Pith. sign in

REVIEW 4 major objections 5 minor 10 cited by

The paper argues diffusion transformers fail on encoder features because Euclidean flow paths cut through the sphere's interior, not from capacity limits, and geodesic paths with curvature weighting let a 131M DiT-B reach FID 3.37.

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

2026-08-03 01:15 UTC pith:LUTNDWXZ

load-bearing objection The core fix — spherical flow matching on DINOv2 latents — is real and worth attention, but the paper's lossless-sphere premise is undercut by its own Fig. 6, and the SOTA claims are not supported by its own Table 2. the 4 major comments →

arxiv 2602.10099 v2 pith:LUTNDWXZ submitted 2026-02-10 cs.LG cs.CV

Learning on the Manifold: Unlocking Standard Diffusion Transformers with Representation Encoders

Amandeep Kumar , Vishal M. Patel This is my paper
classification cs.LG cs.CV
keywords diffusion transformersflow matchingRiemannian flow matchinggeometric interferencehyperspherical manifoldsrepresentation encodersDINOv2Jacobi fields
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.

Flow-matching diffusion models are known to struggle when trained directly on the feature spaces of representation encoders such as DINOv2, and a recent line of work blames model width. This paper tries to establish that the real cause is geometric: those features sit on a fixed-radius hypersphere, so all semantic information lives in direction, yet the standard Euclidean flow path is a straight chord through the sphere's low-density interior. That mismatch forces the network to spend its capacity on a meaningless radial error and on trajectories where the representation is undefined — what the paper calls Geometric Interference. The proposed fix, Riemannian Flow Matching with Jacobi Regularization (RJF), keeps every intermediate state on the sphere along geodesics and reweights the loss for curvature-induced error growth; with it, an off-the-shelf 131M-parameter DiT-B converges to an FID (a standard image-quality score) of 3.37 with guidance and 4.95 without in 200 epochs, where baselines fail to converge. If the diagnosis is right, the practical message is that the convergence wall is removable by correcting the objective rather than by scaling the architecture.

Core claim

The paper's central claim: standard diffusion transformers fail on representation-encoder latents for a geometric reason, not capacity shortage. These features sit on a fixed-radius hypersphere — all semantics are angular — so the Euclidean interpolation, whose norm collapses to ≈0.7√d at mid-path, cuts through the sphere's interior. They show the model wastes its rank on this radial error: masking it lets even a 384-wide model succeed. The fix, RJF, swaps linear paths for spherical linear interpolation (SLERP) geodesics and adds a Jacobi-field weight concentrating learning where curvature amplifies error. With RJF, a standard DiT-B (131M) reaches FID 3.37 (guidance) and 4.95 (none) in 200 e

What carries the argument

The carrying object: the hyperspherical geometry of representation features — LayerNorm pins them to a sphere, z = r·ẑ, r ≈ √d, so all information lives in ẑ. Two mechanisms do the work. Spherical Linear Interpolation (SLERP) — xt = [sin((1−t)Ω)x + sin(tΩ)ϵ]/sin(Ω), Ω = arccos(x·ϵ) — keeps every intermediate state on the sphere and the target velocity in the tangent space, eliminating the radial error by construction. Jacobi Regularization reweights the loss by λ(t,Ω) = sinc²((1−t)Ω), from the Jacobi equation on the sphere, so errors near the noise endpoint — amplified by geodesic focusing — count most. The radial/angular loss decomposition supplies the evidence: mask the radial part and eve

Load-bearing premise

The argument stands or falls on the claim that representation features are exactly hyperspherical — that the radius carries no information, so unit-sphere projection (which the method itself performs in Algorithm 1) is lossless and a purely angular generative process suffices; if radial variation encodes signal the frozen decoder needs, the geometric model is incomplete.

