Diffusion Transformers

Text-conditioned latent diffusion models perform strongly in video generation and are promising backbones for robotic applications. However, existing approaches rely on pixel-level or VAE-based latent representations that lack explicit semantic structure, leaving the impact of the representation space largely unexplored. Slot-based object-centric representations offer a structured alternative by decomposing scenes into object-level latents, or slots. While they have shown success in dynamics modeling and planning, they have not yet been explored for diffusion-based generative modeling. We introduce SlotDiT, a text-guided Diffusion Transformer (DiT) that operates in a slot-based latent space. Given a reference image and a language instruction, SlotDiT decomposes the scene into object-centric slots representing individual entities. Conditioned on the instruction and observed scene context, the model autoregressively denoises future slot trajectories to predict scene dynamics. To systematically investigate latent-space design for diffusion transformers, we compare slot-based representations against VAE-based and semantics-aligned alternatives within a unified DiT framework. Our experiments show that using slots as DiT latents yields competitive video generation quality while consistently improving task-completion rates across four robotic datasets. Furthermore, their compact representation provides a computationally efficient alternative to VAE-based and semantics-aligned latent spaces. Overall, our results demonstrate that object-centric structure is a powerful inductive bias for diffusion-based generative modeling in robotic environments. The project page is available at https://slot-dit.github.io/.

Multimodal brain state decoding has largely focused on fusing paired modalities for prediction, but has rarely explored how their correspondence can be further exploited to enrich training data and improve multimodal representation learning. To address this gap, we propose CoMA-DiT, a bidirectional cross-modal Diffusion Transformer for latent augmentation that treats paired modalities as sources of mutual generative supervision rather than merely as inputs to be fused. CoMA-DiT conditions velocity prediction on the paired modality through cross-modal attention and adaptively injects the resulting variation via a reliability-gated residual mechanism. Experiments on multimodal auditory attention decoding and emotion recognition showed that CoMA-DiT consistently outperformed 20 representative baselines, achieving absolute gains of 4.28% and 6.70% in accuracy and macro-F1 over the no-augmentation baseline, respectively. Extensive ablation, sensitivity, visualization, and interpretability analyses further demonstrated its robustness, generalizability, and ability to capture functionally relevant cross-modal interactions. These findings support a broader view of multimodal learning: Paired modalities can serve not only as inputs for fusion but also as supervision sources that augment one another.

Multi-subject video generation faces two key challenges: uncontrollable fidelity strength and potential semantic drift. We address these by analyzing the internal mechanisms of Diffusion Transformers (DiTs). We found that certain attention blocks naturally form an Intrinsic Spatial Grounding Map (ISGM) that precisely locates reference subjects. Building on this insight, we propose Dual-phase Intrinsic Attention Leveraging (DIAL), a framework that uses these internal signals for both training and inference. In low-noise stages, we use ISGM to guide the attention mechanism, allowing precise control over fidelity strength during inference without retraining. In high-noise stages, we use these same maps to automatically build preference pairs at no additional cost for Reinforcement Learning (RL). This RL procedure effectively anchors the model's attention to reference subjects and mitigates semantic drift. Extensive experiments show that DIAL significantly outperforms baseline models on the OpenS2V-Eval benchmark, consistently improving identity consistency and enabling controllable fidelity strength.

Automated plankton imaging produces severely long-tailed datasets, where the rare taxa of greatest ecological interest have too few images to train or evaluate classifiers reliably. We generate synthetic plankton imagery conditioned on taxonomy: a CLIP encoder is adapted on a large plankton corpus with a ranked contrastive objective extended to deep, ragged taxonomies, then frozen to condition a parameter-efficient diffusion transformer. We evaluate synthetic sample quality on distributional fidelity and downstream classifier utility.

