Diffusion Transformer

Latent diffusion models (LDMs) have recently achieved strong performance in 3D medical image synthesis. However, modalities like cine cardiac MRI (CMR), representing a temporally synchronized 3D volume across the cardiac cycle, add an additional dimension that most generative approaches do not model directly. Instead, they factorize space and time or enforce temporal consistency through auxiliary mechanisms such as anatomical masks. Such strategies introduce structural biases that may limit global context integration and lead to subtle spatiotemporal discontinuities or physiologically inconsistent cardiac dynamics. We investigate whether a unified 4D generative model can learn continuous cardiac dynamics without architectural factorization. We propose CardioDiT, a fully 4D latent diffusion framework for short-axis cine CMR synthesis based on diffusion transformers. A spatiotemporal VQ-VAE encodes 2D+t slices into compact latents, which a diffusion transformer then models jointly as complete 3D+t volumes, coupling space and time throughout the generative process. We evaluate CardioDiT on public CMR datasets and a larger private cohort, comparing it to baselines with progressively stronger spatiotemporal coupling. Results show improved inter-slice consistency, temporally coherent motion, and realistic cardiac function distributions, suggesting that explicit 4D modeling with a diffusion transformer provides a principled foundation for spatiotemporal cardiac image synthesis. Code and models trained on public data are available at https://github.com/Cardio-AI/cardiodit.

Diffusion models have become a leading approach for high-fidelity medical image synthesis. However, most existing methods for 3D medical image generation rely on convolutional U-Net backbones within latent diffusion frameworks. While effective, these architectures impose strong locality biases and limited receptive fields, which may constrain scalability, global context integration, and flexible conditioning. In this work, we introduce VolDiT, the first purely transformer-based 3D Diffusion Transformer for volumetric medical image synthesis. Our approach extends diffusion transformers to native 3D data through volumetric patch embeddings and global self-attention operating directly over 3D tokens. To enable structured control, we propose a timestep-gated control adapter that maps segmentation masks into learnable control tokens that modulate transformer layers during denoising. This token-level conditioning mechanism allows precise spatial guidance while preserving the modeling advantages of transformer architectures. We evaluate our model on high-resolution 3D medical image synthesis tasks and compare it to state-of-the-art 3D latent diffusion models based on U-Nets. Results demonstrate improved global coherence, superior generative fidelity, and enhanced controllability. Our findings suggest that fully transformerbased diffusion models provide a flexible foundation for volumetric medical image synthesis. The code and models trained on public data are available at https://github.com/Cardio-AI/voldit.

Concept erasure serves as a vital safety mechanism for removing unwanted concepts from text-to-image (T2I) models. While extensively studied in U-Net and dual-stream architectures (e.g., Flux), this task remains under-explored in the recent emerging paradigm of single-stream diffusion transformers (e.g., Z-Image). In this new paradigm, text and image tokens are processed as a single unified sequence via shared parameters. Consequently, directly applying prior erasure methods typically leads to generation collapse. To bridge this gap, we introduce Z-Erase, the first concept erasure method tailored for single-stream T2I models. To guarantee stable image generation, Z-Erase first proposes a Stream Disentangled Concept Erasure Framework that decouples updates and enables existing methods on single-stream models. Subsequently, within this framework, we introduce Lagrangian-Guided Adaptive Erasure Modulation, a constrained algorithm that further balances the sensitive erasure-preservation trade-off. Moreover, we provide a rigorous convergence analysis proving that Z-Erase can converge to a Pareto stationary point. Experiments demonstrate that Z-Erase successfully overcomes the generation collapse issue, achieving state-of-the-art performance across a wide range of tasks.

In this paper, we uncover the hidden potential of Diffusion Transformers (DiTs) to significantly enhance generative tasks. Through an in-depth analysis of the denoising process, we demonstrate that introducing a single learned scaling parameter can significantly improve the performance of DiT blocks. Building on this insight, we propose Calibri, a parameter-efficient approach that optimally calibrates DiT components to elevate generative quality. Calibri frames DiT calibration as a black-box reward optimization problem, which is efficiently solved using an evolutionary algorithm and modifies just ~100 parameters. Experimental results reveal that despite its lightweight design, Calibri consistently improves performance across various text-to-image models. Notably, Calibri also reduces the inference steps required for image generation, all while maintaining high-quality outputs.

