Pith. sign in

REVIEW 3 major objections 5 minor 48 references

SPARE: Structural Parameter-Free Affinity Regularization for Flow Matching

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read SPARE shows that the pairwise similarity structure of clean latent tokens can serve as a free training target for diffusion transformers, beating all parameter-free regularizers on ImageNet and improving REPA when combined.

desk verdict SPARE is a clean, parameter-free regularizer with a strong empirical story; the unaddressed affinity-dimension mismatch and quoted baselines keep it from being fully established. read the letter →

arxiv 2608.01990 v1 pith:HWJRBLLO submitted 2026-08-03 cs.CV cs.AI

classification cs.CVcs.AI
keywords flowmatchingrepresentationregularizationdiffusiontransformersaffinityparameter-freecleanlatenttargetsspatialstructureimagegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that the spatial structure needed to accelerate diffusion transformer training is already present in the clean VAE latent, encoded in the pairwise similarity relations among its tokens. Because a pairwise similarity is a single scalar, it is comparable across feature spaces of different widths without a learnable projection head. SPARE regularizes intermediate token features by matching their within-image and cross-image affinities to those of the clean latent. On ImageNet 256×256 with SiT backbones, it attains the lowest FID among parameter-free regularizers, recovers 37–54% of REPA's gains, and improves REPA when added.

What carries the argument

The central object is the affinity matrix A(u,u′) whose entries A_pq are cosine similarities between ℓ2-normalized token rows of two token matrices. This map collapses the ambient dimension: regardless of feature width, the output lives in the same P×P space, eliminating the need for a learnable projection head. The loss is a KL divergence between softmaxed affinity distributions of the intermediate tokens and those of the clean latent, computed over a per-anchor candidate set that includes other positions within the same image and the same position in other images of the batch.

What would settle it

Permute the spatial positions of the clean-latent tokens before computing the SPARE target, keeping everything else equal; if the FID improvement over the baseline remains as large as with the correct positions, then the specific spatial alignment (object boundaries and same-position cross-image correspondences) is not the active signal and the paper's structural claim is false.

Watch

Extended reading notes

Core claim

The clean data latent already carries spatial structure in the relations among its tokens—similarity between two tokens is a scalar comparable across feature spaces without a projection head. This structure exists both within an image (following object boundaries) and across images at corresponding positions. SPARE matches intermediate token affinities to these clean-latent affinities via a single KL divergence over joint affinity distributions. It adds no encoder, head, or parameters, and only 0.08 GB of training memory. On ImageNet 256×256 with SiT backbones under matched 400K-iteration budgets, it attains the lowest FID among parameter-free regularizers in every tested setting, recovers 3

Load-bearing premise

The cosine affinities computed in the clean VAE latent space remain directly comparable to affinities in the intermediate transformer feature space of noised inputs, despite different widths, a large distribution shift, and a learned nonlinear transform.

Editorial extensions

If this is right

  • Parameter-free regularization can capture spatial structure that target-free repulsion destroys, showing that repulsion is not the only head-free option.
  • The clean latent's affinity structure is informative at every noise level, so no timestep scheduling is needed.
  • SPARE composes with encoder-based alignment (REPA), indicating the two signals shape complementary aspects of the representation.
  • Cross-image same-position affinities carry consistent scene layout structure that can be calibrated rather than repelled.
  • Training cost is essentially unchanged, adding only 0.08 GB of memory and no parameters or forward FLOPs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If affinity structure is robust across the stochastic interpolation path as the paper's ablations suggest, SPARE may transfer to other latent diffusion backbones and to higher resolutions with only pooling adjustments to align token grids.
  • The 2.88 FID swing between matching and repelling cross-image affinities suggests that batch-level relational structure is a real inductive bias; a natural extension would be to learn which cross-image pairs to include rather than fixing same-position pairs.
  • Because the target is read off the clean latent, SPARE could be combined with latent-codebook or tokenizer-aware training to propagate structure during generation, not just during training.
  • The success of head-free affinity matching may inform knowledge distillation: relational targets across different widths could replace feature distillation in other generative models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes SPARE, a regularizer for flow-matching diffusion transformers that matches pairwise cosine affinities among intermediate tokens to the affinities of clean VAE latents, both within each image and at matching positions across images in a batch. The affinity target is parameter-free and requires no external encoder or projection head; the loss is a per-anchor KL divergence between softmax distributions over candidate token positions (Eqs. 4–7). On ImageNet 256×256 with SiT backbones under a 400K-iteration training budget, the authors report that SPARE achieves the lowest FID among parameter-free regularizers, recovers 37–54% of REPA's FID reduction, adds only 0.08 GB training memory, and further improves REPA when combined with it, reaching FID 1.90 under CFG at 1M iterations. Ablations, especially the controlled comparison of matching versus repelling cross-image affinities in Table 5d, support the claim that the cross-image signal is beneficial when used as a target and harmful when used for repulsion.

