance (Cor. 1), which is a self-contained tradeoff statement rather than circularity. The empirical ImageNet numbers are external FID/FDr6/IS evaluations, so the paper is not predicting its own training targets. The clearest gaps are disclosed by the paper itself: Sec. 6 concedes random-init ImageNet training is untested and that lambda=1 properness holds only for projected distributions; Sec. 2.4 (Eq. 11) notes lambda<1 shifts the optimum; App. B says all ImageNet runs warm-start from multi-step checkpoints. These are theory-practice mismatches or external-validity caveats, not circular reductions. The only self-citation of note (UCGM checkpoints, App. B) is used solely to initialize parameters and, per the paper, supplies no consistency target, teacher prediction, score, or noise-time target to the TBSM update, so it is not load-bearing. No step in the derivation reduces to its own input by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central method introduces no new physical entities. Its free parameters are the explicit algorithm controls ρ and λ, plus the smoothing ε and standard training hyperparameters; the theory relies on standard but non-trivial regularity and slope assumptions that the paper acknowledges are not automatic.

free parameters (4)
  • λ (intra-source weight) = 0.9 (best latent DiT-XL/2 run), 1.0 or 0.9 in pixel runs
    Controls repulsion strength and tracker query corridor; tuned via FID. Values below 1 make the optimized functional non-proper (Sec. 3, Tab. 3).
  • ρ (tracked-supervision weight) = 0.9 (ImageNet benchmarks), 1.0 (text-to-image)
    Mixes instant and tracked vectors; chosen by validation, not derived; affects the bias-variance tradeoff (Alg. 2, Tab. 3).
  • ε (bearing denominator smoothing) = 1e-6
    Fixed untuned numerical stabilizer; changes the objective to Fε but keeps properness at λ=1 (App. E.5.1).
  • Training hyperparameters (generator LR 1e-5, tracker LR 1e-3, EMA decay 0.999) = As stated
    Chosen training hyperparameters; not part of theoretical identities but needed for the empirical results.
axioms (5)
  • domain assumption Finite first moments, zero collision probability, and valid differentiation under expectation for bearing identities
    Stated in Sec. E; needed for Eq. (7) and Thm. 1.
  • standard math 2-Wasserstein chain rule holds for the energy-distance flow
    Used in E.3.1 dissipation identity (25); requires regularity of Pt.
  • ad hoc to paper Trajectory-wise slope inequality ∥vt∥2 ≥ cF(Pt)^q and relative tracker error bound (Eq. 27)
    Assumed for Thm. 3; the paper states it is substantive and not automatic for kernel discrepancies.
  • domain assumption Realizability, PL inequality, gradient dominance, bounded variance for finite-generator convergence
    Used in Cor. 3 and Prop. 2; not verified for neural generators.
  • domain assumption Frozen representation encoders retain enough information for full image matching
    Limitation 4: properness in representation space holds only for projected distributions.

pith-pipeline@v1.3.0-alltime-deepseek · 30161 in / 13587 out tokens · 121536 ms · 2026-08-01T15:41:48.506708+00:00 · methodology

0 comments
read the original abstract

Modern generative models typically rely on an adversarial critic, a prescribed noise-to-data path, or an autoregressive factorization. Instead, we show that a proper distributional energy can induce sample-level motion and provide direct regression supervision for a one-step generator. Three-Body Scattering Modeling (TBSM) for generation turns the energy distance into a constant-size per-projectile interaction: each projectile is attracted toward one real source and repelled from one independently generated source. Conditioned on the projectile and its condition, its expectation equals the $2$-Wasserstein gradient-flow velocity of $\frac12D_E^2(P_{\theta},Q)$. A batch of $B$ frozen-target events yields $O(B)$ sample-level losses, each using one reference for its condition instead of the minibatch-wide all-pairs field used by methods such as Drifting Models. Tracking this conditional expectation online can reduce field noise. Using scattering in frozen image features, TBSM trains one-step generators on ImageNet-256, achieving FID${}=2.23$ with pixel-space PixelDiT-XL and FID${}=1.63$ with latent-space DiT-XL at NFE${}=1$. We provide a design map relating diffusion-related supervision, Drift-like dynamics, and GAN-like objectives. These results establish tracked scattering as a route to high-dimensional one-step generation. Code: https://github.com/sp12138/TBSM.