Diffusion models have significantly reshaped the field of generative artificial intelligence and are now increasingly explored for their capacity in discriminative representation learning. Diffusion Transformer (DiT) has recently gained attention as a promising alternative to conventional U-Net-based diffusion models, demonstrating a promising avenue for downstream discriminative tasks via generative pre-training. However, its current training efficiency and representational capacity remain largely constrained due to the inadequate timestep searching and insufficient exploitation of DiT-specific feature representations. In light of this view, we introduce Automatically Selected Timestep (A-SelecT) that dynamically pinpoints DiT's most information-rich timestep from the selected transformer feature in a single run, eliminating the need for both computationally intensive exhaustive timestep searching and suboptimal discriminative feature selection. Extensive experiments on classification and segmentation benchmarks demonstrate that DiT, empowered by A-SelecT, surpasses all prior diffusion-based attempts efficiently and effectively.

Diffusion-based video editing has emerged as an important paradigm for high-quality and flexible content generation. However, despite their generality and strong modeling capacity, Diffusion Transformers (DiT) remain computationally expensive due to the iterative denoising process, posing challenges for practical deployment. Existing video diffusion acceleration methods primarily exploit denoising timestep-level feature reuse, which mitigates the redundancy in denoising process, but overlooks the architectural redundancy within the DiT that many attention operations over spatio-temporal tokens are redundantly executed, offering little to no incremental contribution to the model output. This work introduces HetCache, a training-free diffusion acceleration framework designed to exploit the inherent heterogeneity in diffusion-based masked video-to-video (MV2V) generation and editing. Instead of uniformly reuse or randomly sampling tokens, HetCache assesses the contextual relevance and interaction strength among various types of tokens in designated computing steps. Guided by spatial priors, it divides the spatial-temporal tokens in DiT model into context and generative tokens, and selectively caches the context tokens that exhibit the strongest correlation and most representative semantics with generative ones. This strategy reduces redundant attention operations while maintaining editing consistency and fidelity. Experiments show that HetCache achieves a noticeable acceleration, including a 2.67$\times$ latency speedup and FLOPs reduction over commonly used foundation models, with negligible degradation in editing quality.

Recent approaches for segmentation have leveraged pretrained generative models as feature extractors, treating segmentation as a downstream adaptation task via indirect feature retrieval. This implicit use suffers from a fundamental misalignment in representation. It also depends heavily on indirect feature extraction pipelines, which complicate the workflow and limit adaptation. In this paper, we argue that instead of indirect adaptation, segmentation tasks should be trained directly in a generative manner. We identify a key obstacle to this unified formulation: VAE latents of binary masks are sharply distributed, noise robust, and linearly separable, distinct from natural image latents. To bridge this gap, we introduce timesteps sampling strategy for binary masks that emphasizes extreme noise levels for segmentation and moderate noise for image generation, enabling harmonious joint training. We present GenMask, a DiT trains to generate black-and-white segmentation masks as well as colorful images in RGB space under the original generative objective. GenMask preserves the original DiT architecture while removing the need of feature extraction pipelines tailored for segmentation tasks. Empirically, GenMask attains state-of-the-art performance on referring and reasoning segmentation benchmarks and ablations quantify the contribution of each component.

Transformer-based video diffusion models rely on 3D attention over spatial and temporal tokens, which incurs quadratic time and memory complexity and makes end-to-end training for ultra-high-resolution videos prohibitively expensive. To overcome this bottleneck, we propose a pure image adaptation framework that upgrades a video Diffusion Transformer pre-trained at its native scale to synthesize higher-resolution videos. Unfortunately, naively fine-tuning with high-resolution images alone often introduces noticeable noise due to the image-video modality gap. To address this, we decouple the learning objective to separately handle modality alignment and spatial extrapolation. At the core of our approach is Relay LoRA, a two-stage adaptation strategy. In the first stage, the video diffusion model is adapted to the image domain using low-resolution images to bridge the modality gap. In the second stage, the model is further adapted with high-resolution images to acquire spatial extrapolation capability. During inference, only the high-resolution adaptation is retained to preserve the video generation modality while enabling high-resolution video synthesis. To enhance fine-grained detail synthesis, we further propose a High-Frequency-Awareness-Training-Objective, which explicitly encourages the model to recover high-frequency components from degraded latent representations via a dedicated reconstruction loss. Extensive experiments demonstrate that our method produces ultra-high-resolution videos with rich visual details without requiring any video training data, even outperforming previous state-of-the-art models trained on high-resolution videos by 0.8 on the VBench benchmark. Code will be available at https://github.com/WillWu111/ViBe.