What would settle it

Two checks would settle it: (1) measure the per-token radial distribution of DINOv2-B features actually fed to the transformer — if ‖z‖₂ variance is non-negligible or correlates with image class or statistics, the lossless-projection premise fails; (2) decode the frozen RAE at the true feature norm (R≈27.7) versus the paper's optimal R≈45 and compare image fidelity — if the decoder is strongly scale-sensitive, the radius is not inert and a purely angular model passes the decoder a scale it must then compensate for, qualifying the geometric claim.

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

If this is right

  • Width scaling is unnecessary: the same standard DiT-B/L/XL architectures that fail under Euclidean flow matching converge under RJF — DiT-B reaches FID 3.37 (guided) and 4.95 (unguided) at 200 epochs, and DiT-XL reaches 3.62 unguided at 80 epochs.
  • The diagnosis transfers across encoders: RJF takes SigLIP from FID 130.21 to 10.39 and MAE from 50.48 to 19.82 on DiT-B, consistent with the paper's attribution of the shared failure to LayerNorm and contrastive normalization-induced spherical geometry.
  • Fixing the endpoints is not enough: projecting the noise onto the sphere without changing the path improves FID only marginally (24.32 to 21.99), confirming that the chord trajectory itself, not the endpoint mismatch, is the failure mechanism.
  • The geometric fix accelerates schedules: DiT-XL with RJF reaches FID 6.32 in 24 epochs, beating a strong baseline trained for 80 epochs (FID 6.94), implying much shorter training runs are viable for representation-latent diffusion.
  • Semantic fidelity rises with geometric alignment: RJF reports Inception Score 186.2 and precision 0.82 on DiT-XL, which the authors read as the model capturing high-fidelity semantic modes rather than averaging textures.

Where Pith is reading between the lines

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

  • The shell-radius premise is partially strained by the paper's own Fig. 6: best generation happens at inference radius R≈45, well above the true feature norm √d≈27.7. A natural extension is to make the decoder scale-insensitive or train it at the true radius, which would either erase the need for this rescaling or reveal that the radius carries more information than the projection premise admits.
  • If the geometric diagnosis is correct, the same chord-versus-geodesic failure should appear in any generative setting whose latent space is a thin shell or sphere — contrastive embeddings in text or audio models, or normalized hidden states in autoregressive models — and the SLERP-plus-Jacobi recipe should transfer there without architectural change.
  • For a fixed data point, a uniform noise vector is nearly orthogonal to it in high dimension, so Ω concentrates near π/2 and the per-sample Jacobi weight is almost a deterministic function of time; a clean test is to replace λ(t,Ω) with λ(t,π/2) and check whether FID changes — if it does not, curvature correction is effectively a time-reweighting rather than a per-sample geometric adjustment.

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

4 major / 5 minor

Summary. The paper argues that standard Euclidean flow matching on representation-encoder latents (DINOv2, SigLIP, MAE) fails not because of a capacity bottleneck, but because linear interpolation paths cut through the low-density interior of the approximately hyperspherical feature distribution. The authors propose Riemannian Flow Matching with Jacobi Regularization (RJF): normalize latents to the unit sphere, use SLERP geodesic interpolation, regress tangent-space velocities, and reweight the loss by sinc²((1−t)Ω), with sampling via exponential-map integration and a final radius scaling R. They report that this enables a standard 131M DiT-B to reach FID 3.37 with guidance in 200 epochs and DiT-XL to reach FID 3.62 in 80 epochs without guidance, whereas Euclidean baselines fail to converge.

Significance. If validated, the paper makes a useful and nontrivial contribution: it challenges the width-scaling explanation of RAE and provides a simple geometric fix that materially improves convergence. The central ablation in Table 3 is strong—replacing Euclidean interpolation with SLERP reduces DiT-B FID from 24.32 to 7.06—and the paper gives explicit algorithms and a geometric derivation of the Jacobi weighting. These are concrete, reproducible assets. However, the geometric premise is not fully supported by the paper's own radius experiment, and the reported gains are single-run FIDs without uncertainty quantification. The core idea is worth serious further consideration, but the manuscript needs substantial revision before the headline claims can be accepted.