Figures

Figures reproduced from arXiv: 2607.18198 by Deyuan Liu, Jun Xie, Peng Sun, Tao Lin, Xinyi Shang, Zhenglin Cheng.

Figure 1
Figure 1. Figure 1: Samples generated in one step by TBSM-trained models. From left to right: MNIST, Fashion￾MNIST, and CIFAR-10 (top to bottom); pixel-space ImageNet-1K samples; and text-to-image samples generated at NFE = 1 by a Qwen-Image-20B (Wu et al., 2025) model fine-tuned using only our TBSM (see App. C). 1 arXiv:2607.18198v1 [cs.LG] 20 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of TBSM. Left: a real–generated–generated event yields a noisy scattering vector. Center: instant scattering drives convergence toward the target along noisy trajectories; tracking smooths the motion. Right: at the fully tracked energy-distance endpoint, an online tracker estimates the conditional expectation of the instant scattering vector at each projectile and supervises the generator. Center-… view at source ↗
Figure 3
Figure 3. Figure 3: One-step generated samples across the (ρ, λ) design map. See Tab. 4 for optimization settings. 3 THE GENERATIVE DESIGN MAP The (ρ, λ) map organizes four reference configurations and interior settings within Alg. 2. 3.1 REFERENCE CONFIGURATIONS AND MATCHED GENERATOR UPDATES Here ρ mixes the instantaneous stochastic vector and learned tracker field, while λ jointly sets the intra-source coefficient in vbλ = … view at source ↗
Figure 4
Figure 4. Figure 4: Converting a multi-step ImageNet-512 model into a one-step generator. Single-pass samples from a pretrained multi-step DiT-XL/4 before (left) and after TBSM training with ρ = λ = 0.9 (right). Both grids are random and uncurated. (a) FID = 0.99 FDr6 = 5.64 IS = 328.0. (b) FID = 1.84 FDr6 = 8.34 IS = 247.3 [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FD-loss-trained JiT-B checkpoint refinement on ImageNet-256. Random, uncurated single-pass samples before (left) and after continued TBSM training with ρ = λ = 0.9 (right). A ADDITIONAL IMAGENET RESULTS Tab. 4 gives optimizer, batch-size, and training-step settings for both studies. Multi-step-to-one-step conversion at 512 × 512. The left panel of [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: One-step text-to-image generation. NFE = 1 samples generated at 1024 × 1024 resolution by a TBSM-fine-tuned Qwen-Image-20B model. and then ℓ2-normalized along C at each spatial location. For scattering, we then identify each normalized C × H × W sample with its vectorization in R CHW , so every bearing denominator uses the global Euclidean norm (equivalently, the tensor Frobenius norm) across all channels … view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Amortized Moment Matching for Visual Generation

    cs.LG 2026-07 accept novelty 6.0

    Amortized Fréchet Distance uses neural nets to match conditional means and covariances, yielding stronger one-step visual generators than explicit FD-loss or multi-step teachers.

Reference graph

Works this paper leans on

76 extracted references · 37 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Gradient Flows in Metric Spaces and in the Space of Probability Measures

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar \'e . Gradient Flows in Metric Spaces and in the Space of Probability Measures. Birkh \"a user, 2 edition, 2008

    2008

  2. [2]

    Maximum mean discrepancy gradient flow

    Michael Arbel, Anna Korba, Adil Salim, and Arthur Gretton. Maximum mean discrepancy gradient flow. In Advances in Neural Information Processing Systems, volume 32, 2019

    2019

  3. [3]

    Qwen2.5-VL technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL technical report. a...

    Pith/arXiv arXiv 2025

  4. [4]

    Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Lakshminarayanan, Stephan Hoyer, and R \'e mi Munos

    Marc G. Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Lakshminarayanan, Stephan Hoyer, and R \'e mi Munos. The Cram \'e r distance as a solution to biased Wasserstein gradients. arXiv preprint arXiv:1705.10743, 2017

    Pith/arXiv arXiv 2017

  5. [5]

    On the global convergence of Wasserstein gradient flow of the Coulomb discrepancy

    Siwan Boufad \`e ne and Fran c ois-Xavier Vialard. On the global convergence of Wasserstein gradient flow of the Coulomb discrepancy. SIAM Journal on Mathematical Analysis, 57 0 (4): 0 4556--4587, 2025

    2025

  6. [6]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018

    Pith/arXiv arXiv 2018

  7. [7]

