Diffusion Transformer

Also known as DiT

Ask about this topic

Overview

Diffusion Transformers (DiTs) use transformer networks to turn noise into data, typically by repeatedly predicting denoising updates or following a learned flow. Work explores latent- and pixel-space models, conditioning through text, images, geometry, or other modalities, and representation-learning methods that improve generation and downstream perception. To lower the cost of training and inference, researchers use sparse attention, token reduction, caching, quantization, distillation, and hardware-aware designs. A central challenge is balancing speed, memory, and model size against fidelity, temporal or spatial consistency, and reliable performance across tasks and benchmarks.

Latest papers

CardsList

473 papers78 papers in the last four weeks, up 100% on the four weeks before. 0.5% of all new papers.

May 13, 2026

Test-time Sparsity for Extreme Fast Action Diffusion

Action diffusion excels at high-fidelity action generation but incurs heavy computational costs owing to its iterative denoising nature. Despite current technologies showing promise in accelerating diffusion transformers by reusing the cached features, they struggle to adapt to policy dynamics arising from diverse perceptions and multi-round rollout iterations in open environments. We propose test-time sparsity to tackle this challenge, which aims to accelerate action diffusion by dynamically predicting prunable residual computations for each model forward at test time. However, two bottlenecks remain in this paradigm: 1) repetitive conditional encoding and pruning offset most potential speed gains, and 2) the features cached from previous denoising timesteps cannot constrain large pruning errors under aggressive sparsity. To address the first bottleneck, we design a highly parallelized inference pipeline that minimizes the non-decoder delay to milliseconds. Specifically, we first design a lightweight pruner that shares the encoder with the diffusion transformer. Then, we decouple the encoding and pruning from the autoregressive denoising loop by processing all denoising timesteps in parallel, and overlap the pruner with the decoder forward inference through asynchronism. To overcome the second bottleneck, we introduce an omnidirectional reusing strategy, which achieves 95% sparsity by selectively reusing features cached from the current forward, previous denoising timesteps, and earlier rollout iterations. To learn the rollout-level reusing strategies, we sample a few action trajectories to supervise the sparsified diffusion step by step. Extensive experiments demonstrate that our method reduces FLOPs by 92% and accelerates action generation by 5x, achieving lossless performance with an inference frequency of 47.5 Hz. Our code is available at https://github.com/ky-ji/Test-time-Sparsity.

Kangye Ji, Yuan Meng, Jianbo Zhou +3

Tsinghua University · The Chinese University of Hong Kong

Diffusion TransformerInference-Time OptimizationDiffusion PolicyDiffusion Model CachingDiffusion Model Inference Acceleration
AskPDF
May 12, 2026

Sparse Mixture-of-Experts Routing in Visual Diffusion Transformers:Diagnosis, Boundary Calibration and Evolutionary Roadmap from Routing Collapse to Selective Deadlock

This paper systematically diagnoses the training failure modes of Token-Choice sparse Mixture-of-Experts (MoE) on video Diffusion Transformers. Starting from a pretrained dense model of about 5 billion parameters, we convert it into an MoE architecture following three laws: routed experts exactly clone the original FFN weights, shared experts are initialized to zero for verification and then to extremely small non-zero noise for actual training, while only the gating networks start from random initialization. Experiments reveal a hierarchy of five failure modes: (1) linear routers suffer global soft saturation with complete expert homogenization; (2) MLP routers introduce selective deadlock, where roughly one-third of layers degenerate into a single-expert mode that cannot be prevented by increasing the auxiliary loss; (3) cross-attention routers exhibit preliminary self-recovery, yet about nine layers remain stubbornly deadlocked; (4) deadlocked layers display a U-shaped distribution, concentrated in shallow visual processing layers and deep semantic integration layers; (5) bfloat16 mixed precision causes tiny weight updates to be truncated to zero by hardware. Based on routing decision time series over 65 million tokens across 5,000 training steps, we propose the Functional Redundancy Hypothesis: deadlock is a rational waiting strategy before the shared expert matures within the gate-shared expert-routed expert triadic system. This hypothesis is supported by the theory of functional redundancy in systems biology. On the engineering side, we summarize the Three Laws of dense-to-MoE conversion and provide a complete solution for the bfloat16 precision trap. We calibrate the current capability boundary of the Token-Choice paradigm and outline a three-step evolutionary roadmap from visual unification to a world model.

Haiying Sha

Independent Researcher

Video Diffusion ModelsDiffusion TransformerSparse Mixture-of-ExpertsExpert RoutingMixture-of-Experts Models
AskPDF
May 12, 2026

FIS-DiT: Breaking the Few-Step Video Inference Barrier via Training-Free Frame Interleaved Sparsity

While the overall inference latency of Video Diffusion Transformers (DiTs) can be substantially reduced through model distillation, per-step inference latency remains a critical bottleneck. Existing acceleration paradigms primarily exploit redundancy across the denoising trajectory; however, we identify a limitation where these step-wise strategies encounter diminishing returns in few-step regimes. In such scenarios, the scarcity of temporal states prevents effective feature reuse or predictive modeling, creating a formidable barrier to further acceleration. To overcome this, we propose Frame Interleaved Sparsity DiT (FIS-DiT), a training-free and operator-agnostic framework that shifts the optimization focus from the temporal trajectory to the latent frame dimension. Our approach is motivated by an intrinsic duality within this dimension: the existence of frame-wise sparsity that permits reduced computation, coupled with a structural consistency where each frame position remains equally vital to the global spatiotemporal context. Leveraging this insight, we implement Frame Interleaved Sparsity (FIS) as an execution strategy that manipulates frame subsets across the model hierarchy, refreshing all latent positions without requiring full-scale block computation. Empirical evaluations on Wan 2.2 and HunyuanVideo 1.5 demonstrate that FIS-DiT consistently achieves 2.11--2.41×\times speedup with negligible degradation across VBench-Q and CLIP metrics, providing a scalable and robust pathway toward real-time high-definition video generation.