major comments (4)
  1. [Sec. 5.3, Fig. 6] The lossless-sphere premise is contradicted by the paper's own inference-radius experiment. Eq. (1) and Algorithm 1 treat the radius as fixed at √d and normalize it away, but Algorithm 2 restores a single R at the output. Fig. 6 reports FID 7.79 at R≈27.7 (the natural DINOv2-B norm) and FID 6.77 at R≈45. This shows that the RAE decoder is scale-sensitive, or that the learned angular distribution is biased, so 'all information is encoded in angular vectors' is not valid at the decoding stage. The headline FIDs therefore depend on an empirically fitted radius that is not predicted by the manifold model. Please report how R was selected, provide a validation protocol, and reconcile the result with the claim that the generative process stays on the true data manifold.
  2. [Algorithm 2, lines 4–5] The inference algorithm samples t_raw from a LogitNormal and applies a time shift immediately before a loop that uses undefined t_i and t_{i+1}. A 50-step deterministic sampler should use a fixed schedule (e.g., t_i = i/N), not a random draw. As written, the sampling procedure is not reproducible and the loop is ill-defined. Please state the inference time schedule explicitly or remove the copied training-time sampling lines.
  3. [Tables 1–5, Sec. 4] All quantitative claims are based on single training runs with no error bars or multiple seeds. Some differences that support secondary claims are small: Table 4 shows DiT DH 6.33 vs. 6.20 with RJF, and Table 3 shows RFM 7.06 vs. RJF 6.77. These margins could easily be within run-to-run FID variability. The main SLERP-vs-Euclidean improvement is large, but the Jacobi-regularization gain and the radius sweep in Fig. 6 need error bars or seeded repeats before they can support the stated conclusions.
  4. [Sec. 3.3, Eq. (14), Appendix B] The Jacobi weighting λ(t,Ω)=sinc²((1−t)Ω) is introduced as a correction to an RFM objective that 'treats errors uniformly,' but RFM on the sphere already optimizes in the Riemannian metric. The derivation in Eqs. (20)–(23) quantifies endpoint displacement per unit tangent perturbation, yet the final weight is a per-sample scalar multiplier; it reweights the loss rather than changing the conditional target. Its benefit may be real, but the framing as a necessary curvature correction is not fully justified. Please clarify why the standard Riemannian metric is insufficient, and if the gain over plain RFM is marginal, present the weighting as a training heuristic rather than a derived geometric correction.
minor comments (5)
  1. [Sec. 2.1, Eq. (3)] The approximation ⟨x,ϵ⟩≈0 is valid at high dimension, but the finite-d dimension fluctuations are O(√d). A short sentence quantifying this would make the norm-collapse argument more precise.
  2. [Algorithm 1] When Ω≈0 or Ω≈π, the SLERP denominator sin(Ω) is numerically unstable. Please state the fallback used for near-degenerate pairs.
  3. [Table 4] The first two rows are both labeled 'DiT-XL/1' but report different FIDs (4.29 and 4.28). Likely a typo; also unify model naming with Table 1 and Table 3 ('DiT-B/1' vs 'LightningDiT-B/1').
  4. [Fig. 6] The y-axis label 'gFID' is unclear (guidance FID? FID without guidance?). Add axis ticks and a caption explaining the metric and whether this is the same 80-epoch DiT-B model used elsewhere.
  5. [Global] Typos and naming inconsistencies: 'DiNO' for DINOv2, 'SigLip' for SigLIP, 'high-dimesion'; also Algorithm 2 restores radius R at line 17, but the main text does not state in Sec. 4.1 which R was used for the headline results.

