MaskAlign: Token-Subset Representation Alignment for Efficient Diffusion Training

Hugging Face Daily Papers Papers

Summary

MaskAlign proposes a token-subset representation alignment method that improves diffusion transformer training by reducing reliance on complete token sets and maintaining stable alignment under perturbations.

Representation alignment with pretrained vision models has recently shown strong potential for accelerating diffusion transformer training. By aligning intermediate diffusion features with clean-image representations from self-supervised vision encoders, existing methods improve convergence and generation quality. However, such alignment also introduces a non-trivial constraint: diffusion models operate on noisy inputs whose usable information varies across timesteps, while the reference features are extracted from clean images. In this paper, we revisit this mismatch from a token-level perspective. We find that, under full-token representation alignment, tokens with large alignment-gradient norms exhibit a stable spatial preference, suggesting that the alignment objective does not affect all tokens uniformly and may encourage the model to rely on the complete set of clean-image tokens. To address this issue, we propose MaskAlign, a token-subset representation alignment method that applies alignment to randomly sampled token subsets during training. By exposing the model to different token subsets across iterations, MaskAlign reduces the dependence of representation alignment on the complete token set and encourages alignment behavior that is more stable under token-subset perturbations. To mitigate the information loss caused by directly dropping tokens, we further introduce a lightweight pre-mask token mixing block that shares information across tokens before masking.
Original Article
View Cached Full Text

Cached at: 06/12/26, 06:51 AM

Paper page - MaskAlign: Token-Subset Representation Alignment for Efficient Diffusion Training

Source: https://huggingface.co/papers/2606.08788 Published on Jun 7

·

Submitted byhttps://huggingface.co/pangly

panglyon Jun 12

Abstract

Token-subset representation alignment method called MaskAlign improves diffusion transformer training by reducing reliance on complete token sets and maintaining stable alignment behavior under perturbations.

Representation alignmentwith pretrainedvision modelshas recently shown strong potential for accelerating diffusion transformer training. By aligning intermediate diffusion features with clean-image representations fromself-supervised vision encoders, existing methods improve convergence and generation quality. However, such alignment also introduces a non-trivial constraint: diffusion models operate on noisy inputs whose usable information varies across timesteps, while the reference features are extracted from clean images. In this paper, we revisit this mismatch from a token-level perspective. We find that, under full-tokenrepresentation alignment, tokens with largealignment-gradient normsexhibit a stable spatial preference, suggesting that the alignment objective does not affect all tokens uniformly and may encourage the model to rely on the complete set of clean-image tokens. To address this issue, we proposeMaskAlign, atoken-subset representation alignmentmethod that applies alignment to randomly sampled token subsets during training. By exposing the model to different token subsets across iterations,MaskAlignreduces the dependence ofrepresentation alignmenton the complete token set and encourages alignment behavior that is more stable under token-subset perturbations. To mitigate the information loss caused by directly dropping tokens, we further introduce a lightweightpre-mask token mixingblock that shares information across tokens before masking.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2606\.08788

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.08788 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.08788 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.08788 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

PreDiff-LM: Pretrained Discrete Masked Diffusion Language Modeling with Hybrid Attention

arXiv cs.AI

PreDiff-LM proposes a hybrid attention mechanism that preserves causal attention for prompt tokens and bidirectional attention for masked target tokens, enabling adaptation of pretrained autoregressive models for discrete masked diffusion language modeling, achieving improvements in perplexity and downstream tasks over prior diffusion baselines.