Jian Tang, Jiawei Fan, Qingbin Liu +1

Platform and Content Group, Tencent

Video Diffusion ModelsDiffusion TransformerFew-Step Diffusion SamplingDiffusion Model Inference Acceleration
AskPDF
May 11, 2026

ChunkFlow: Communication-Aware Chunked Prefetching for Layerwise Offloading in Distributed Diffusion Transformer Inference

Layerwise offloading reduces the GPU memory footprint of large diffusion transformer (DiT) inference by prefetching upcoming layers from host memory, but its effectiveness hinges on hiding prefetch latency behind per-layer computation. This assumption breaks down when the per-GPU compute workload is small. Moreover, on PCIe-only nodes, prefetch and inter-GPU collective communications such as all-reduce and all-to-all contend on the shared PCIe path, exposing prefetch latency even when compute would otherwise hide it. We revisit layerwise offloading as a co-scheduling problem between prefetch and communication, guided by a first-order analytical model that predicts when prefetch can be hidden by computation. Building on this model, we design ChunkFlow, a communication-aware, chunk-granular offloading runtime that adaptively yields to collective communication and smoothly trades GPU memory for prefetch volume. On three representative diffusion transformers running on two H100 GPUs over PCIe with Ulysses sequence parallelism, ChunkFlow delivers up to 1.28x step-time speedup over SGLang's existing layerwise offloading, reduces peak GPU memory by up to 49% over the no-offload baseline at near-identical step time once the workload is large enough, and exposes a tunable memory-latency tradeoff that recovers near-zero step-time overhead in the small-workload regime.

Han Meng, Danny Willow Liu, Dong Li

University of California, Merced · University of Chicago · University of California, Merced and Yotta Labs +1

Diffusion TransformerGPU AccelerationMemory-Efficient InferenceDistributed InferenceDiffusion Model Inference Acceleration
AskPDF
May 11, 2026

HiDream-O1-Image: A Natively Unified Image Generative Foundation Model with Pixel-level Unified Transformer

The evolution of visual generative models has long been constrained by fragmented architectures relying on disjoint text encoders and external VAEs. In this report, we present HiDream-O1-Image, a natively unified generative foundation model via pixel-space Diffusion Transformer, that pioneers a paradigm shift from modular architectures to an end-to-end in-context visual generation engine. By mapping raw image pixels, text tokens, and task-specific conditions into a single shared token space, HiDream-O1-Image achieves a structural unification of multimodal inputs within an Unified Transformer (UiT) architecture. This native encoding paradigm eliminates the need for separate VAEs or disjoint pre-trained text encoders, allowing the model to treat diverse generation and editing tasks as a consistent in-context reasoning process. Extensive experiments show that HiDream-O1-Image excels across various generation tasks, including text-to-image generation, instruction-based editing, and subject-driven personalization. Notably, with only 8B parameters, HiDream-O1-Image (8B) achieves performance parity with or even surpasses established state-of-the-art models with significantly larger parameters (e.g., 27B Qwen-Image). Crucially, to validate the immense scalability of this paradigm, we successfully scale the architecture up to over 200B parameters. Experimental results demonstrate that this massive-scale version HiDream-O1-Image-Pro (200B+) unlocks unprecedented generative capabilities and superior performance, establishing new state-of-the-art benchmarks. Ultimately, HiDream-O1-Image highlights the immense potential of natively unified architectures and charts a highly scalable path toward next-generation multimodal AI.

Qi Cai, Jingwen Chen, Chengmin Gao +22

Diffusion TransformerText-Guided Image EditingUnified Multimodal ModelsPixel-Space Diffusion ModelsMultimodal Foundation ModelsT2I Diffusion Models+1
AskPDF
May 11, 2026

What Concepts Lie Within? Detecting and Suppressing Risky Content in Diffusion Transformers

The rise of text-to-image (T2I) models has increasingly raised concerns regarding the generation of risky content, such as sexual, violent, and copyright-protected images, highlighting the need for effective safeguards within the models themselves. Although existing methods have been proposed to eliminate risky concepts from T2I models, they are primarily developed for earlier U-Net architectures, leaving the state-of-the-art Diffusion-Transformer-based T2I models inadequately protected. This gap stems from a fundamental architectural shift: Diffusion Transformers (DiTs) entangle semantic injection and visual synthesis via joint attention, which makes it difficult to isolate and erase risky content within the generation. To bridge this gap, we investigate how semantic concepts are represented in DiTs and discover that attention heads exhibit concept-specific sensitivity. This property enables both the detection and suppression of risky content. Building on this discovery, we propose AHV-D&S, a training-free inference-time safeguard for image generation in DiTs. Specifically, AHV-D&S quantifies each textual token's sensitivity across all attention heads as an Attention Head Vector (AHV), which serves as a discriminative signature for detecting risky generation tendencies. In the inference stage, we propose a momentum-based strategy to dynamically track token-wise AHVs across denoising steps, and a sensitivity-guided adaptive suppression strategy that suppresses the attention weights of identified risky tokens based on head-specific risk scores. Extensive experiments demonstrate that AHV-D&S effectively suppresses sexual, copyrighted-style, and various harmful content while preserving visual quality, and further exhibits strong robustness against adversarial prompts and transferability across different DiT-based T2I models.

Chenyu Zhang

Tianjin University · Tianjin, China

Diffusion TransformerT2I GenerationT2I Diffusion Models
AskPDF
May 10, 2026

Attention Sinks in Diffusion Transformers: A Causal Analysis

Attention sinks -- tokens that receive disproportionate attention mass -- are assumed to be functionally important in autoregressive language models, but their role in diffusion transformers remains unclear. We present a causal analysis in text-to-image diffusion, dynamically identifying dominant attention recipients per timestep and suppressing them via paired, training-free interventions on the score and value paths. Across 553 GenEval prompts on Stable Diffusion~3 (with SDXL corroboration), removing these sinks does not degrade text-image alignment (CLIP-T) or preference proxies (ImageReward, HPS-v2) at k=1k{=}1; only under stronger interventions (k ⁣≥ ⁣10k\!\geq\!10) does HPS-v2 exhibit a metric-dependent boundary, while CLIP-T remains robust throughout. The perceptual shifts induced by suppression are nonetheless sink-specific -- ∼ ⁣6×\sim\!6\times larger than equal-budget random masking -- revealing an empirical dissociation between trajectory-level perturbation and semantic alignment in diffusion transformers. \footnote{Code available at https://github.com/wfz666/ICML26-attention-sink.}

Fangzheng Wu, Brian Summa

Department of Computer Science, Tulane University, New Orleans, LA, USA.