Significance. If the reported results hold, SPARE is a practically valuable contribution: it extracts a structured training signal directly from the clean latent already present in latent diffusion pipelines, eliminating the need for external encoders and learnable projection heads used by prior alignment methods. The strengths of the paper include the explicit cost measurements (Table 4), the transparent ablations that isolate the cross-image target choice (Table 5d), the honest statements of limitations, and the fact that the main effect is demonstrated on a standard benchmark (ImageNet 256×256) with a standard backbone. The idea of using clean-latent affinity as a parameter-free target is simple and novel, and the paper gives credible evidence that the mechanism differs from repulsion-based regularization. However, the central comparability assumption for cross-width affinities is not theoretically established, and several headline comparisons rely on quoted rather than matched baseline numbers; these issues need to be addressed before the paper's strongest claims can be accepted.

major comments (3)
  1. [§Method, Eq. (4)–(7)] The load-bearing premise that cosine affinities are 'single scalars comparable across feature spaces without a projection head' is not established. For random d-dimensional vectors the standard deviation of cosine similarity scales as ~d^{-1/2}; the VAE latent has C=4 while SiT hidden states have C_s=768 or 1152. With a fixed temperature τ=0.5, the target softmax is substantially sharper than the student softmax, so the per-anchor KL in Eq. (7) is initially dominated by this marginal scale mismatch rather than by whether the student's ranking matches the target's ranking. Matching the target probability vector forces the student to amplify the variance of its affinities in addition to aligning their spatial pattern. Table 5c varies τ but does not isolate structure from sharpness. Please provide either a theoretical argument for invariance of the affinity distribution under the interpolat
  2. [Table 2 and Table 1] The composition result in Table 2 is a headline claim, but the REPA baseline rows are quoted from the original paper while the '+SPARE' rows are the authors' own runs. The caption states that only the 1M CFG pair is 'our run for both rows'; at 400K the comparison is therefore not matched in environment, sampler, seed, or hardware. FID differences of 0.41 (XL/2, 400K) and 0.06 (1M) could easily be within run-to-run variation. The same issue affects the 'recovers 37–54% of REPA's reduction' statement in Table 1, where REPA, DiverseDiT, and SRA2 numbers are quoted. Please rerun the quoted baselines under the identical protocol used for the method rows — at minimum REPA for the composition experiment — or clearly report confidence intervals and soften the claim that SPARE 'improves over REPA when combined with it.'
  3. [§Ablation Studies and Fig. 5] The paper relies on single runs for most quantitative claims, including the small but important composition improvements in Table 2 and some ablation differences in Table 5 (e.g., τ=0.5 vs τ=1.0 is only 0.09 FID; λ=0.5 vs λ=1.0 is 0.06 FID). No error bars or multiple-seed summaries are reported, so the reader cannot assess whether differences of this size are significant. This is particularly relevant for the claim that SPARE improves over REPA at 400K, where the gap is 0.41 FID. Please provide at least two or three seeds for the key comparisons, or report the range/variance and discuss whether the differences exceed run-to-run noise.
minor comments (5)
  1. [Abstract] The phrase 'a single scalar comparable across feature spaces' is too strong given the distributional issue discussed in the major comments; consider rephrasing to 'comparable after softmax calibration' or similar.
  2. [Eq. (7), Fig. 2] The loss is described as a KL between 'joint affinity distributions,' but Eq. (7) is actually an average of per-anchor KL divergences over the candidate set. Suggest renaming it 'anchor-wise KL' or 'averaged per-anchor KL' to avoid confusion.
  3. [Fig. 2 caption] The zoom-in mentions self-pairs excluded and the split into within-image and cross-image pairs, but does not mention the stop-gradient on cross-image keys in Eq. (6); adding a note would help readers follow the implementation.
  4. [Table 1 caption] For clarity, please state explicitly in the caption that all non-‡ numbers are quoted from the respective papers and are not matched runs under the authors' protocol; the current wording says this in the body text but the table alone could be misinterpreted.
  5. [Related Work] The reference to Wu et al. 2025 includes a co-author of the present paper (Yao Tang). This is not a problem scientifically, but it may warrant a self-citation disclosure in the final version.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: SPARE's target is a frozen, data-derived affinity matrix; the minor self-citations are related-work only.