    A unifying view of variational generative wasserstein flows

    Paul Caucheteux, Cl \'e ment Bonet, and Anna Korba. A unifying view of variational generative wasserstein flows. arXiv preprint arXiv:2605.31369, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    Quantitative convergence of Wasserstein gradient flows of kernel mean discrepancies

    L \'e na \"i c Chizat, Maria Colombo, Roberto Colombo, and Xavier Fern \'a ndez-Real. Quantitative convergence of Wasserstein gradient flows of kernel mean discrepancies. arXiv preprint arXiv:2603.01977, 2026

    arXiv 2026

  9. [9]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp.\ 248--255. Ieee, 2009

    2009

  10. [10]

    Generative modeling via drifting

    Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He. Generative modeling via drifting. arXiv preprint arXiv:2602.04770, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 0 8780--8794, 2021

    2021

  12. [12]

    Wasserstein gradient flows of MMD functionals with distance kernel and cauchy problems on quantile functions

    Richard Duong, Viktor Stein, Robert Beinert, Johannes Hertrich, and Gabriele Steidl. Wasserstein gradient flows of MMD functionals with distance kernel and cauchy problems on quantile functions. ESAIM: Control, Optimisation and Calculus of Variations, 32: 0 10, 2026. doi:10.1051/cocv/2025097

    arXiv 2026

  13. [13]

    Bekkers, and Floor Eijkelboom

    Maria Esteban-Casadevall, Jorge Carrasco-Pollo, Max Welling, Jan-Willem van de Meent, Erik J. Bekkers, and Floor Eijkelboom. Kernel-gradient drifting models. arXiv preprint arXiv:2605.10727, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Representation distribution matching for one-step visual generation

    Lan Feng, Wuyang Li, Eloi Zablocki, Matthieu Cord, and Alexandre Alahi. Representation distribution matching for one-step visual generation. arXiv preprint arXiv:2607.02375, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    One step diffusion via shortcut models

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. arXiv preprint arXiv:2410.12557, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Drifting fields are not conservative

    Leonard T Franz, Sebastian Hoffmann, Tim Weiland, Bernhard Sch \"o lkopf, and Georg Martius. Drifting fields are not conservative. arXiv preprint arXiv:2604.06333, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    Deep MMD gradient flow without adversarial training

    Alexandre Galashov, Valentin de Bortoli, and Arthur Gretton. Deep MMD gradient flow without adversarial training. In International Conference on Learning Representations, 2025

    2025

  18. [18]

    Zico Kolter, and Kaiming He

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J. Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling. arXiv preprint arXiv:2505.13447, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014

    2014

  20. [20]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Sch \"o lkopf, and Alexander Smola. A kernel two-sample test. Journal of Machine Learning Research, 13 0 (25): 0 723--773, 2012

    2012

  21. [21]

    Posterior sampling based on gradient flows of the MMD with negative distance kernel

    Paul Hagemann, Johannes Hertrich, Fabian Altekr \"u ger, Robert Beinert, Jannis Chemseddine, and Gabriele Steidl. Posterior sampling based on gradient flows of the MMD with negative distance kernel. In International Conference on Learning Representations, 2024

    2024

  22. [22]

    Cand \`e s

    Jiaqi Han, Puheng Li, Qiushan Guo, Renyuan Xu, Stefano Ermon, and Emmanuel J. Cand \`e s. One-step generative modeling via wasserstein gradient flows. arXiv preprint arXiv:2605.11755, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

    2016

  24. [24]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll \'a r, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 16000--16009, 2022

    2022

  25. [25]

    Sinkhorn-drifting generative models

    Ping He, Om Khangaonkar, Hamed Pirsiavash, Yikun Bai, and Soheil Kolouri. Sinkhorn-drifting generative models. arXiv preprint arXiv:2603.12366, 2026

    arXiv 2026

  26. [26]

    Generative sliced MMD flows with Riesz kernels

    Johannes Hertrich, Christian Wald, Fabian Altekr \"u ger, and Paul Hagemann. Generative sliced MMD flows with Riesz kernels. In International Conference on Learning Representations, 2024

    2024

  27. [27]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  28. [28]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  29. [29]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  30. [30]

    The variational formulation of the Fokker--Planck equation

    Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the Fokker--Planck equation. SIAM Journal on Mathematical Analysis, 29 0 (1): 0 1--17, 1998