Transformer InterpretabilityDiffusion TransformerSelf-AttentionAttention SinksT2I Diffusion ModelsDiffusion Models
AskPDF
May 10, 2026

Remix the Timbre: Diffusion-Based Style Transfer Across Polyphonic Stems

Timbre transfer aims to modify the timbral identity of a musical recording while preserving its original melody and rhythm. While single-instrument timbre transfer has made substantial progress, existing approaches to multi-instrument settings commonly rely on separate-then-transfer pipelines that propagate source-separation artifacts and process the resulting stems independently. This paper proposes MixtureTT, to the best of our knowledge the first system for flexible per-stem timbre transfer directly from a polyphonic mixture. Given a mixture and a separate timbre reference for each target voice, MixtureTT jointly transfers all stems to the specified instruments. Its Joint Stem Diffusion Transformer first models cross-stem harmonic context and then refines each output under its own content and timbre conditions. All stems are generated within a single joint diffusion process rather than through separate per-stem diffusion runs. This design avoids propagating waveform-level separation errors through a cascade and produces more coherent multi-stem outputs. Despite operating under a harder input condition, MixtureTT outperforms practical separate-then-transfer pipelines on objective and subjective evaluations, while remaining competitive with single-instrument methods given ground-truth isolated stems. Results on both synthesized and real-audio SATB benchmarks further demonstrate its generalization. Controlled ablations show that jointly modeling cross-stem dependencies improves over an equivalent independent-stem variant, while the structure and timing of this coordination are also critical for effective mixture-level timbre transfer.

Leduo Chen, Junchuan Zhao, Shengchen Li

XJTLU

Diffusion TransformerControllable Music GenerationAudio Diffusion Models
AskPDF
May 8, 2026

BRIDGE: Background Routing and Isolated Discrete Gating for Coarse-Mask Local Editing

Coarse-mask local editing must preserve the background while allowing new content to depart from the mask contour. Existing mask-conditioned editors often entangle these goals and treat the coarse mask as a shape prior. We introduce BRIDGE, a dual-path architecture with a full-canvas Main Path, a local Subject Path, and a learned binary gate that routes subject-token positional embeddings between subject-centric and background-anchored coordinates. On Qwen-Image-Edit-2511, BRIDGE keeps the mask outside the DiT backbone and trains both targets with independent noise. On BRIDGE-Bench, it achieves 0.503 local SigLIP2-T, 0.902 DINO, and 0.175 DreamSim, outperforming ACE++ and FLUX.1-Fill in local precision. We then instantiate BRIDGE on FLUX.2 [klein] 9B Base with full-transformer training and subject-reference conditioning. Dense BBox supervision preserves Sub-image quality; optional mask-cropped inference removes out-of-mask Sub tokens, tightening user support without training a sparse model that overfits the mask silhouette. On supported ICE-Bench tasks, FLUX.2 BRIDGE averages 0.5941 over four local edits and 0.6254 over subject-guided edits, leading the compared open-source models on the latter. Across Qwen and FLUX.2, BRIDGE separates support localization from subject geometry and improves local generation while preserving the surrounding scene. Code is available at https://github.com/XIONGPEILIN/BRIDGE

Peilin Xiong, Honghui Yuan, Junwen Chen +1

Department of Informatics, The University of Electro-Communications, Tokyo, Japan

Diffusion TransformerPositional EncodingImage InpaintingDiffusion-Based Image EditingImage Editing
AskPDF
May 8, 2026

TextLDM: Language Modeling with Continuous Latent Diffusion

Diffusion Transformers (DiT) trained with flow matching in a VAE latent space have unified visual generation across images and videos. A natural next step toward a single architecture for both generation (visual synthesis) and understanding (text generation) is to apply this framework to language modeling. We propose TextLDM, which transfers the visual latent diffusion recipe to text generation with minimal architectural modification. A Transformer-based VAE maps discrete tokens to continuous latents, enhanced by Representation Alignment (REPA) with a frozen pretrained language model to produce representations effective for conditional denoising. A standard DiT then performs flow matching in this latent space, identical in architecture to its visual counterpart. The central challenge we address is obtaining high-quality continuous text representations: we find that reconstruction fidelity alone is insufficient, and that aligning latent features with a pretrained language model via REPA is critical for downstream generation quality. Trained from scratch on OpenWebText2, TextLDM substantially outperforms prior diffusion language models and matches GPT-2 under the same settings. Our results establish that the visual DiT recipe transfers effectively to language, taking a concrete step toward unified diffusion architectures for multimodal generation and understanding.

Jiaxiu Jiang, Jingjing Ren, Wenbo Li +10

Joy Future Academy · HIT · HKUST(GZ)

Diffusion TransformerLatent Diffusion ModelsDiffusion Language ModelsContinuous Diffusion Language ModelsDiffusion Language Model Training
AskPDF
May 8, 2026

EditTransfer++: Toward Faithful and Efficient Visual-Prompt-Guided Image Editing

Visual-prompt-guided edit transfer aims to learn image transformations directly from example pairs, offering more precise and controllable editing than purely text-driven approaches. However, existing diffusion transformer-based methods often fail to faithfully reproduce the demonstrated edits due to structural mismatches between the task and the backbone, including a pretrained bias toward textual conditioning and inherent stochastic instability during sampling. To bridge this gap, we present EditTransfer++, a framework that combines progressively structured training with an efficient conditioning scheme to improve both visual prompt faithfulness and inference efficiency. We first mitigate textual dominance with a text-decoupled training strategy that removes text conditioning during fine-tuning, compelling the model to infer transformations solely from visual evidence while still supporting optional text guidance at inference. On top of this visually grounded model, a best-worst contrastive refinement mechanism reshapes the denoising trajectories to suppress unfaithful generations and improve consistency across random seeds. To alleviate the computational bottleneck of high-resolution in-context editing, we further introduce a condition compression and reuse strategy that reduces token redundancy and enables efficient generation of images with a 1024-pixel long edge. Extensive experiments on existing benchmarks and the proposed EditTransfer-Bench show that EditTransfer++ achieves state-of-the-art visual prompt faithfulness with substantially faster inference than prior methods, suggesting a promising direction for scalable prompt-guided image editing and broader visual in-context learning.