Diffusion Transformers (DiTs) excel at generation, but their global self-attention makes controllable, reference-image-based editing a distinct challenge. Unlike U-Nets, naively injecting local appearance into a DiT can disrupt its holistic scene structure. We address this by proposing the first training-free framework specifically designed to tame DiTs for high-fidelity appearance transfer. Our core is a synergistic system that disentangles structure and appearance. We leverage high-fidelity inversion to establish a rich content prior for the source image, capturing its lighting and micro-textures. A novel attention-sharing mechanism then dynamically fuses purified appearance features from a reference, guided by geometric priors. Our unified approach operates at 1024px and outperforms specialized methods on tasks ranging from semantic attribute transfer to fine-grained material application. Extensive experiments confirm our state-of-the-art performance in both structural preservation and appearance fidelity.

Text-to-image synthesis for remote sensing (RS) lacks an accessible, high-performance generative foundation, as directly training diffusion models at large, high resolutions is computationally prohibitive. Training-free resolution promotion via Rotary Position Embedding (RoPE) extrapolation offers an efficient alternative, but existing algorithms apply static scaling rules tailored to natural scenes, whereas RS imagery is dominated by dense, minute instances that hinge on high-frequency structural integrity. We present a comprehensive framework for large-scale RS synthesis. First, we curate over 100{,}000 image-text pairs to train RS-FLUX, a domain-specialized generative prior. Second, we propose SHARP, a training-free extrapolation algorithm that introduces a Rational Decay Scheduler to modulate RoPE frequencies throughout denoising: strong positional extrapolation early on enforces coherent global layouts, and progressive relaxation later recovers dense high-frequency details. Extensive experiments show that SHARP consistently achieves state-of-the-art performance across multiple promoted resolutions with negligible ($<$4%) overhead, offering an efficient and structurally faithful solution for large-scale RS generation.

Recent advances in generative models, such as diffusion and flow matching, have shown strong performance in audio tasks. However, speech enhancement (SE) models are typically trained on limited datasets and evaluated under narrow conditions, limiting real-world applicability. To address this, we propose DiT-Flow, a flow matching-based SE framework built on the latent Diffusion Transformer (DiT) backbone and trained for robustness across diverse distortions, including noise, reverberation, and compression. DiT-Flow operates on compact variational auto-encoders (VAEs)-derived latent features. We validated our approach on StillSonicSet, a synthetic yet acoustically realistic dataset composed of LibriSpeech, FSD50K, FMA, and 90 Matterport3D scenes. Experiments show that DiT-Flow consistently outperforms state-of-the-art generative SE models, demonstrating the effectiveness of flow matching in multi-condition speech enhancement. Despite ongoing efforts to expand synthetic data realism, a persistent bottleneck in SE is the inevitable mismatch between training and deployment conditions. By integrating LoRA with the MoE framework, we achieve both parameter-efficient and high-performance training for DiT-Flow robust to multiple distortions with using 4.9% percentage of the total parameters to obtain a better performance on five unseen distortions.