Circularity Check

0 steps flagged

No significant circularity: the core derivation is self-contained and tested against independent baselines; the tuned inference radius is a reported limitation, not a circular reduction.

full rationale

The paper's derivation chain is not circular. The geometric premise (Eq. 1) is an empirical observation about DINOv2 features, and the chord/norm-collapse argument (Eqs. 2-3) follows from high-dimensional orthogonality; neither is defined in terms of the conclusions. The Riemannian path (Eq. 9), tangent velocity (Eq. 10), and Jacobi weight (Eq. 13; Appendix B, Eqs. 16-23) are closed-form mathematical constructions with no fitted parameters and no reliance on the authors' own prior work. The Jacobi weight is derived from standard Jacobi-field equations, not imported as an unverified self-citation. The main empirical claims are validated against independent baselines (Euclidean FM, +SN, REPA, LightningDiT) using the same DiT backbones and decoder, and the ablation in Table 3 isolates the contributions of RFM vs. Jacobi regularization. No self-citation is load-bearing, and no uniqueness theorem from the authors is invoked. The one genuinely fitted element is the inference-time radius R in Sec. 5.3/Fig. 6: the paper transparently reports that the natural DINOv2-B norm (R≈27.7) gives FID 7.79 while R≈45 gives 6.77, attributing this to decoder scale sensitivity. This is a limitation and a robustness/correctness concern, because it undercuts the 'radius is information-free' premise, but it is a hyperparameter choice, not a prediction reduced by construction to its input. The paper does not claim R is derived from the manifold model, so this does not constitute circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on the hyperspherical geometry of the features and a handful of training/inference hyperparameters; the geometry is empirically motivated, the Jacobi weighting is derived, and one hyperparameter (inference radius) is fitted to the evaluation metric.

free parameters (3)
  • Inference radius R = ≈45 (best FID); natural feature norm ≈√768≈27.7
    Section 5.3/Fig. 6 sweeps radius and picks the value minimizing FID; all headline FID numbers (6.77/4.95/3.37) rely on this tuned radius.
  • LogitNormal time-sampling parameters (µ, σ) and shift s(n=4096) = not reported (inherited from RAE)
    Algorithm 1 requires these; the paper only states it follows RAE's setting, so exact values are not reproducible from the text.
  • Classifier-free guidance scale = not reported
    The with-guidance result FID 3.37 (Table 3) does not state the guidance weight used; a standard hyperparameter with strong effect on FID.
axioms (4)
  • domain assumption DINOv2/SigLIP/MAE features are confined to a hypersphere of fixed radius with negligible radial variance (Eq. 1, Fig. 3).
    The method projects features to the unit sphere and discards radius; if radial variance carries signal, information is lost.
  • domain assumption Uniform-on-sphere noise is a sufficient source prior for generating the target feature distribution.
    Algorithm 1 samples projected Gaussian noise; this assumes uniform geodesic transport can match the target distribution.
  • standard math Jacobi-field analysis correctly maps per-step velocity errors to endpoint error on S^{d-1}, justifying the sinc^2 weighting (Appendix B).
    Derivation uses standard Jacobi equations; the application to the loss assumes errors are perturbations along geodesics from x_t to the noise endpoint, which is an approximation.
  • domain assumption The frozen RAE decoder remains a valid decoder for unit-sphere latents rescaled to radius R.
    The decoder was trained on original features (norm √d) but receives latents scaled to R≈45 at inference; Fig. 6 shows the decoder is scale-sensitive.

pith-pipeline@v1.3.0-alltime-deepseek · 15536 in / 21279 out tokens · 216522 ms · 2026-08-03T01:15:19.200777+00:00 · methodology

0 comments
read the original abstract

Leveraging representation encoders for generative modeling offers a path for efficient, high-fidelity synthesis. However, standard diffusion transformers fail to converge on these representations directly. While recent work attributes this to a capacity bottleneck proposing computationally expensive width scaling of diffusion transformers we demonstrate that the failure is fundamentally geometric. We identify Geometric Interference as the root cause: standard Euclidean flow matching forces probability paths through the low-density interior of the hyperspherical feature space of representation encoders, rather than following the manifold surface. To resolve this, we propose Riemannian Flow Matching with Jacobi Regularization (RJF). By constraining the generative process to the manifold geodesics and correcting for curvature-induced error propagation, RJF enables standard Diffusion Transformer architectures to converge without width scaling. Our method RJF enables the standard DiT-B architecture (131M parameters) to converge effectively, achieving an FID of 3.37 where prior methods fail to converge. Code: https://github.com/amandpkr/RJF

Figures

Figures reproduced from arXiv: 2602.10099 by Amandeep Kumar, Vishal M. Patel.