Lan Chen, Qi Mao, Yiren Song +2

School of Information and Communication Engineering and the State Key Laboratory of Media Convergence and Communication, Communication University of China, Beijing 100024, China · ShowLab, National University of Singapore, 119077, Singapore · State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, Beijing 100871, China

Diffusion TransformerDiffusion Model Fine-TuningDiffusion-Based Image Editing
AskPDF
May 7, 2026

Not All Tokens Need 40 Steps: Heterogeneous Step Allocation in Diffusion Transformers for Efficient Video Generation

Diffusion Transformers (DiTs) have achieved state-of-the-art video generation quality, but they incur immense computational cost because standard inference applies the same number of denoising steps uniformly to every token in the sequence. It is well known that human vision ignores vast amounts of redundant motion. Why, then, do our densest models treat every spatiotemporal token with equal priority? In this paper, we introduce Heterogeneous Step Allocation (HSA), a training-free inference algorithm that assigns varying step budgets to different spatiotemporal tokens based on their velocity dynamics. To resolve the resulting sequence-length mismatch without sacrificing global context, HSA introduces a KV-cache synchronization mechanism that allows active tokens to attend to the full sequence while entirely bypassing inactive tokens. Furthermore, we derive a cached Euler update that advances the latent states of skipped tokens in a single operation without additional model evaluations. We evaluate HSA on the Wan-2 and LTX-2 models for both text-to-video (T2V) and image-to-video (I2V) generation. Our results demonstrate that HSA significantly outperforms previous state-of-the-art caching methods and the vanilla Flow Matching baseline, especially at aggressive acceleration regimes (e.g., 50% and 25% runtimes). Crucially, HSA achieves a superior quality-runtime Pareto frontier without the need for expensive offline profiling, robustly preserving structural integrity and generation quality even under tight computational budgets. Project page: https://ernestchu.github.io/hsa

Ernie Chu, Vishal M. Patel

Johns Hopkins University Baltimore, MD 21218

Video Diffusion ModelsDiffusion TransformerDiffusion Model CachingDiffusion Timestep SchedulingDiffusion Model Inference Acceleration
AskPDF
May 7, 2026

Continuous Latent Diffusion Language Model

Large language models have achieved remarkable success under the autoregressive paradigm, yet high-quality text generation need not be tied to a fixed left-to-right order. Existing alternatives still struggle to jointly achieve generation efficiency, scalable representation learning, and effective global semantic modeling. We propose Cola DLM, a hierarchical latent diffusion language model that frames text generation through hierarchical information decomposition. Cola DLM first learns a stable text-to-latent mapping with a Text VAE, then models a global semantic prior in continuous latent space with a block-causal DiT, and finally generates text through conditional decoding. From a unified Markov-path perspective, its diffusion process performs latent prior transport rather than token-level observation recovery, thereby separating global semantic organization from local textual realization. This design yields a more flexible non-autoregressive inductive bias, supports semantic compression and prior fitting in continuous space, and naturally extends to other continuous modalities. Through experiments spanning 4 research questions, 8 benchmarks, strictly matched ~2B-parameter autoregressive and LLaDA baselines, and scaling curves up to about 2000 EFLOPs, we identify an effective overall configuration of Cola DLM and verify its strong scaling behavior for text generation. Taken together, the results establish hierarchical continuous latent prior modeling as a principled alternative to strictly token-level language modeling, where generation quality and scaling behavior may better reflect model capability than likelihood, while also suggesting a concrete path toward unified modeling across discrete text and continuous modalities.

Hongcan Guo, Qinyu Zhao, Yian Zhao +8

ByteDance Seed · The Australian National University · Peking University +2

Non-Autoregressive Text GenerationDiffusion TransformerLatent Diffusion ModelsDiffusion Language ModelsContinuous Diffusion Language Models
AskPDF
May 7, 2026

Why Do DiT Editors Drift? Plug-and-Play Low Frequency Alignment in VAE Latent Space

Recent advances in diffusion transformers (DiTs) have enabled promising single-turn image editing capabilities. However, multi-turn editing often leads to progressive semantic drift and quality degradation.In this work, we study this problem from a latent-space frequency perspective by decomposing the editing process into two functional components: VAE and DiT. Through systematic analysis in the VAE latent space, we uncover that the DiT introduces dominant low-frequency drift that accumulates as semantic misalignment across editing rounds, while the VAE contributes comparatively stable reconstruction bias.Based on this insight, we propose VAE-LFA (Low Frequency Alignment), a training-free, plug-and-play method that performs alignment in VAE latent space. VAE-LFA decomposes latent discrepancies across editing rounds via low-pass filtering, and aligns low-frequency statistics to an exponential moving average of previous rounds, effectively suppressing accumulated semantic drift while preserving high-frequency details.Our method requires no retraining, ground-truth priors, or access to diffusion parameters, making it applicable to both white-box and black-box DiT editors. For white-box models, VAE-LFA is seamlessly integrated into the editing pipeline by eliminating redundant VAE round trips; for black-box models, it operates via an off-the-shelf VAE to perform inter-round latent alignment.Extensive experiments demonstrate that VAE-LFA improves semantic consistency and visual fidelity across diverse multi-turn editing scenarios, including both controlled and in-the-wild images.