Monocular depth estimation is a ubiquitous yet highly ill-posed computer vision task, with downstream applications in scene reconstruction, computational photography, and robotics, among others. Despite the field's maturity, recent models still struggle to generalize to out-of-distribution inputs and to produce sharp and detailed depth maps. In this paper, we revisit Marigold, a set of techniques for repurposing modern image generation and editing models, powered by the diffusion transformer (DiT) architecture, into state-of-the-art monocular depth estimators. Our recipes target single-step inference from pretrained multi-step flow-matching models, with quantization where needed, preserving model capacity while remaining cheap to run. We analyze the artifacts of naive training and identify two effective remedies: aligning the model's internal representations with semantic features extracted from ground-truth, and adopting a 2-stage fine-tuning protocol built around a novel Sinkhorn-based loss. The results are crisper, cleaner depth maps that generalize well out-of-distribution, with 16-26% improvement in AbsRel over the previous best on KITTI and ETH3D. Qualitatively, our model resolves fur, foliage, and hair-thin edges that have eluded prior models. Furthermore, Marigold V2 achieves state-of-the-art results when applied to other dense regression tasks, such as surface normals estimation and intrinsic image decomposition. Project website: https://hf.co/spaces/huawei-bayerlab/marigold-v2-web

Current text-to-image systems typically employ a "text encoder plus diffusion decoder" paradigm, in which text semantics directly modulate continuous latent noise. Despite their success, these methods lack an explicit, interpretable intermediate representation that effectively bridges high-level linguistic semantics and low-level visual signals. In this paper, we propose Vision-of-Thought (VoT), a framework that introduces a discrete visual-thinking layer between vision-language models (VLMs) and diffusion transformers (DiTs). Instead of treating VLMs merely as text encoders, we use them as multimodal planners that generate discrete VoT tokens representing high-level visual plans, such as objects and layouts, before rendering pixels. We train a specialized VoT tokenizer in the VLM semantic space with a closed-loop objective that combines VLM alignment, feature reconstruction, and vector-quantization losses. These objectives make the tokens semantically readable by the VLM while preserving the visual information needed for generation. Experimental results demonstrate that VoT improves semantic alignment and provides a structured interface for interpretable and controllable generation.

We present Alignment-Free Text-Audiobox (Text-AB), a unified framework for high-quality voice dubbing and full-duplex dialogue synthesis. Building on a Diffusion Transformer trained with a flow-matching objective, Text-AB departs from the Audiobox system along three dimensions. First, it operates in a latent diffusion framework using DAC-VAE features that encode 48 kHz waveforms into a 25 Hz latent sequence, giving over 10x higher compression than previous EnCodec representations while improving resynthesis quality. Second, Text-AB is alignment-free: it consumes raw text via an off-the-shelf text encoder and learns text-speech alignment through cross-attention, removing the need for forced alignment and explicit duration prediction. Third, we scale model and data substantially, pretraining a 3B-parameter model on 480k hours of monolingual speech, followed by supervised fine-tuning on three downstream tasks: cross-lingual voice dubbing, full-duplex dialogue synthesis, and emotional full-duplex dialogue synthesis. At inference, Text-AB supports one-shot generation for up to ~1 min of speech and arbitrarily long-form generation via a multi-diffusion scheme, plus a multi-stage reranking strategy that enhances quality based on automated metrics. On a real-world dubbing benchmark, Text-AB delivers a step-change improvement over the latest internal dubbing system, with large gains in prosody similarity, voice similarity, naturalness, and shareability. For full-duplex dialogue synthesis, it approaches human recordings on short-form conversations and substantially outperforms the latest internal model on long-form human-likeness and expressivity, while natively modeling turn-taking, back-channeling, and emotional dynamics. For emotional dialogue synthesis, emotion conditioning significantly improves emotion alignment and emotional interaction quality over the unconditioned baseline.

Real-world image super-resolution (SR) increasingly relies on Diffusion Transformer (DiT) backbones, whose internal activations can be dominated by a small number of massive channels. Yet improving perceptual quality in these models still typically requires fine-tuning the network or attaching additional adapters, leaving this structured activation space largely unexplored for adaptation. We investigate whether dominant channels can instead serve as a compact adaptation interface for frozen DiT-based SR models. We first characterize their behavior in pretrained SR backbones and show through controlled interventions that they strongly affect reconstruction quality. Building on this observation, we introduce SPARK, a lightweight input-conditioned controller that predicts bounded per-channel affine transformations for only the selected channels, while keeping the SR backbone and VAE frozen. Dominant channels are identified through an online activation-ranking procedure, and only a small predictor conditioned on the low-resolution VAE latent is optimized. Experiments on three DiT-based SR backbones across DIV2K, RealSR, and DRealSR show consistent gains in both fidelity and perceptual quality while modulating only eight channels per stream and block. Controlled comparisons further show that these gains cannot be explained by parameter budget or access to the selected channels alone.