Figure 1
Figure 1. Figure 1: Bridging the Geometric Gap. We demonstrate that respecting the intrinsic geometry of pre-trained representations en￾coders enables the use of standard Diffusion Transformers without any architectural modification such as Width Scaling (Zheng et al., 2025). Our method, Riemannian Flow Matching with Jacobi Regu￾larization (+DiNO+RJF), achieves an FID of 4.95 using standard LightingDiT-B (Yao et al., 2025) ar… view at source ↗
Figure 2
Figure 2. Figure 2: Geometric Trajectories on the Hypersphere. Visual￾ization of flow matching paths on the manifold S d−1 . Standard Euclidean Flow Matching constructs linear paths that ignore the manifold geometry. Whether targeting standard Gaussian noise ϵ (orange) or projecting noise onto the sphere ϵs (purple) to strictly learn the angular component, the linear interpolation forms a chord that cuts through the low-densi… view at source ↗
Figure 3
Figure 3. Figure 3: The Geometry Gap. A comparison of radial feature norms (r = ∥z∥2) between DINOv2-B representations and a stan￾dard Gaussian prior in R 768. While the Gaussian prior (blue) is distributed across a diffuse shell, DINOv2-B features (orange) are rigidly constrained to a hypersphere with near-zero radial variance. This extreme geometric mismatch prevents standard diffusion mod￾els from converging effectively. A… view at source ↗
Figure 4
Figure 4. Figure 4: Geometric Interference vs. Capacity. We train DiT-S models of varying widths on DINOv2 tokens (d = 768). Top Row: When minimizing Euclidean MSE, narrower models (d < 768) suffer from collapse; the Angular Loss (semantics) gets stuck. Bottom Row: When the radial loss is ignored, even narrow models (d = 384) converge perfectly on the angular component. This proves the bottleneck is not the dimensionality of … view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of LightingDiT-XL+RJF trained for 80 epochs on ImageNet 256×256. We show uncurated results on the five classes . 4.3. Ablation Study We investigate the impact of each geometric component by training a standard LightingDiT-B/1 model on DINOv2- B latents. The results are summarized in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance across different radius projection We further analyze the impact of the projection radius R during the in￾ference stage ( [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results of LightingDiT-XL+RJF trained for 80 epochs on ImageNet 256×256. We show uncurated results on the 12 classes . 14 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results of LightingDiT-XL+RJF trained for 80 epochs on ImageNet 256×256. We show uncurated results on the 12 classes . 15 [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] 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 10 Pith papers

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

  1. Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention

    cs.CV 2026-07 unverdicted novelty 7.0

    The subspace intervention framework reveals that pre-training objectives shape how ViTs encode geometric information in compressible low-rank subspaces, with peak precision at intermediate layers.

  2. STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation

    cs.CV 2026-06 unverdicted novelty 7.0

    STREAM applies stochastic Riemannian flow matching on VFM-derived unit hypersphere latents with a novel anisotropic decoder to achieve SOTA reconstruction and generation on breast and colorectal cancer histopathology ...

  3. Latent Riemannian Flow Matching for Geometry-Grounded 3D Foundation Models

    cs.CV 2026-07 conditional novelty 6.0

    Using a product-of-hyperspheres Riemannian flow matching model on VGGT's latent codes, the authors generate plausible depth, point maps, and RGB for target views from one to four unposed context images.