    1998

  31. [31]

    Scaling up gans for text-to-image synthesis

    Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10124--10134, 2023

    2023

  32. [32]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 24174--24184, 2024

    2024

  33. [33]

    Kingma and Ruiqi Gao

    Diederik P. Kingma and Ruiqi Gao. Understanding diffusion objectives as the ELBO with simple data augmentation. In Advances in Neural Information Processing Systems, 2023

    2023

  34. [34]

    MMD GAN : Towards deeper understanding of moment matching network

    Chun-Liang Li, Wei-Cheng Chang, Yu Cheng, Yiming Yang, and Barnab \'a s P \'o czos. MMD GAN : Towards deeper understanding of moment matching network. In Advances in Neural Information Processing Systems, volume 30, 2017

    2017

  35. [35]

    Back to basics: Let denoising generative models denoise

    Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise. arXiv preprint arXiv:2511.13720, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Autoregressive image generation without vector quantization

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37: 0 56424--56445, 2024

    2024

  37. [37]

    Generative moment matching networks

    Yujia Li, Kevin Swersky, and Richard Zemel. Generative moment matching networks. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pp.\ 1718--1727, 2015

    2015

  38. [38]

    Adversarial flow models

    Shanchuan Lin, Ceyuan Yang, Zhijie Lin, Hao Chen, and Haoqi Fan. Adversarial flow models. arXiv preprint arXiv:2511.22475, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Continuous adversarial flow models

    Shanchuan Lin, Ceyuan Yang, Zhijie Lin, Hao Chen, and Haoqi Fan. Continuous adversarial flow models. arXiv preprint arXiv:2604.11521, 2026

    Pith/arXiv arXiv 2026

  40. [40]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  41. [41]

    Stein variational gradient descent: A general purpose Bayesian inference algorithm

    Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose Bayesian inference algorithm. In Advances in Neural Information Processing Systems, volume 29, 2016

    2016

  42. [42]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  43. [43]

    Simplifying, stabilizing and scaling continuous-time consistency models

    Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    One-step latent-free image generation with pixel mean flows

    Yiyang Lu, Susie Lu, Qiao Sun, Hanhong Zhao, Zhicheng Jiang, Xianbang Wang, Tianhong Li, Zhengyang Geng, and Kaiming He. One-step latent-free image generation with pixel mean flows. arXiv preprint arXiv:2601.22158, 2026

    Pith/arXiv arXiv 2026

  45. [45]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision, pp.\ 23--40. Springer, 2024

    2024

  46. [46]

    PixelGen : Improving pixel diffusion with perceptual supervision

    Zehong Ma, Ruihan Xu, and Shiliang Zhang. PixelGen : Improving pixel diffusion with perceptual supervision. arXiv preprint arXiv:2602.02493, 2026

    Pith/arXiv arXiv 2026

  47. [47]

    PyTorch : An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K \"o pf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch : An imperative style, high-per...

    Pith/arXiv arXiv 1912

  48. [48]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 4195--4205, 2023

    2023

  49. [49]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  50. [50]

    Smoothed distance kernels for MMD s and applications in Wasserstein gradient flows

    Nicolaj Rux, Michael Quellmalz, and Gabriele Steidl. Smoothed distance kernels for MMD s and applications in Wasserstein gradient flows. Advances in Computational Mathematics, 52: 0 24, 2026

    2026

  51. [51]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022

    Pith/arXiv arXiv 2022

  52. [52]

    Stylegan-xl: Scaling stylegan to large diverse datasets

    Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pp.\ 1--10, 2022

    2022

  53. [53]

    Schilling, Renming Song, and Zoran Vondra c ek

    Ren \'e L. Schilling, Renming Song, and Zoran Vondra c ek. Bernstein Functions: Theory and Applications. De Gruyter, 2 edition, 2012. doi:10.1515/9783110269338

    doi:10.1515/9783110269338 2012

  54. [54]

    Equivalence of distance-based and RKHS -based statistics in hypothesis testing

    Dino Sejdinovic, Bharath Sriperumbudur, Arthur Gretton, and Kenji Fukumizu. Equivalence of distance-based and RKHS -based statistics in hypothesis testing. The Annals of Statistics, 41 0 (5): 0 2263--2291, 2013

    2013

  55. [55]

    Oriane Sim \'e oni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha \"e l Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timoth \'e e Darcet, Th \'e o Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille ...