We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior through image-only pre-training and mid-training. The generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image into LLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.

Representational similarity is foundational to analyses of deep networks, yet distances between point-valued representations are not intrinsically tied to downstream function: nearby states may produce different behaviors, while distant states may behave similarly. We instead give representations volume, turning similarity into statistical distinguishability. Overlapping stochastic representations necessarily induce overlapping downstream distributions, grounding latent comparison in model function and bringing it under information-theoretic tools such as the data-processing inequality. We realize this idea in pretrained transformers through a light-touch modification to LayerNorm: at each residual-stream read, we normalize the state, add isotropic Gaussian noise, and renormalize. During distillation fine-tuning, one learned allocation parameter per residual-stream read distributes a fixed global rate budget across the processing stack. The resulting model can be viewed as transformer blocks reading the residual stream with learned finite precision under a shared global rate budget. Using the Bhattacharyya coefficient, we trace which counterfactual distinctions are preserved through MLP blocks or selectively exposed to the query, key, and value computations of individual attention heads. Experiments on ViT-S and GPT-2 small reveal the depthwise propagation of continuous visual perturbations and head-specific sensitivity to token distinctions aligned with known attention motifs. These results establish distinguishability as a functionally grounded lens on transformer computation that complements existing interpretability approaches.

Deep learning-based side-channel analysis has historically focused on single-byte targets and manually cropped traces, which risks discarding exploitable leakage. While recent work has proposed specialized architectures and resampling techniques to address this gap, the literature lacks a simple transformer baseline for simultaneous full-key attacks on uncropped traces. We present an open-source transformer implementation for uncropped full-key attacks which uses the standard transformer encoder backbone, adapting only the input and output layers to the side-channel setting. We release our implementation, training recipes, and pretrained weights for uncropped ASCADv1f, ASCADv1r, and CHES-CTF-2018 which achieve performance competitive with previously-reported results, while using less than 10GB of VRAM and requiring at most 3.34 hours of training on a single NVIDIA A6000.

Real-world image generation often involves multi-turn editing, where users iteratively modify small regions while most image content remains unchanged. However, existing diffusion transformer (DiT)-based editing pipelines recompute the entire image at every turn, causing substantial redundant computation. Existing DiT acceleration methods further ignore semantic correspondence across prompts, leading to unnecessary recomputation or unsafe reuse that harms editing quality. To address this, we propose RegionCache, a semantic-aware reuse framework for multi-turn image editing that selectively reuses diffusion states from unchanged regions. RegionCache detects reusable regions through semantic overlap between consecutive prompts and cross-attention localization, and adopts an adaptive reuse schedule based on prompt similarity and contextual consistency. Experiments on PixArt-alpha demonstrate that RegionCache achieves 1.43x--2.55x end-to-end speedup while maintaining comparable image quality. Code is available at https://github.com/hebutBryant/RegionCache.

Multimodal Diffusion Transformers (MM-DiTs) have demonstrated remarkable text-to-image generation performance, surpassing traditional U-Net-based diffusion models. Nevertheless, their powerful generative capabilities also raise significant safety concerns, as they may generate sensitive or inappropriate content. While existing concept erasure methods aim to mitigate such risks, most require modifying model parameters, which are often architecture-specific and impractical for deployed larger models. Several tuning-free approaches face challenges when applied to advanced large-scale MM-DiTs due to their deeply embedded knowledge, broad semantic space, and context-dependent text encoders. To address these challenges, we propose to erase concepts by directly manipulating the model's internal representations. Our key insight, derived from an in-depth analysis of MM-DiT's block-wise generative roles, is that text-conditioned semantic representations are most salient in the middle blocks of MM-DiTs. Based on this, we extract representations of an unwanted concept and a desirable safe one from the middle block, construct a steering vector from their difference, and inject this single vector into consecutive early and middle blocks. By operating exclusively on the sparse text-branch tokens and leveraging the straight sampling trajectory of rectified flow, our method achieves effective concept erasure with negligible overhead and without any training. Extensive experiments across MM-DiT models demonstrate that our method achieves state-of-the-art performance in erasing diverse concepts, enables effective control over the final output, and remains robust to adversarial attacks.