  4. Orbis 2: A Hierarchical World Model for Driving

    cs.CV 2026-07 conditional novelty 6.0

    A hierarchical driving world model — planning in compressed DINO space at 2 Hz and rendering detailed frames at 10 Hz — achieves state-of-the-art long-horizon stability, steering response, and representation quality.

  5. Diffusion Image Generation with Explicit Modeling of Data Manifold Geometry

    cs.CV 2026-05 unverdicted novelty 6.0

    MIND integrates discrete patch tokenization into diffusion score functions via soft top-k and dual-branch layers, achieving FID 22.73 (no guidance) and 2.06 (with guidance) on ImageNet-256 after 80 epochs, outperformi...

  6. Diffusion Image Generation with Explicit Modeling of Data Manifold Geometry

    cs.CV 2026-05 unverdicted novelty 6.0

    MIND integrates discrete patch tokenization into diffusion score functions via soft top-k aggregation and dual-branch embeddings, achieving FID 1.95 on ImageNet-256 with 715M parameters while outperforming larger baselines.

  7. RADAR: Relative Angular Divergence Across Representations

    cs.LG 2026-05 unverdicted novelty 6.0

    RADAR is a geometrically grounded metric that predicts cross-domain transferability by comparing layer-wise representation trajectory distributions in foundation models.

  8. RiT: Vanilla Diffusion Transformers Suffice in Representation Space

    cs.CV 2026-05 conditional novelty 6.0

    A vanilla Diffusion Transformer trained via x-prediction on frozen DINOv2 features reaches FID 1.14 on ImageNet 256x256 with fewer parameters and faster sampling than prior DiT variants.

  9. Aligning Latent Geometry for Spherical Flow Matching in Image Generation

    cs.CV 2026-05 unverdicted novelty 6.0

    Projecting VAE latents to a fixed spherical radius and replacing linear interpolation with spherical linear interpolation improves class-conditional ImageNet-256 FID while leaving the diffusion architecture unchanged.

  10. Geometry-Aware Representation Denoising for Robust Multi-view 3D Reconstruction

    cs.CV 2026-05 unverdicted novelty 5.0

    GARD performs diffusion-based multi-view restoration in the feature space of a feed-forward 3D reconstructor to recover scene geometry and RGB images under degraded conditions, shown effective on the DA3 benchmark.

Reference graph

Works this paper leans on

30 extracted references · 18 linked inside Pith · cited by 9 Pith papers

  1. [1]

    Albergo, M. S. and Vanden-Eijnden, E. Building normal- izing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571,

    Pith/arXiv arXiv

  2. [5]

    Mdtv2: Masked diffusion transformer is a strong image synthe- sizer.arXiv preprint arXiv:2303.14389,

    Gao, S., Zhou, P., Cheng, M.-M., and Yan, S. Mdtv2: Masked diffusion transformer is a strong image synthe- sizer.arXiv preprint arXiv:2303.14389,

    Pith/arXiv arXiv

  3. [7]

    Scalable adaptive computation for iterative generation.arXiv preprint arXiv:2212.11972,

    Jabri, A., Fleet, D., and Chen, T. Scalable adaptive computation for iterative generation.arXiv preprint arXiv:2212.11972,

    Pith/arXiv arXiv

  4. [8]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes.CoRR, abs/1312.6114,

    Pith/arXiv arXiv

  5. [11]

    Repa-e: Unlocking vae for end-to-end tuning with latent diffusion transformers.arXiv preprint arXiv:2504.10483,

    Leng, X., Singh, J., Hou, Y ., Xing, Z., Xie, S., and Zheng, L. Repa-e: Unlocking vae for end-to-end tuning with latent diffusion transformers.arXiv preprint arXiv:2504.10483,

    arXiv

  6. [12]

    Flow matching meets biology and life science: a survey.arXiv preprint arXiv:2507.17731,

    Li, Z., Zeng, Z., Lin, X., Fang, F., Qu, Y ., Xu, Z., Liu, Z., Ning, X., Wei, T., Liu, G., et al. Flow matching meets biology and life science: a survey.arXiv preprint arXiv:2507.17731,

    arXiv

  7. [13]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y ., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    Pith/arXiv arXiv