Xiaoce Wang, Sifan Zhou, Kaifei Wang +4

♠Tsinghua University · ♡Carnegie Mellon University · ♢Peking University +3

Diffusion TransformerMulti-Turn Image EditingDiffusion-Based Image EditingImage Editing
AskPDF
May 7, 2026

Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers

Scaling Diffusion Transformers (DiTs) to hundreds of layers introduces a structural vulnerability: networks can enter a silent, mean-dominated collapse state that homogenizes token representations and suppresses centered variation. Through mechanistic auditing, we isolate the trigger event of this collapse as Mean Mode Screaming (MMS). MMS can occur even when training appears stable, with a mean-coherent backward shock on residual writers that opens deep residual branches and drives the network into a mean-dominated state. We show this behavior is driven by an exact decomposition of these gradients into mean-coherent and centered components, compounded by the structural suppression of attention-logit gradients through the null space of the Softmax Jacobian once values homogenize. To address this, we propose Mean-Variance Split (MV-Split) Residuals, which combine a separately gained centered residual update with a leaky trunk-mean replacement. On a 400-layer single-stream DiT, MV-Split prevents the divergent collapse that crashes the un-stabilized baseline; it tracks close to the baseline's pre-crash trajectory while remaining substantially better than token-isotropic gating methods such as LayerScale across the full schedule. Finally, we present a 1000-layer DiT as a scale-validation run at boundary scales, establishing that the architecture remains stably trainable at extreme depth.

Pengqi Lu

Beijing, China

Deep Learning OptimizationRepresentation CollapseDiffusion TransformerResidual Learning
AskPDF
May 6, 2026

Taming Outlier Tokens in Diffusion Transformers

We study outlier tokens in Diffusion Transformers (DiTs) for image generation. Prior work has shown that Vision Transformers (ViTs) can produce a small number of high-norm tokens that attract disproportionate attention while carrying limited local information, but their role in generative models remains underexplored. We show that this phenomenon appears in both the encoder and denoiser of modern Representation Autoencoder (RAE)-DiT pipelines: pretrained ViT encoders can produce outlier representations, and DiTs themselves can develop internal outlier tokens, especially in intermediate layers. Moreover, simply masking high-norm tokens does not improve performance, indicating that the problem is not only caused by a few extreme values, but is more closely related to corrupted local patch semantics. To address this issue, we introduce Dual-Stage Registers (DSR), a register-based intervention for both components: trained registers when available, recursive test-time registers otherwise, and diffusion registers for the denoiser. Across ImageNet and large-scale text-to-image generation, these interventions consistently reduce outlier artifacts and improve generation quality. Our results highlight outlier-token control as an important ingredient in building stronger DiTs.

Xiaoyu Wu, Yifei Wang, Tsu-Jui Fu +3

Rice University · Apple

Diffusion TransformerActivation OutliersDiffusion-Based Image GenerationT2I Diffusion Models
AskPDF
May 6, 2026

Concurrence of Symmetry Breaking and Nonlocality Phase Transitions in Diffusion Models

Diffusion models undergo a phase transition in a critical time window during generation dynamics, with two complementary diagnoses of criticality. The symmetry breaking picture views the critical window as when trajectories bifurcate into different semantic minima of the energy landscape, whereas the nonlocality picture views the critical window as when local denoising fails. We study whether two notions of such phase transitions are concurrent in modern diffusion transformers. By evaluating the dynamics and outcomes of the generation trajectory, we observe a near-simultaneous occurrence of the non-locality and symmetry breaking critical times. Our work is the first to unify the two notions of phase transitions in practice: it provides a concrete diagnostic for when and why diffusion models rely on conditioning and global denoising, enabling principled evaluation of model efficiency and guiding the design of architectures and sampling schemes that avoid unnecessary computation.

Yifan F. Zhang, Fangjun Hu, Guangkuo Liu +2

Department of Electrical and Computer Engineering Princeton University Princeton, NJ 08544, USA · QuEra Computing Inc. 1284 Soldiers Field Road, Boston, MA 02135, USA · JILA and Department of Physics University of Colorado Boulder Boulder, CO 80309, USA +1

Diffusion TransformerSymmetry BreakingPhase TransitionsDiffusion Models
AskPDF
May 5, 2026

AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers

Representation alignment has recently emerged as an effective paradigm for accelerating Diffusion Transformer training. Despite their success, existing alignment methods typically impose a fixed supervision target or a fixed alignment granularity throughout the entire denoising trajectory, whether the guidance is provided by external vision encoders, internal self-representations, or VAE-derived features. We argue that such timestep-agnostic alignment is suboptimal because the useful granularity of representation supervision changes systematically with the signal-to-noise ratio. In high-noise regimes, diffusion models benefit more from coarse semantic and layout-level anchoring, whereas in low-noise regimes, the training signal should emphasize spatially detailed and structurally faithful refinement. This non-stationary alignment behavior creates a representational mismatch for static single-level supervisors. To address this issue, we propose Adaptive Hierarchical Prior Alignment (AHPA), a lightweight alignment framework that exploits the hierarchical representations naturally embedded in the frozen VAE encoder. Instead of using only a single compressed latent as the alignment target, AHPA extracts multi-level VAE features that provide complementary priors ranging from local geometry and spatial topology to coarse semantic layout. A timestep-conditioned Dynamic Router adaptively selects and weights these hierarchical priors along the denoising trajectory, thereby synchronizing the alignment granularity with the model's evolving training needs. Extensive experiments show that AHPA improves convergence and generation quality over baselines and incurs no additional inference cost while avoiding external encoder supervision during training.