full rationale

SPARE's objective (Eq. 7) matches the transformer's intermediate-token cosine affinities to the clean VAE latent's affinities. The target matrix A(v_i) is computed from the frozen data latent (Eq. 4 with u=v), not from the model's own output, so the derivation is not self-referential: the regularizer is an auxiliary training signal rather than a prediction of its own input. Hyperparameters (block ℓ, temperature τ, weight λ) are tuned in Table 5 ablations, which is standard model selection, not fitting-then-predicting. The central comparability premise—that scalar cosine affinities bridge different feature widths without a projection head—is an assumption tested empirically (Figs 1 and 3, Table 5d), not a conclusion imported from a cited theorem. The self-citations (Wu et al. 2025 with co-author Yao Tang; Wang et al. 2026/SRA2 with co-author Linze Li) appear only as related work and comparison baselines; neither supplies the load-bearing argument that affinity structure transfers across the interpolation path. The paper explicitly acknowledges its breadth limitation ('Most findings use SiT backbones on ImageNet 256×256 at 400K iterations, with single runs at 1M and on MS-COCO text-to-image'), which is a generality concern, not circularity. The sharpness/entropy confound raised by the skeptic is a correctness risk about whether the KL gradient encodes structural alignment, but it does not make the method circular: the target is still data-derived and external to the model. Overall, no derivation step reduces by construction to its input; the only issue is minor self-citation overlap that is not load-bearing.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method introduces no new entities. Its free parameters are the hyperparameters lambda, tau, and the block index, chosen by ablations. The main assumptions concern the transferability of affinity structure from the clean latent space to the noised intermediate feature space, and the informativeness of cross-image same-position affinities.

free parameters (3)
  • lambda (loss weight) = 1.0
    Weight of L_SPARE in L = L_FM + lambda * L_SPARE. Chosen by ablation on ImageNet FID (Table 5e).
  • tau (softmax temperature) = 0.5
    Temperature in sigma_tau for the KL calibration. Chosen by ablation (Table 5c); tau=2 degrades noticeably.
  • ell (injection block) = 3 for SPARE alone, 2 for REPA composition
    Block index at which the affinity loss is applied. Chosen by ablation (Table 5b); ell=8 loses all benefit.
assumptions (5)
  • domain assumption The clean VAE latent's token affinity matrix is a valid, informative target for intermediate diffusion transformer features.
    Central premise of the paper, supported by Figs. 1 and 3 and Table 5d, but not proven theoretically.
  • domain assumption Cosine affinity between token vectors is comparable across feature spaces of different widths without any projection.
    Key enabler for removing the projection head; Eq. (4) maps any feature matrix to the same P x P affinity space.
  • domain assumption Cross-image same-position affinity in clean latents carries meaningful spatial structure, so calibrating to it is beneficial.
    Supported by the non-uniform map in Fig. 3 and the align-versus-repel ablation in Table 5d.
  • domain assumption The frozen VAE encoder (sd-vae-ft-ema) provides a spatially faithful representation across the noise schedule.
    The method relies on the VAE latent as a stable spatial reference; no proof is given that its affinity structure survives interpolation to noised inputs.
  • standard math Flow matching / stochastic interpolant framework with linear schedules.
    Background formulation used to define the denoising objective and the intermediate token features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPARE: Structural Parameter-Free Affinity Regularization for Flow Matching." pith.science (2026). https://pith.science/paper/HWJRBLLO