Recent theoretical models of diffusion processes, conceptualized as coupled Ornstein-Uhlenbeck systems, predict a hierarchy of interaction timescales, and consequently, the existence of a synchronization gap between modes that commit at different stages of the reverse process. However, because these predictions rely on continuous time and analytically tractable score functions, it remains unclear how this phenomenology manifests in the deep, discrete architectures deployed in practice. In this work, we investigate how the synchronization gap is mechanistically realized within pretrained Diffusion Transformers (DiTs). We construct an explicit architectural realization of replica coupling by embedding two generative trajectories into a joint token sequence, modulated by a symmetric cross attention gate with variable coupling strength g. Through a linearized analysis of the attention difference, we show that the replica interaction decomposes mechanistically. We empirically validate our theoretical framework on a pretrained DiT-XL/2 model by tracking commitment and per layer internal mode energies. Our results reveal that: (1) the synchronization gap is an intrinsic architectural property of DiTs that persists even when external coupling is turned off; (2) as predicted by our spatial routing bounds, the gap completely collapses under strong coupling; (3) the gap is strictly depth localized, emerging sharply only within the final layers of the Transformer; and (4) global, low frequency structures consistently commit before local, high frequency details. Ultimately, our findings provide a mechanistic interpretation of how Diffusion Transformers resolve generative ambiguity, isolating speciation transitions to the terminal layers of the network.

Diffusion Transformers (DiTs) have significantly enhanced text-to-image (T2I) generation quality, enabling high-quality personalized content creation. However, fine-tuning these models requires substantial computational complexity and memory, limiting practical deployment under resource constraints. To tackle these challenges, we propose a memory-efficient fine-tuning framework called DiT-BlockSkip, integrating timestep-aware dynamic patch sampling and block skipping by precomputing residual features. Our dynamic patch sampling strategy adjusts patch sizes based on the diffusion timestep, then resizes the cropped patches to a fixed lower resolution. This approach reduces forward & backward memory usage while allowing the model to capture global structures at higher timesteps and fine-grained details at lower timesteps. The block skipping mechanism selectively fine-tunes essential transformer blocks and precomputes residual features for the skipped blocks, significantly reducing training memory. To identify vital blocks for personalization, we introduce a block selection strategy based on cross-attention masking. Evaluations demonstrate that our approach achieves competitive personalization performance qualitatively and quantitatively, while reducing memory usage substantially, moving toward on-device feasibility (e.g., smartphones, IoT devices) for large-scale diffusion transformers.

Millimeter-wave (mmWave) radar enables contactless respiratory sensing,yet fine-grained monitoring is often degraded by nonstationary interference from body micromotions.To achieve micromotion interference removal,we propose mmWave-Diffusion,an observation-anchored conditional diffusion framework that directly models the residual between radar phase observations and the respiratory ground truth,and initializes sampling within an observation-consistent neighborhood rather than from Gaussian noise-thereby aligning the generative process with the measurement physics and reducing inference overhead. The accompanying Radar Diffusion Transformer (RDT) is explicitly conditioned on phase observations, enforces strict one-to-one temporal alignment via patch-level dual positional encodings, and injects local physical priors through banded-mask multi-head cross-attention, enabling robust denoising and interference removal in just 20 reverse steps. Evaluated on 13.25 hours of synchronized radar-respiration data, mmWave-Diffusion achieves state-of-the-art waveform reconstruction and respiratory-rate estimation with strong generalization. Code repository:https://github.com/goodluckyongw/mmWave-Diffusion.

Despite the rapid advancement of Virtual Try-On (VTON) and Try-Off (VTOFF) technologies, existing VTON methods face challenges with fine-grained detail preservation, generalization to complex scenes, complicated pipeline, and efficient inference. To tackle these problems, we propose OmniDiT, an omni Virtual Try-On framework based on the Diffusion Transformer, which combines try-on and try-off tasks into one unified model. Specifically, we first establish a self-evolving data curation pipeline to continuously produce data, and construct a large VTON dataset Omni-TryOn, which contains over 380k diverse and high-quality garment-model-tryon image pairs and detailed text prompts. Then, we employ the token concatenation and design an adaptive position encoding to effectively incorporate multiple reference conditions. To relieve the bottleneck of long sequence computation, we are the first to introduce Shifted Window Attention into the diffusion model, thus achieving a linear complexity. To remedy the performance degradation caused by local window attention, we utilize multiple timestep prediction and an alignment loss to improve generation fidelity. Experiments reveal that, under various complex scenes, our method achieves the best performance in both the model-free VTON and VTOFF tasks and a performance comparable to current SOTA methods in the model-based VTON task.