Ruibin Min, Yexin Liu, Aimin Pan +5

Sun Yat-sen University · The Hong Kong University of Science and Technology · VNET Group +2

Diffusion Model AlignmentDiffusion TransformerDiffusion Model Acceleration
AskPDF
May 5, 2026

Ortho-Hydra: Orthogonalized Experts for DiT LoRA

LoRA fine-tuning of diffusion transformers (DiT) on multi-style data suffers from style bleed: a single low-rank residual cannot represent several distinct artist fingerprints, and the optimizer converges to their average. Mixture-of-experts LoRA in the HydraLoRA style replaces the up-projection with EE heads under a router, but when every expert is zero-initialized the router receives identical gradient from each head and remains at the uniform prior. The experts then evolve permutation-symmetrically, and the network trains as a single rank-rr LoRA at E×E{\times} the cost. We present Ortho-Hydra, a re-parameterisation that combines an OFT-style Cayley-orthogonal shared basis with per-expert disjoint output subspaces carved from the top-(Er)(Er) left singular vectors of the pretrained weight. Disjointness makes the router's per-expert score non-degenerate at step~00, so specialization receives gradient signal before any expert has trained. We test the predicted deadlock on a DiT pipeline by comparing two HydraLoRA baselines, a zero-initialized shared-basis variant and the original σ=0.1σ{=}0.1 Gaussian-jitter mitigation, against Ortho-Hydra under a matched optimiser, dataset, and step budget. Neither baseline leaves the uniform prior within the first 1k1\text{k} steps; Ortho-Hydra begins de-uniformising within the first few hundred. End-task generation quality on multi-style data is out of scope; we report the construction, the cold-start mechanism, and the routing dynamics it changes. Code: https://github.com/sorryhyun/anima_lora.

Seunghyun Ji

Fine-TuningDiffusion TransformerDiffusion Model Fine-TuningLow-Rank AdaptationMixture-of-Experts Models
AskPDF
May 4, 2026

Mamoda2.5: Enhancing Unified Multimodal Model with DiT-MoE

We present Mamoda2.5, a unified AR-Diffusion framework that seamlessly integrates multimodal understanding and generation within a single architecture. To efficiently enhance the model's generation capability, we equip the Diffusion Transformer backbone with a fine-grained Mixture-of-Experts (MoE) design (128 experts, Top-8 routing), yielding a 25B-parameter model that activates only 3B parameters, significantly reducing training costs while scaling up the model capacity. Mamoda2.5 achieves top-tier generation performance on VBench 2.0 and sets a new record in video editing quality, surpassing evaluated open-source models and matching the performance of current top-tier proprietary models, including the Kling O1 on OpenVE-Bench. Furthermore, we introduce a joint few-step distillation and reinforcement learning framework that compresses the 30-step editing model into a 4-step model and greatly accelerates model inference. Compared to open-source baselines, Mamoda2.5 achieves up to 95.9×95.9\times faster video editing inference. In real-world applications, Mamoda2.5 has been successfully deployed for content moderation and creative restoration tasks in advertising scenarios, achieving a 98% success rate in internal advertising video editing scenario.

Yangming Shi, Shixiang Zhu, Tao Shen +14

Mamoda Team, ByteDance

Diffusion Model Inference AccelerationUnified Multimodal ModelsDiffusion Model DistillationVideo EditingDiffusion TransformerMultimodal Foundation Models+2
AskPDF
May 2, 2026

AttnRouter: Per-Category Attention Routing for Training-Free Image Editing on MMDiT

We study training-free image editing on Qwen-Image-Edit-2511, a 60-block multi-modal diffusion transformer (MMDiT) that concatenates noise and source-image tokens within a single attention stream. We make three contributions. (i) We introduce KVInject, a single-forward attention manipulation that alpha-blends source-half key/value projections into the noise-half within a localized layer/step band. KVInject is simpler than the classical two-pass MasaCtrl recipe and avoids the prompt-mismatch failure mode that disables MasaCtrl on MMDiT (composite score drops 31% versus baseline). (ii) We show that no single attention operation dominates across edit types, motivating AttnRouter, a per-category routing table that dispatches edits to the operation that best preserves source structure for that type. With ground-truth categories the router improves the CLIP-T+DINO-I composite by 6.4% over the editing baseline; an automatic CLIP zero-shot classifier closes 98% of this gap despite only 55% category accuracy. (iii) Through layer-, step-, and alpha-band ablations we localize the editing-effective attention sub-circuit: K/V injection in early denoising steps (S0-7) recovers nearly all of the gain of full-step injection, while injection in early (L0-15) or late (L45-60) layer bands fails to drive editing entirely; alpha in [0.3, 0.5] is a stable sweet spot. We also report negative results that highlight what does not transfer from the UNet folklore: simple K/V rescaling never beats baseline and aggressive variants collapse generation entirely (composite 0.084). We release code, pre-computed routing tables, and a 100-sample stratified subset of ImgEdit-Bench used in all ablations.

Guandong Li, Mengxia Ye

iFLYTEK · Aegon THTF

Diffusion TransformerText-Guided Image EditingDiffusion-Based Image EditingTraining-Free Image Editing
AskPDF
May 1, 2026

Soft Graph Diffusion Transformer for MIMO Detection

Learning-based MIMO detection has shown strong empirical performance, yet existing methods typically rely on fixed-depth architectures without explicitly modeling the progressive refinement of symbol estimates. In this paper, we revisit MIMO detection from a flow matching perspective and propose the Soft Graph Diffusion Transformer (SGDiT), which reformulates detection as a noise-level-conditioned denoising process that progressively transforms a Gaussian initialization toward the posterior conditioned on channel observations. An adaptive layer normalization (AdaLN)-conditioned soft graph transformer is employed to parameterize the denoising dynamics, enabling stage-aware information integration between observation and symbol domains. To better align with the discrete nature of symbol detection, we further adopt a cross-entropy-based training objective that directly models bit-wise posterior probabilities, providing a more suitable inductive bias than conventional regression-based formulations. Experimental results across various MIMO system configurations demonstrate that SGDiT achieves competitive bit error rate (BER) performance compared with representative baselines. Furthermore, the proposed model exhibits good generalization capability across different channel conditions. Overall, the SGDiT framework provides an effective and practical approach for neural MIMO detection.