  8. [14]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003,

    Pith/arXiv arXiv

  9. [15]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193,

    Oquab, M., Darcet, T., Moutakanni, T., V o, H., Szafraniec, M., Khalidov, V ., Fernandez, P., Haziza, D., Massa, F., El- Nouby, A., et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193,

    Pith/arXiv arXiv

  10. [16]

    Semantics lead the way: Harmonizing semantic and texture modeling with asynchronous latent diffusion.arXiv preprint arXiv:2512.04926,

    Pan, Y ., Feng, R., Dai, Q., Wang, Y ., Lin, W., Guo, M., Luo, C., and Zheng, N. Semantics lead the way: Harmonizing semantic and texture modeling with asynchronous latent diffusion.arXiv preprint arXiv:2512.04926,

    arXiv

  11. [17]

    The intrinsic dimension of images and its impact on learning.arXiv preprint arXiv:2104.08894,

    Pope, P., Zhu, C., Abdelkader, A., Goldblum, M., and Gold- stein, T. The intrinsic dimension of images and its impact on learning.arXiv preprint arXiv:2104.08894,

    Pith/arXiv arXiv

  12. [19]

    P., Kumar, A., Er- mon, S., and Poole, B

    Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Er- mon, S., and Poole, B. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456,

    Pith/arXiv arXiv 2011

  13. [21]

    Represen- tation entanglement for generation: Training diffusion transformers is much easier than you think.arXiv preprint arXiv:2507.01467,

    Wu, G., Zhang, S., Shi, R., Gao, S., Chen, Z., Wang, L., Chen, Z., Gao, H., Tang, Y ., Yang, J., et al. Represen- tation entanglement for generation: Training diffusion transformers is much easier than you think.arXiv preprint arXiv:2507.01467,

    arXiv

  14. [22]

    Y ., Gastegger, M., Jim´enez-Luna, J., Lewis, S., Satorras, V

    Yim, J., Campbell, A., Foong, A. Y ., Gastegger, M., Jim´enez-Luna, J., Lewis, S., Satorras, V . G., Veeling, B. S., Barzilay, R., Jaakkola, T., et al. Fast protein back- bone generation with se (3) flow matching.arXiv preprint arXiv:2310.05297,

    Pith/arXiv arXiv

  15. [23]

    Representation alignment for generation: Training diffusion transformers is easier than you think

    Yu, S., Kwak, S., Jang, H., Jeong, J., Huang, J., Shin, J., and Xie, S. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940,

    Pith/arXiv arXiv

  16. [24]

    Zaghen, O., Eijkelboom, F., Pouplin, A., and Bekkers, E. J. Towards variational flow matching on general geometries. arXiv preprint arXiv:2502.12981,

    arXiv

  17. [25]

    Diffusion trans- formers with representation autoencoders.arXiv preprint arXiv:2510.11690,

    Zheng, B., Ma, N., Tong, S., and Xie, S. Diffusion trans- formers with representation autoencoders.arXiv preprint arXiv:2510.11690,

    Pith/arXiv arXiv

  18. [26]

    Fast training of diffusion models with masked transformers

    Zheng, H., Nie, W., Vahdat, A., and Anandkumar, A. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305,

    Pith/arXiv arXiv

  19. [27]

    accelerate convergence by aligning intermediate diffusion features with pretrained representations such as DINOv2 (Oquab et al., 2023). This paradigm has been extended by architectures like DDT (Wang et al., 2025b), which applies alignment to a decoupled encoder-decoder structure, and REG (Wu et al., 2025), which introduces a learnable class token for exp...