Diffusion Transformers (DiTs) achieve strong video generation quality but suffer from high inference cost due to dense 3D attention, motivating sparse attention techniques for improving efficiency. However, existing training-free sparse attention methods for video generation still face two unresolved limitations: ignoring layer heterogeneity in attention pruning and ignoring query-key coupling in block partitioning, which hinder a better quality-speedup trade-off. In this work, we uncover a critical insight: attention sparsity is an intrinsic layer-wise property, with only minor variation across different inputs. Motivated by this observation, we propose SVOO, a training-free sparse attention framework for fast video generation via offline layer-wise sparsity profiling and online bidirectional co-clustering. Specifically, SVOO adopts a two-stage paradigm: (i) offline layer-wise sensitivity profiling to derive intrinsic per-layer pruning levels, and (ii) online block-wise sparse attention via a bidirectional co-clustering algorithm. Extensive experiments on seven widely used video generation models demonstrate that SVOO achieves a superior quality-speedup trade-off over state-of-the-art methods, delivering up to 1.93x speedup while maintaining a PSNR of up to 29 dB on Wan2.1. Code is available at: https://github.com/Mutual-Luo/SVOO.

In this work, we study the role of Massive Activations (MAs), which are rare, high-magnitude spikes confined to a few fixed hidden dimensions in video diffusion transformers (DiTs). We uncover a structured positional hierarchy: MA magnitudes peak at first-frame tokens and recur at the spatial boundary tokens of latent frames, with this pattern being most pronounced during early denoising. We trace this organization to an encoding asymmetry of the video VAEs, whose causal temporal padding and zero spatial padding cause the first latent frame and frame borders to carry reduced content load. Elevated MAs consistently align with these lower-content structural positions. To understand their function, we analyze intermediate representations and find that MAs act as implicit rescalers of residual computation: enlarging MAs suppresses the corresponding self-attention and feed-forward updates, while erasing them amplifies these updates. Together, these observations suggest that MAs serves as a token-level rescaler of residual computation, which video DiTs deploy unevenly, placing the strongest damping at the encoding-asymmetric structural positions. Motivated by this native rescaling behavior, we propose Structured Activation Steering (STAS), a training-free technique that steers MAs at the observed structural positions toward a scaled, model-derived reference during early denoising. STAS requires no additional forward passes and consistently improves video quality and temporal coherence across text-to-video models with negligible overhead.

Recent advances in video generation have revealed an unexpected phenomenon: diffusion-based video models exhibit non-trivial reasoning capabilities. Prior work attributes this to a Chain-of-Frames (CoF) mechanism, where reasoning is assumed to unfold sequentially across video frames. In this work, we challenge this assumption and uncover a fundamentally different mechanism. We show that reasoning in video models instead primarily emerges along the diffusion denoising steps. Through qualitative analysis and targeted probing experiments, we find that models explore multiple candidate solutions in early denoising steps and progressively converge to a final answer, a process we term Chain-of-Steps (CoS). Beyond this core mechanism, we identify several emergent reasoning behaviors critical to model performance: (1) working memory that supports tasks requiring consistent reference, such as object permanence; (2) self-correction and enhancement, allowing recovery from incorrect intermediate solutions; and (3) perception before action, where early steps establish semantic grounding and later steps perform structured manipulation. Moreover, analysis of Diffusion Transformer layers shows that middle layers conduct key reasoning procedures. Motivated by these insights, we present a simple Training-Free Ensemble (TFE) as a proof-of-concept, demonstrating how reasoning can be improved by ensembling latent trajectories from identical models with different random seeds. Overall, our work provides the first systematic dissection of the mechanisms underlying video reasoning, offering a foundation to guide future research in better exploiting the inherent reasoning dynamics of video models as a new substrate for intelligence.