Nan Jiang, Jiadong Hong, Lei Liu +3

College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China · Theory Lab, Central Research Institute, 2012 Labs, Huawei Technologies Co., Ltd., Hong Kong, China

Diffusion TransformerWireless CommunicationsGraph Transformer
AskPDF
Apr 30, 2026

YOSE: You Only Select Essential Tokens for Efficient DiT-based Video Object Removal

Recent advances in Diffusion Transformer (DiT)-based video generation technologies have shown impressive results for video object removal. However, these methods still suffer from substantial inference latency. For instance, although MiniMax Remover achieves state-of-the-art visual quality, it operates at only around 10FPS, primarily due to dense computations over the entire spatiotemporal token space, even when only a small masked region actually requires processing. In this paper, we present YOSE, You Only Select Essential Tokens, an efficient fine-tuning framework. YOSE introduces two key components: Batch Variable-length Indexing (BVI) and Diffusion Process Simulator (DiffSim) Module. BVI is a differentiable dynamic indexing operator that adaptively selects essential tokens based on mask information, enabling variable-length token processing across samples. DiffSim provides a diffusion process approximation mechanism for unmasked tokens, which simulates the influence of unmasked regions within DiT self-attention to maintain semantic consistency for masked tokens. With these designs, YOSE achieves mask-aware acceleration, where the inference time scales approximately linearly with the masked regions, in contrast to full-token diffusion methods whose computation remains constant regardless of the mask size. Extensive experiments demonstrate that YOSE achieves up to 2.5X speedup in 70% of cases while maintaining visual quality comparable to the baseline. Code is available at: https://github.com/Wucy0519/YOSE-CVPR26.

Chenyang Wu, Lina Lei, Fan Li +6

1VCIP, CS, Nankai University · 3NKIARI, Shenzhen Futian · 2Huawei Noah’s Ark Lab

Diffusion TransformerVideo EditingVisual Token PruningDiffusion Model Inference Acceleration
AskPDF
Apr 29, 2026

Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models

To address the high sampling cost of Diffusion Transformers (DiTs), feature caching offers a training-free acceleration method. However, existing methods rely on hand-crafted forecasting formulas that fail under aggressive skipping. We propose L2P (Learnable Linear Predictor), a simple data-driven caching framework that replaces fixed coefficients with learnable per-timestep weights. Rapidly trained in ~20 seconds on a single GPU, L2P accurately reconstructs current features from past trajectories. L2P significantly outperforms existing baselines: it achieves a 4.55x FLOPs reduction and 4.15x latency speedup on FLUX.1-dev, and maintains high visual fidelity under up to 7.18x acceleration on Qwen-Image models, where prior methods show noticeable quality degradation. Our results show learning linear predictors is highly effective for efficient DiT inference. Code is available at https://github.com/Aredstone/L2P-Cache.

Zhirong Shen, Rui Huang, Jiacheng Liu +6

Shanghai Jiao Tong University · University of Electronic Science and Technology of China · Shandong University +2

Diffusion TransformerDiffusion Model CachingDiffusion Model Inference AccelerationDiffusion Models
AskPDF
Apr 28, 2026

The Thinking Pixel: Recursive Sparse Reasoning in Multimodal Diffusion Latents

Diffusion models have achieved success in high-fidelity data synthesis, yet their capacity for more complex, structured reasoning like text following tasks remains constrained. While advances in language models have leveraged strategies such as latent reasoning and recursion to enhance text understanding capabilities, extending these to multimodal text-to-image generation tasks is challenging due to the continuous and non-discrete nature of visual tokens. To tackle this problem, we draw inspiration from modular human cognition and propose a recursive, sparse mixture-of-experts framework integrated into conventional diffusion models. Our approach introduces a recursive component within joint attention layers that iteratively refines visual tokens over multiple latent steps while efficiently sharing parameters via sparse selection of neural modules. At each step, a gating network is devised to dynamically select specialized neural modules, conditioned on the current visual tokens, the diffusion timestep, and the conditioning information. Comprehensive evaluation on class-conditioned ImageNet image generation tasks and additional studies on the GenEval and DPG benchmark demonstrate the superiority of the proposed method in enhancing model image generation performance.

Yuwei Sun, Yuxuan Yao, Hui Li +1

Shanghai Academy of AI for Science · Fudan University

Diffusion TransformerConditional Image GenerationSparse Mixture-of-ExpertsDiffusion ModelsMultimodal Diffusion Models
AskPDF
Apr 26, 2026

Edit Where You Mean: Region-Aware Adapter Injection for Mask-Free Local Image Editing

Large diffusion transformers (DiTs) follow global editing instructions well but consistently leak local edits into unrelated regions, because joint-attention architectures offer no explicit channel telling the network where to apply the edit. We introduce AdaptEdit, a co-trained, instruction- and region-aware adapter framework that retro-fits a frozen DiT into a precise local editor without modifying its backbone weights. A lightweight Block Adapter at every transformer block injects a structured condition stream that factorizes what to edit (instruction semantics) from where to edit (spatial mask); a learned SpatialGate routes the adapter signal selectively into the edit region while keeping the rest of the image near-identical to the source; and a Region-Aware Loss focuses the training objective on the changing pixels. Because these components make the backbone's internal representation mask-aware end-to-end, a thin MaskPredictor head trained jointly with the editor can ground the edit region directly from the instruction and source image -- eliminating any user-mask requirement at deployment. We evaluate on two complementary benchmarks: MagicBrush (paired ground-truth targets) to measure pixel-level preservation and edit accuracy, and Emu-Edit Test (no ground-truth images, 9 diverse edit categories) to stress-test instruction following and generalization across edit types. On both, AdaptEdit achieves state-of-the-art results, simultaneously outperforming mask-free and oracle-mask baselines. A seven-variant ablation cleanly isolates the contribution of each component.