Modern image editors combine vision-language models (VLMs) with diffusion transformer backbones to modify a single reference image according to instructions without fine-tuning. This capability also enables unauthorized manipulation of publicly released images. Existing inference-time defenses either invalidate edits through conspicuous corruption, thereby exposing the protection, or allow them to proceed with identity or reference content drift, thereby failing to prevent the editing behavior itself. We instead target a stealthy and harmless no-op in which the requested edit is suppressed, the output remains natural and source-preserving without conspicuous artifacts or identity replacement, and harmful semantics requested by malicious instructions are absent. We propose NullEdit, which targets the VLM representation jointly formed from the reference image and instruction before it conditions the downstream DiT backbone. Using normal-edit and no-edit anchors, NullEdit redirects this representation, while cross-prompt gradient averaging transfers protection to held out instructions. Across Step1X-Edit and Qwen-Image-Edit on CelebA-HQ and VGGFace2, NullEdit reduces the EditReward IF score by 0.813 on average relative to the SOTA baseline while preserving subject identity and source content.

Vision Transformers demonstrate remarkable global modeling capacity but often underperform in data-scarce regimes. Distilling convolutional inductive biases from a CNN teacher provides an effective remedy while leaving the deployed model unchanged. However, general-purpose feature distillation transfers little in this setting. In CNN-to-CNN distillation, pooling, flattening, and logit-space projections remove the spatial grid that encodes locality and translation equivariance. Unlike a convolutional student, a ViT cannot readily reconstruct this structure on its own. In this paper, we propose iBKD, a distillation framework that preserves the spatial grid throughout the entire transfer process. Its core module, the Inductive Bias Attention Module, aggregates features from all student layers onto the teacher's grid using learned weights. It then enhances structural cues through channel and deformable spatial attention and injects them via convolutional cross-attention operating directly between spatial grids rather than token sets. The module is used only during training, leaving the deployed model as an unmodified ViT with no inference overhead. Across seven Transformer backbones and six data-scarce benchmarks, iBKD consistently outperforms both locality-guidance methods and general knowledge distillation baselines, with its advantage increasing as the amount of training data decreases.

Latent diffusion models have recently advanced video frame interpolation by synthesizing intermediate frames between input images. However, handling large temporal gaps and complex motion remains challenging, often resulting in motion blur, structural distortions, and temporal inconsistencies. Event cameras provide high-temporal-resolution motion cues that are well suited for bridging these gaps and improving interpolation quality. To exploit this advantage without training an event-assisted model from scratch, we propose an adapter-based framework that incorporates event-derived cues into a pre-trained image-to-video diffusion model with minimal architectural changes. Specifically, our method leverages Image Warped Events (IWEs) and bidirectional sparse optical flow to provide spatially and temporally aligned guidance during generation. By injecting these event-guided structural and motion cues into the diffusion process, our approach reduces interpolation artifacts and improves both reconstruction fidelity and temporal coherence. Experimental results on real and synthetic benchmarks show that our method consistently outperforms existing state-of-the-art approaches. The project page is at https://joseph-lin-tech.github.io/BridgeEventDiT-VFI/.