Diffusion Transformers (DiTs) have demonstrated remarkable scalability and quality in image and video generation, prompting growing interest in extending them to controllable generation and editing tasks. However, compared to the image counterparts, progress in video control and editing remains limited, mainly due to the scarcity of paired video data and the high computational cost of training video diffusion models. To address this issue, in this paper, we propose a video-free tuning framework termed ViFeEdit for video diffusion transformers. Without requiring any forms of video training data, ViFeEdit achieves versatile video generation and editing, adapted solely with 2D images. At the core of our approach is an architectural reparameterization that decouples spatial independence from the full 3D attention in modern video diffusion transformers, which enables visually faithful editing while maintaining temporal consistency with only minimal additional parameters. Moreover, this design operates in a dual-path pipeline with separate timestep embeddings for noise scheduling, exhibiting strong adaptability to diverse conditioning signals. Extensive experiments demonstrate that our method delivers promising results of controllable video generation and editing with only minimal training on 2D image data. Codes are available https://github.com/Lexie-YU/ViFeEdit.

While recent Flow Matching models avoid the reconstruction bottlenecks of latent autoencoders by operating directly in pixel space, the raw pixel manifold provides little explicit semantic organization, making target-specific transport directions difficult for a shared finite-capacity vector field to distinguish. When different transport requirements are locally entangled in this way, their learning signals can interfere and hinder optimization, a phenomenon we refer to as trajectory conflict. To alleviate trajectory conflict while preserving direct generation in pixel space, we propose Waypoint Diffusion Transformers (WiT), which introduces explicit semantic routing into pixel-space generation. WiT structures pixel-space prediction through compact semantic waypoints projected from pre-trained vision representations, providing a semantically organized intermediate representation that complements direct pixel prediction. During ODE integration, a lightweight Waypoints Predictor dynamically infers the semantic waypoint from the current noisy state, and the predicted waypoint continuously conditions the primary Pixel Space Generator through our Just-Pixel AdaLN mechanism. This spatially varying semantic modulation reorganizes the finite-capacity learning problem and helps the model preserve target-specific transport directions while retaining generation entirely in pixel space. Experiments on ImageNet 2012 demonstrate consistent improvements over strong pixel-space baselines across model scales, with WiT-H/16 achieving an FID of 1.79. Code will be released at https://hainuo-wang.github.io/WiT.

Representation Alignment (REPA) has emerged as a simple way to accelerate Diffusion Transformers training in latent space. At the same time, pixel-space diffusion transformers such as Just image Transformers (JiT) have attracted growing attention because they remove a dependency on a pretrained tokenizer, and then avoid the reconstruction bottleneck of latent diffusion. This paper shows that the REPA can fail for JiT. REPA yields worse FID for JiT as training proceeds and collapses diversity on image subsets that are tightly clustered in the representation space of pretrained semantic encoder on ImageNet. We trace the failure to an information asymmetry: denoising occurs in the high dimensional image space, while the semantic target is strongly compressed, making direct regression a shortcut objective. We propose PixelREPA, which transforms the alignment target and constrains alignment with a Masked Transformer Adapter that combines a shallow transformer adapter with partial token masking. PixelREPA improves both training convergence and final quality. PixelREPA reduces FID from 3.66 to 3.17 for JiT-B$/16$ and improves Inception Score (IS) from 275.1 to 284.6 on ImageNet $256 \times 256$, while achieving $> 2\times$ faster convergence. Finally, PixelREPA-H$/16$ achieves FID$=1.81$ and IS$=317.2$. Our code is available at https://github.com/kaist-cvml/PixelREPA.

Scene Text Image Super-Resolution (STISR) aims to restore high-resolution details in low-resolution text images, which is crucial for both human readability and machine recognition. Existing methods, however, often depend on external Optical Character Recognition (OCR) models for textual priors or rely on complex multi-component architectures that are difficult to train and reproduce. In this paper, we introduce DualTSR, a unified end-to-end framework that addresses both issues. DualTSR employs a single multimodal transformer backbone trained with a dual diffusion objective. It simultaneously models the continuous distribution of high-resolution images via Conditional Flow Matching and the discrete distribution of textual content via discrete diffusion. This shared design enables visual and textual information to interact at every layer, allowing the model to infer text priors internally instead of relying on an external OCR module. Compared with prior multi-branch diffusion systems, DualTSR offers a simpler end-to-end formulation with fewer hand-crafted components. Experiments on synthetic Chinese benchmarks and a curated real-world evaluation protocol show that DualTSR achieves strong perceptual quality and text fidelity.