    2023

  20. [28]

    jointly learns low-level and high-level semantic distributions, and approaches like V A-V AE (Yao et al., 2025), and SVG (Shi et al.,

    2025

  21. [29]

    Standard FM constructs probability paths via linear interpolation in Euclidean space, regressing a velocity field to guide samples from a source distribution to the data

    has emerged as a robust, simulation-free alternative for training Continuous Normalizing Flows (CNFs). Standard FM constructs probability paths via linear interpolation in Euclidean space, regressing a velocity field to guide samples from a source distribution to the data. However, as noted in recent geometric deep learning literature (Rozen et al., 2021;...

    2021

  22. [30]

    generalizes the framework by replacing Euclidean straight lines with geodesic paths defined by the Riemannian metric. By defining the conditional probability path as a geodesic interpolation, RFM ensures that the flow remains strictly on the manifold, avoiding regions of low density such as the interior of a hypersphere. This formulation has been widely a...

    2023

  23. [2013]

    9 Kouzelis, T., Karypidis, E., Kakogeorgiou, I., Gidaris, S., and Komodakis, N

    URL https: //api.semanticscholar.org/CorpusID: 216078090. 9 Kouzelis, T., Karypidis, E., Kakogeorgiou, I., Gidaris, S., and Komodakis, N. Boosting generative image mod- eling via joint image-feature synthesis.arXiv preprint arXiv:2504.16064,

    arXiv

  24. [2015]

    Latent diffusion model without variational autoencoder.arXiv preprint arXiv:2510.15301,

    Shi, M., Wang, H., Zheng, W., Yuan, Z., Wu, X., Wang, X., Wan, P., Zhou, J., and Lu, J. Latent diffusion model without variational autoencoder.arXiv preprint arXiv:2510.15301,

    arXiv

  25. [2020]

    Simpler diffusion (sid2): 1.5 fid on imagenet512 with pixel-space diffusion.arXiv preprint arXiv:2410.19324,

    Hoogeboom, E., Mensink, T., Heek, J., Lamerigts, K., Gao, R., and Salimans, T. Simpler diffusion (sid2): 1.5 fid on imagenet512 with pixel-space diffusion.arXiv preprint arXiv:2410.19324,

    Pith/arXiv arXiv

  26. [2021]

    Pixnerd: Pixel neural field diffusion.arXiv preprint arXiv:2507.23268, 2025a

    Wang, S., Gao, Z., Zhu, C., Huang, W., and Wang, L. Pixnerd: Pixel neural field diffusion.arXiv preprint arXiv:2507.23268, 2025a. Wang, S., Tian, Z., Huang, W., and Wang, L. Ddt: Decoupled diffusion transformer.arXiv preprint arXiv:2504.05741, 2025b. Wang, T. and Isola, P. Understanding contrastive represen- tation learning through alignment and uniformit...

    Pith/arXiv arXiv

  27. [2022]

    J., Akhound-Sadegh, T., Huguet, G., Fatras, K., Rector-Brooks, J., Liu, C.-H., Nica, A

    Bose, A. J., Akhound-Sadegh, T., Huguet, G., Fatras, K., Rector-Brooks, J., Liu, C.-H., Nica, A. C., Korablyov, M., Bronstein, M., and Tong, A. Se (3)-stochastic flow matching for protein backbone generation.arXiv preprint arXiv:2310.02391,

    Pith/arXiv arXiv

  28. [2023]

    Pixelflow: Pixel-space generative models with flow.arXiv preprint arXiv:2504.07963,

    Chen, S., Ge, C., Zhang, S., Sun, P., and Luo, P. Pixelflow: Pixel-space generative models with flow.arXiv preprint arXiv:2504.07963,

    Pith/arXiv arXiv

  29. [2024]

    Chen, R. T. and Lipman, Y . Flow matching on general geometries.arXiv preprint arXiv:2302.03660,

    Pith/arXiv arXiv

  30. [2025]

    A., Hu, V

    Krause, F., Phan, T., Gui, M., Baumann, S. A., Hu, V . T., and Ommer, B. Tread: Token routing for efficient architecture-agnostic diffusion training.arXiv preprint arXiv:2501.04765,

    arXiv