Honghao Cai, Xiangyuan Wang, Yunhao Bai +8

The Chinese University of Hong Kong, Shenzhen · Xiaohongshu Inc. · Peking University +2

Diffusion TransformerText-Guided Image EditingDiffusion Model Fine-TuningImage EditingParameter-Efficient Fine-Tuning
AskPDF
Apr 26, 2026

Talker-T2AV: Joint Talking Audio-Video Generation with Autoregressive Diffusion Modeling

Joint audio-video generation models have shown that unified generation yields stronger cross-modal coherence than cascaded approaches. However, existing models couple modalities throughout denoising via pervasive attention, treating high-level semantics and low-level details in a fully entangled manner. This is suboptimal for talking head synthesis: while audio and facial motion are semantically correlated, their low-level realizations (acoustic signals and visual textures) follow distinct rendering processes. Enforcing joint modeling across all levels causes unnecessary entanglement and reduces efficiency. We propose Talker-T2AV, an autoregressive diffusion framework where high-level cross-modal modeling occurs in a shared backbone, while low-level refinement uses modality-specific decoders. A shared autoregressive language model jointly reasons over audio and video in a unified patch-level token space. Two lightweight diffusion transformer heads decode the hidden states into frame-level audio and video latents. Experiments on talking portrait benchmarks show Talker-T2AV outperforms dual-branch baselines in lip-sync accuracy, video quality, and audio quality, achieving stronger cross-modal consistency than cascaded pipelines.

Zhen Ye, Xu Tan, Aoxiong Yin +8

Hong Kong University of Science and Technology · Independent Researcher · Zhejiang University +3

Talking Head GenerationAudio-Video GenerationDiffusion TransformerCross-Modal LearningMultimodal Diffusion ModelsAutoregressive Video Generation+1
AskPDF
Apr 24, 2026

OccDirector: Language-Guided Behavior and Interaction Generation in 4D Occupancy Space

Generative world models increasingly rely on 4D occupancy for realistic autonomous driving simulation. However, existing generation frameworks depend on rigid geometric conditions (e.g., explicit trajectories) or simplistic attribute-level text, failing to orchestrate complex, sequential multi-agent interactions. To address this semantic-spatiotemporal gap, we propose OccDirector, a pioneering framework that generates 4D occupancy dynamics conditioned solely on natural language. Operating as a ``scenario director'', OccDirector maps natural language scripts into physically plausible voxel dynamics without requiring geometric priors. Technically, it employs a VLM-driven Spatio-Temporal MMDiT equipped with a history-prefix anchoring strategy to ensure long-horizon interaction consistency. Furthermore, we introduce OccInteract-85k, a novel dataset uniquely annotated with multi-level language instructions: ranging from static layouts to intricate multi-agent behaviors, alongside a novel VLM-based evaluation benchmark. Extensive experiments demonstrate that OccDirector achieves state-of-the-art generation quality and unprecedented instruction-following capabilities, successfully shifting the paradigm from appearance synthesis to language-driven behavior orchestration.

Zhuding Liang, Tianyi Yan, Dubing Chen +6

SKL-IOTSC, CIS, University of Macau, Macau, China · Li Auto Inc, Beijing, China

Diffusion TransformerMulti-Agent SystemsConditional GenerationAutonomous Driving SimulationAutonomous Driving Scenario Generation4D Scene Generation
AskPDF
Apr 24, 2026

UniSonate: A Unified Model for Speech, Music, and Sound Effect Generation with Text Instructions

Generative audio modeling has largely been fragmented into specialized tasks, text-to-speech (TTS), text-to-music (TTM), and text-to-audio (TTA), each operating under heterogeneous control paradigms. Unifying these modalities remains a fundamental challenge due to the intrinsic dissonance between structured semantic representations (speech/music) and unstructured acoustic textures (sound effects). In this paper, we introduce UniSonate, a unified flow-matching framework capable of synthesizing speech, music, and sound effects through a standardized, reference-free natural language instruction interface. To reconcile structural disparities, we propose a novel dynamic token injection mechanism that projects unstructured environmental sounds into a structured temporal latent space, enabling precise duration control within a phoneme-driven Multimodal Diffusion Transformer (MM-DiT). Coupled with a multi-stage curriculum learning strategy, this approach effectively mitigates cross-modal optimization conflicts. Extensive experiments demonstrate that UniSonate achieves state-of-the-art performance in instruction-based TTS (WER 1.47%) and TTM (SongEval Coherence 3.18), while maintaining competitive fidelity in TTA. Crucially, we observe positive transfer, where joint training on diverse audio data significantly enhances structural coherence and prosodic expressiveness compared to single-task baselines. Audio samples are available at https://qiangchunyu.github.io/UniSonate/.

Chunyu Qiang, Xiaopeng Wang, Kang Yin +11

1Tianjin University, Tianjin, China · 2Kling Team, Kuaishou Technology, Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China

Text-to-Music GenerationText-to-Audio GenerationTTS SynthesisDiffusion TransformerSound Effects GenerationAudio Generation
AskPDF
Apr 23, 2026

Sculpt4D: Generating 4D Shapes via Sparse-Attention Diffusion Transformers

Recent breakthroughs in 3D generative modeling have yielded remarkable progress in static shape synthesis, yet high-fidelity dynamic 4D generation remains elusive, hindered by temporal artifacts and prohibitive computational demand. We present Sculpt4D, a native 4D generative framework that seamlessly integrates efficient temporal modeling into a pretrained 3D Diffusion Transformer (Hunyuan3D 2.1), thereby mitigating the scarcity of 4D training data. At its core lies a Block Sparse Attention mechanism that preserves object identity by anchoring to the initial frame while capturing rich motion dynamics via a time-decaying sparse mask. This design faithfully models complex spatiotemporal dependencies with high fidelity, while sidestepping the quadratic overhead of full attention and reducing network total computation by 56%. Consequently, Sculpt4D establishes a new state-of-the-art in temporally coherent 4D synthesis and charts a path toward efficient and scalable 4D generation.

Minghao Yin, Wenbo Hu, Jiale Xu +2

1The University of Hong Kong · 2ARC Lab, Tencent PCG

Diffusion Transformer3D Shape GenerationSparse Attention
AskPDF