The success of vision transformers-especially for generative modeling-is limited by the quadratic cost and weak spatial inductive bias of self-attention. We propose PDE-SSM, a spatial state-space block that replaces attention with a learnable convection-diffusion-reaction partial differential equation. This operator encodes a strong spatial prior by modeling information flow via physically grounded dynamics rather than all-to-all token interactions. Solving the PDE in the Fourier domain yields global coupling with near-linear complexity of $O(N \log N)$, delivering a principled and scalable alternative to attention. We integrate PDE-SSM into a flow-matching generative model to obtain the PDE-based Diffusion Transformer PDE-SSM-DiT. Empirically, PDE-SSM-DiT matches or exceeds the performance of state-of-the-art Diffusion Transformers while substantially reducing compute. Our results show that, analogous to 1D settings where SSMs supplant attention, multi-dimensional PDE operators provide an efficient, inductive-bias-rich foundation for next-generation vision models.

Imbalanced datasets pose a difficulty in fraud detection, as classifiers are often biased toward the majority class and perform poorly on rare fraudulent transactions. Synthetic data generation is therefore commonly used to mitigate this problem. In this work, we propose the Clustered Embedding Diffusion-Transformer (EmDT), a diffusion model designed to generate fraudulent samples. Our key innovation is to leverage UMAP clustering to identify distinct fraudulent patterns, and train a Transformer denoising network with sinusoidal positional embeddings to capture feature relationships throughout the diffusion process. Once the synthetic data has been generated, we employ a standard decision-tree-based classifier (e.g., XGBoost) for classification, as this type of model remains better suited to tabular datasets. Experiments on a credit card fraud detection dataset demonstrate that EmDT significantly improves downstream classification performance compared to existing oversampling and generative methods, while maintaining comparable privacy protection and preserving feature correlations present in the original data.

Diffusion-based image compression has recently shown outstanding perceptual fidelity, yet its practicality is hindered by prohibitive sampling overhead and high memory usage. Most existing diffusion codecs employ U-Net architectures, where hierarchical downsampling forces diffusion to operate in shallow latent spaces (typically with only 8x spatial downscaling), resulting in excessive computation. In contrast, conventional VAE-based codecs work in much deeper latent domains (16x - 64x downscaled), motivating a key question: Can diffusion operate effectively in such compact latent spaces without compromising reconstruction quality? To address this, we introduce DiT-IC, an Aligned Diffusion Transformer for Image Compression, which replaces the U-Net with a Diffusion Transformer capable of performing diffusion in latent space entirely at 32x downscaled resolution. DiT-IC adapts a pretrained text-to-image multi-step DiT into a single-step reconstruction model through three key alignment mechanisms: (1) a variance-guided reconstruction flow that adapts denoising strength to latent uncertainty for efficient reconstruction; (2) a self-distillation alignment that enforces consistency with encoder-defined latent geometry to enable one-step diffusion; and (3) a latent-conditioned guidance that replaces text prompts with semantically aligned latent conditions, enabling text-free inference. With these designs, DiT-IC achieves state-of-the-art perceptual quality while offering up to 30x faster decoding and drastically lower memory usage than existing diffusion-based codecs. Remarkably, it can reconstruct 2048x2048 images on a 16 GB laptop GPU.

Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11$\times$ speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.

Recently, Multimodal Large Language Models (MLLMs) have been widely integrated into diffusion frameworks primarily as text encoders to tackle complex tasks such as spatial reasoning. However, this paradigm suffers from two critical limitations: (i) MLLMs text encoder exhibits insufficient reasoning depth. Single-step encoding fails to activate the Chain-of-Thought process, which is essential for MLLMs to provide accurate guidance for complex tasks. (ii) The guidance remains invariant during the decoding process. Invariant guidance during decoding prevents DiT from progressively decomposing complex instructions into actionable denoising steps, even with correct MLLM encodings. To this end, we propose Endogenous Chain-of-Thought (EndoCoT), a novel framework that first activates MLLMs' reasoning potential by iteratively refining latent thought states through an iterative thought guidance module, and then bridges these states to the DiT's denoising process. Second, a terminal thought grounding module is applied to ensure the reasoning trajectory remains grounded in textual supervision by aligning the final state with ground-truth answers. With these two components, the MLLM text encoder delivers meticulously reasoned guidance, enabling the DiT to execute it progressively and ultimately solve complex tasks in a step-by-step manner. Extensive evaluations across diverse benchmarks (e.g., Maze, TSP, VSP, and Sudoku) achieve an average accuracy of 92.1%, outperforming the strongest baseline by 8.3 percentage points. The code and dataset are publicly available at https://internlm.github.io/EndoCoT/.