    2025

  56. [56]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020 a

    Pith/arXiv arXiv 2010

  57. [57]

    Improved techniques for training consistency models

    Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. arXiv preprint arXiv:2310.14189, 2023

    Pith/arXiv arXiv 2023

  58. [58]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020 b

    Pith/arXiv arXiv 2011

  59. [59]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023

    Pith/arXiv arXiv 2023

  60. [60]

    Autoregressive model beats diffusion: Llama for scalable image generation

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024

    Pith/arXiv arXiv 2024

  61. [61]

    Unified continuous generative models

    Peng Sun, Yi Jiang, and Tao Lin. Unified continuous generative models. arXiv preprint arXiv:2505.07447, 2025

    Pith/arXiv arXiv 2025

  62. [62]

    Sz \'e kely and Maria L

    G \'a bor J. Sz \'e kely and Maria L. Rizzo. Energy statistics: A class of statistics based on distances. Journal of Statistical Planning and Inference, 143 0 (8): 0 1249--1272, 2013

    2013

  63. [63]

    Sriperumbudur, Arthur Gretton, and Zonghao Chen

    Chenyang Tian, Bharath K. Sriperumbudur, Arthur Gretton, and Zonghao Chen. Sobolev regularized MMD gradient flow. arXiv preprint arXiv:2605.11884, 2026

    Pith/arXiv arXiv 2026

  64. [64]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37: 0 84839--84865, 2024

    2024

  65. [65]

    SigLIP 2 : Multilingual vision-language encoders with improved semantic understanding, localization, and dense features

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier H \'e naff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. SigLIP 2 : Multilingual vision-language encoders with improved semantic understanding, localization, and dense feature...

    Pith/arXiv arXiv 2025

  66. [66]

    Coulomb GAN s: Provably optimal Nash equilibria via potential fields

    Thomas Unterthiner, Bernhard Nessler, Calvin Seward, G \"u nter Klambauer, Martin Heusel, Hubert Ramsauer, and Sepp Hochreiter. Coulomb GAN s: Provably optimal Nash equilibria via potential fields. In International Conference on Learning Representations, 2018

    2018

  67. [67]

    Pixel recurrent neural networks

    A \"a ron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of the 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pp.\ 1747--1756, 2016

    2016

  68. [68]

    Qwen-image technical report, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun...

    Pith/arXiv arXiv 2025

  69. [69]

    Representation Fr \'e chet loss for visual generation

    Jiawei Yang, Zhengyang Geng, Xuan Ju, Yonglong Tian, and Yue Wang. Representation Fr \'e chet loss for visual generation. arXiv preprint arXiv:2604.28190, 2026

    Pith/arXiv arXiv 2026

  70. [70]

    Improved distribution matching distillation for fast image synthesis

    Tianwei Yin, Micha \"e l Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024 a

    Pith/arXiv arXiv 2024

  71. [71]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Micha \"e l Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 6613--6623, 2024 b

    2024

  72. [72]

    Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, Alexander G

    Lijun Yu, Jos \'e Lezama, Nitesh B. Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion--tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023

    Pith/arXiv arXiv 2023

  73. [73]

    Autoregressive image generation with masked bit modeling

    Qihang Yu, Qihao Liu, Ju He, Xinyang Zhang, Yang Liu, Liang-Chieh Chen, and Xi Chen. Autoregressive image generation with masked bit modeling. arXiv preprint arXiv:2602.09024, 2026

    arXiv 2026

  74. [74]

    PixelDiT : Pixel diffusion transformers for image generation

    Yongsheng Yu, Wei Xiong, Weili Nie, Yichen Sheng, Shiqiu Liu, and Jiebo Luo. PixelDiT : Pixel diffusion transformers for image generation. arXiv preprint arXiv:2511.20645, 2025

    Pith/arXiv arXiv 2025

  75. [75]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343, 2023

    Pith/arXiv arXiv 2023

  76. [76]

    Perceptual flow matching for few-step generative modeling

    Chuyang Zhao, Yifei Song, Hongfa Wang, Jianlong Yuan, Yuan Zhang, Siming Fu, Zhineng Chen, Huilin Deng, Haoyang Huang, and Nan Duan. Perceptual flow matching for few-step generative modeling. arXiv preprint arXiv:2607.03524, 2026

    Pith/arXiv arXiv 2026