@misc{pith2026260801990,
  author       = {Pith},
  title        = {Pith review of: SPARE: Structural Parameter-Free Affinity Regularization for Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HWJRBLLO}},
  note         = {Machine review of arXiv:2608.01990}
}
abstract

Denoising diffusion transformers achieve strong generation quality but converge slowly during training. Regularizing their internal representations has emerged as an effective accelerator, yet existing methods split into two families with complementary costs. Target-based methods strengthen representations by aligning them to external features, which requires an external encoder and a learnable projection head to bridge feature spaces. Target-free methods hold no reference at all, and can only repel the model's own features across samples or layers, discarding whatever structure the data contains. Prior work suggests that spatial structure, rather than global semantics, drives the gains of alignment. We therefore ask whether such structure can serve as a target directly, and whether it exists not only within an image but across images. Our key insight is that the clean data latent already carries this structure in the relations among its tokens, where a relation is the similarity between two tokens, a single scalar comparable across feature spaces without a projection head. We propose Structural Parameter-free Affinity Regularization (SPARE), a regularizer that matches the pairwise affinities of intermediate tokens to those of the clean latents. To exploit this structure fully, SPARE extends the matching to token pairs across images, precisely the pairs that prior target-free methods repel by default, and calibrates both relation types with a single learning objective. On ImageNet $256 \times 256$ with SiT backbones under matched 400K-iteration budgets, SPARE adds no encoder, head, or parameters and only 0.08 GB of training memory, yet attains the lowest FID among parameter-free regularizers in every tested setting, recovers 37 to 54\% of REPA's FID reduction, and improves over REPA when combined with it, reaching FID 1.90 under classifier-free guidance at 1M iterations.

Figures

Figures reproduced from arXiv: 2608.01990 by the authors.

Figure 1
Figure 1. Each map shows the cosine similarity between the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Four ways to regularize a diffusion transformer’s representations. (a) Vanilla training, where the latent’s structure goes unused. (b) Target-based alignment (REPA, SRA, SRA2) matches each image’s tokens to reference features through a learnable MLP head (red); the reference may come from a frozen external encoder (blue, drawn) or from internal substitutes. (c) Target-free objectives drop the target and repel the mo… view at source ↗
Figure 3
Figure 3. Cross-image same-position affinity of clean la￾tents. For each spatial position, we average the cosine simi￾larity between the clean VAE latents of two different images at that same position (later formalized as App(vi , vi ′ )), over ∼20K images in class-shuffled batches. Values are shown as deviation from the whole-map average (white = map mean, red above, blue below). The deviations are spatially orga￾nized, elev… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: SiT-XL/2 generations at 100K, 200K, and 400K iterations (CFG [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: FID-50K across training for the baseline (left) and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 12 linked inside Pith

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

Show all 48 references
  1. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  2. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

  3. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  4. [12]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [13]

    Journal of Machine Learning Research , volume=

    Stochastic interpolants: A unifying framework for flows and diffusions , author=. Journal of Machine Learning Research , volume=

  6. [14]

    European Conference on Computer Vision , pages=

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers , author=. European Conference on Computer Vision , pages=. 2024 , organization=

  7. [15]

    arXiv preprint arXiv:2410.06940 , year=

    Representation alignment for generation: Training diffusion transformers is easier than you think , author=. arXiv preprint arXiv:2410.06940 , year=