Diffusion transformers (DiTs) achieve high generative quality but lock FLOPs to image resolution, limiting principled latency-quality trade-offs, and allocate computation uniformly across input spatial tokens, wasting resource allocation to unimportant regions. We introduce Elastic Latent Interface Transformer (ELIT), a drop-in, DiT-compatible mechanism that decouples input image size from compute. Our approach inserts a latent interface, a learnable variable-length token sequence on which standard transformer blocks can operate. Lightweight Read and Write cross-attention layers move information between spatial tokens and latents and prioritize important input regions. By training with random dropping of tail latents, ELIT learns to produce importance-ordered representations with earlier latents capturing global structure while later ones contain information to refine details. At inference, the number of latents can be dynamically adjusted to match compute constraints. ELIT is deliberately minimal, adding two cross-attention layers while leaving the rectified flow objective and the DiT stack unchanged. Across datasets and architectures (DiT, U-ViT, HDiT, MM-DiT), ELIT delivers consistent gains. On ImageNet-1K 512px, ELIT delivers an average gain of $35.3\%$ and $39.6\%$ in FID and FDD scores. Project page: https://snap-research.github.io/elit/

Acceleration methods for diffusion models (e.g., token merging or downsampling) typically optimize synthesis quality under reduced compute, yet often ignore discriminative capacity. We revisit token compression with a joint objective and present BiGain, a training-free, plug-and-play framework that preserves generation quality while improving classification in accelerated diffusion models. Our key insight is frequency separation: mapping feature-space signals into a frequency-aware representation disentangles fine detail from global semantics, enabling compression that respects both generative fidelity and discriminative utility. BiGain reflects this principle with two frequency-aware operators: (1) Laplacian-gated token merging, which encourages merges among spectrally smooth tokens while discouraging merges of high-contrast tokens, thereby retaining edges and textures; and (2) Interpolate-Extrapolate KV Downsampling, which downsamples keys/values via a controllable interextrapolation between nearest and average pooling while keeping queries intact, thereby conserving attention precision. Across DiT- and U-Net-based backbones and ImageNet-1K, ImageNet-100, Oxford-IIIT Pets, and COCO-2017, our operators consistently improve the speed-accuracy trade-off for diffusion-based classification, while maintaining or enhancing generation quality under comparable acceleration. For instance, on ImageNet-1K, with 70% token merging on Stable Diffusion 2.0, BiGain increases classification accuracy by 7.15% while improving FID by 0.34 (1.85%). Our analyses indicate that balanced spectral retention, preserving high-frequency detail and low/mid-frequency semantics, is a reliable design rule for token compression in diffusion models. To our knowledge, BiGain is the first framework to jointly study and advance both generation and classification under accelerated diffusion, supporting lower-cost deployment.

We present a parameter-efficient Diffusion Transformer (DiT) for generating 200bp cell-type-specific regulatory DNA sequences. By replacing the U-Net backbone of DNA-Diffusion with a transformer denoiser equipped with a 2D CNN input encoder, our model matches the U-Net's best validation loss in 13 epochs (60$\times$ fewer) and converges 39% lower, while reducing memorization from 5.3% to 1.7% of generated sequences aligning to training data via BLAT. Ablations show the CNN encoder is essential: without it, validation loss increases 70% regardless of positional embedding choice. We further apply DDPO finetuning using Enformer as a reward model, achieving a 38$\times$ improvement in predicted regulatory activity. Cross-validation against DRAKES on an independent prediction task confirms that improvements reflect genuine regulatory signal rather than reward model overfitting.