  8. [16]

    arXiv preprint arXiv:2512.10794 , year=

    What matters for Representation Alignment: Global Information or Spatial Structure? , author=. arXiv preprint arXiv:2512.10794 , year=

  9. [17]

    arXiv preprint arXiv:2505.02831 , year=

    No Other Representation Component Is Needed: Diffusion Transformers Can Provide Representation Guidance by Themselves , author=. arXiv preprint arXiv:2505.02831 , year=

  10. [18]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    SRA 2: Variational Autoencoder Self-Representation Alignment for Efficient Diffusion Training , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  11. [19]

    arXiv preprint arXiv:2506.09027 , year=

    Diffuse and disperse: Image generation with representation regularization , author=. arXiv preprint arXiv:2506.09027 , year=

  12. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Diversedit: Towards diverse representation learning in diffusion transformers , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [21]

    arXiv preprint arXiv:2010.02502 , year=

    Denoising diffusion implicit models , author=. arXiv preprint arXiv:2010.02502 , year=

  14. [22]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  15. [23]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

  16. [24]

    arXiv:2209.03003 , year =

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author =. arXiv:2209.03003 , year =

  17. [25]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Repa-e: Unlocking vae for end-to-end tuning of latent diffusion transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  18. [26]

    Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think , url =

    Wu, Ge and Zhang, Shen and Shi, Ruijing and Gao, Shanghua and Chen, Zhenyuan and Wang, Lei and Chen, Zhaowei and Gao, Hongcheng and Tang, Yao and Yang, jian and Cheng, Ming-Ming and Li, Xiang , booktitle =. Representation Entanglement for Generation: Training Diffusion Transfo...

  19. [27]

    arXiv preprint arXiv:2011.13456 , year=

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

  20. [28]

    Advances in neural information processing systems , volume=

    Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=

  21. [29]

    Advances in neural information processing systems , volume=

    Video diffusion models , author=. Advances in neural information processing systems , volume=

  22. [30]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Adding conditional control to text-to-image diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  23. [31]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  24. [32]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  25. [33]

    Advances in neural information processing systems , volume=

    Improved techniques for training gans , author=. Advances in neural information processing systems , volume=

  26. [34]

    arXiv preprint arXiv:2103.03841 , year=

    Generating images with sparse representations , author=. arXiv preprint arXiv:2103.03841 , year=

  27. [35]

    Advances in neural information processing systems , volume=

    Improved precision and recall metric for assessing generative models , author=. Advances in neural information processing systems , volume=

  28. [36]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

  29. [37]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

  30. [38]

    arXiv preprint arXiv:2304.07193 , year=

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

  31. [39]

    arXiv preprint arXiv:2510.11690 , year=

    Diffusion transformers with representation autoencoders , author=. arXiv preprint arXiv:2510.11690 , year=

  32. [40]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Relational knowledge distillation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  33. [41]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Similarity-preserving knowledge distillation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  34. [42]

    Advances in Neural Information Processing Systems , volume=

    Ressl: Relational self-supervised learning with weak augmentation , author=. Advances in Neural Information Processing Systems , volume=

  35. [43]

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

    Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  36. [44]

    arXiv preprint arXiv:2605.16949 , year=

    Beyond Point-Wise Matching: Structural Representation Alignment for Accelerating Diffusion Transformers , author=. arXiv preprint arXiv:2605.16949 , year=

  37. [45]

    arXiv preprint arXiv:2605.03317 , year=

    AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers , author=. arXiv preprint arXiv:2605.03317 , year=

  38. [46]

    arXiv preprint arXiv:2509.21565 , year=

    No Alignment Needed for Generation: Learning Linearly Separable Representations in Diffusion Models , author=. arXiv preprint arXiv:2509.21565 , year=

  39. [47]

    Forty-first international conference on machine learning , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first international conference on machine learning , year=

  40. [48]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.