Diffusion transformer (DiT), a rapidly emerging architecture for image generation, has gained much attention. However, despite ongoing efforts to improve its performance, the understanding of DiT remains superficial. In this work, we delve into and investigate a critical conditioning mechanism within DiT, adaLN-Zero, which achieves superior performance compared to adaLN. Our work studies three potential elements driving this performance, including an SE-like structure, zero-initialization, and a "gradual" update order, among which zero-initialization is proved to be the most influential. Building on this understanding, we propose an analysis-guided initialization strategy, termed adaLN-Gaussian, which serves both as an empirical validation of our analysis and as a practical initialization method that consistently improves optimization efficiency. On the other hand, inspired by the SE-like structure, we introduce an improved conditioning mechanism called SE-adaLN-Zero. Extensive experiments following DiT on four datasets, especially on ImageNet1K demonstrate the effectiveness and generalization of adaLN-Gaussian and SE-adaLN-Zero. Beyond class-to-image generation, we also evaluate the generalization of the two improved methods on text-to-image generation.

Data-Free Knowledge Distillation (DFKD) transfers knowledge from a pretrained teacher model to a compact student model by synthesizing semantically informative data, eliminating the need for access to the original training dataset. Existing DFKD methods rely heavily on architecture-specific statistical priors (e.g., Batch Normalization statistics) to guide data synthesis, however, such architecture-dependent priors are often absent in modern architectures such as Vision Transformers (ViTs), resulting in degraded semantic quality of the synthesized data and consequently catastrophic performance degradation. In this paper, we propose \emph{UniDFKD}, a unified data-free knowledge distillation framework that replaces architecture-specific statistics with explicit, architecture-agnostic semantic priors. \emph{UniDFKD} governs the entire synthesis-distillation pipeline along three dimensions: (1) Categorical Semantic Conditioning (CSC) defines \emph{what} to synthesize by persistently modulating the generator with language-derived embeddings to capture semantic diversity; (2) Spatial Semantic Anchoring (SSA) dictates \emph{where} evidence belongs by anchoring the teacher's spatial attributions to a Gaussian prior; and (3) Spatial Semantic Distillation (SSD) controls \emph{how} knowledge is transferred by explicitly aligning teacher-student spatial evidence alongside predictions. Extensive experiments across CNNs and ViTs demonstrate that UniDFKD establishes a new state-of-the-art, outperforming existing methods by an average absolute margin of over 20% in both homogeneous and heterogeneous settings.

Diffusion caching is a lightweight strategy that accelerates Diffusion Transformers (DiTs) by reusing intermediate features across denoising steps, but existing paradigms face a fundamental trade-off: online heuristics lack global budget awareness, whereas static schedules lack instance adaptivity and fail to flexibly adapt to varying runtime budget constraints. To bridge this gap, we present BAG (Budget-Aware Gating), a novel caching policy that unifies global budget pacing with dynamic, instance-adaptive feature reuse. Rather than relying on hand-crafted rules, BAG employs a lightweight gating network that dynamically decides whether to execute a full computation or reuse cached features at each step by jointly conditioning on the budget state and local trajectory feedback. We train this policy via offline-to-online schedule distillation, transferring the decision-making of offline-searched schedules into a compact online gate. Extensive experiments on FLUX.1-dev and Wan2.1 demonstrate that BAG consistently outperforms state-of-the-art caching methods across various speedup tiers while remaining robust across different resolutions, seeds, and guidance scales. Code will be released.

Image super-resolution (SR) with large generative models has recently achieved remarkable perceptual quality, yet maintaining fidelity to the LR observation remains challenging. In particular, we observe that diffusion transformers (DiTs) built on latent representations suffer from a critical limitation: the compression bottleneck of the VAE weakens fine-grained spatial information, leading to hallucinated details that are weakly grounded in the input image. In this work, we revisit generative SR from a representation perspective and propose a pixel-grounded super-resolution (PGSR) framework that preserves LR-observed pixel evidence before VAE compression and reuses it throughout restoration. Instead of relying solely on the compressed latent condition, PGSR extracts pre-VAE pixel evidence from the upsampled LR image and reuses it at two stages. First, Condition-Side Trajectory Guidance fuses LR-derived pixel evidence with the latent LR condition to guide the latent restoration trajectory. Second, Decoder-Side Pixel Grounding injects multi-scale pixel features into the frozen VAE decoder to ground the final rendering with LR-observed cues. To efficiently adapt large pretrained DiT models, we keep the latent autoencoder and main flow-matching backbone frozen, and train only lightweight restoration modules. We further study an efficient local-window attention variant for improved high-resolution efficiency and scalability. Extensive experiments demonstrate that PGSR improves the realism--fidelity trade-off and produces more faithful, visually convincing results than existing latent generative SR approaches.

Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth. Dense attention gives each token broad horizontal access to the past, but the vertical feedback channel between decoding steps remains narrow: only the sampled token returns to the bottom of the stack, while the top-layer hidden state is discarded. We introduce the \emph{full-bandwidth transformer}, which widens this channel with \emph{latent feedback}: at each decoding step, the previous top-layer hidden state is fused with the sampled token embedding through a gated linear unit and fed back as the next input. Latent feedback lets non-verbalized computation re-enter the stack with a renewed depth budget, while preserving the standard transformer architecture, KV cache, and language-modeling objective. To train full-bandwidth transformers without losing parallel teacher forcing, we use a scheduled multi-pass objective that introduces latent feedback late in pretraining and mixes a small fraction of deeper feedback passes for stability. We train 1B-parameter full-bandwidth transformers up to 400B tokens and find that latent feedback improves validation loss, 5-shot language-model evaluation, math and coding generation, and instruction-tuned performance. With negligible per-token decoding overhead, full-bandwidth transformers match or approach standard transformers trained with roughly $1.5\times$ more tokens, and manage to produce shorter reasoning traces at equal or better accuracy.

Neural synthesis for musical instruments has the potential to revolutionize current practices that use concatenative synthesis and a sample library. However, most research focused on piano synthesis and expressive performance generation; little work has been done on continuously articulated instruments like the violin, let alone rendering them with playing techniques and dynamics. We present VIOLET, a latent-diffusion framework for controllable violin synthesis, which uses a Diffusion Transformer (DiT) with rectified flow to synthesize high-fidelity audio from MIDI notes, playing techniques, and continuous dynamics. To train VIOLET, in addition to using a few existing datasets, we curate a new dataset named CSV-TD, which contains 39 h of 48 kHz synthetic audio and time-aligned annotations of MIDI notes, note-level techniques, and continuous dynamics curves. Objective and subjective evaluations show that VIOLET synthesizes violin performances with high technique adherence, accurate pitch and timing alignment, and good dynamics control. It outperforms the current state-of-the-art neural violin synthesis system and approaches a top commercial virtual instrument in terms of technique clarity, naturalness, and dynamics following.

Simulating complex fluid flows requires capturing full equilibrium distributions rather than just mean trajectories, yet high-fidelity solvers remain computationally prohibitive. Recent advances, such as Diffusion Graph Networks (DGNs), have combined diffusion models with graph neural networks to sample equilibrium states directly from unstructured meshes, enabling distributional accuracy even from short simulations. However, graph-based diffusion approaches suffer from hand-crafted architectural constraints, limited receptive fields in message passing, and costly multi-scale designs, which restrict scalability to larger and more complex domains. We propose Fluid-DiT, a Graph-Free Diffusion Transformer that replaces graph message passing with attention-based denoising, eliminating explicit graph design while preserving the ability to model distributions of chaotic flows. Our framework introduces a latent-space formulation that disentangles geometric fidelity from distributional learning, reducing high-frequency artifacts and accelerating sampling. By leveraging the transformer's global receptive field, Fluid-DiT naturally captures both local flow structures and long-range correlations without requiring hierarchical graph coarsening. On canonical benchmarks including laminar cylinder wakes, ellipse-flow systems, and turbulent 3D wing experiments, Fluid-DiT consistently outperforms graph-based diffusion baselines in both sample quality and distributional accuracy, achieving higher $R^2$ correlations and lower Wasserstein distances. Moreover, it generalizes robustly from short, incomplete trajectories to unseen Reynolds numbers and geometries, demonstrating strong scalability.

Training-free text-to-high-resolution image generation has recently attracted growing research attention. However, existing studies on this task primarily focus on adapting off-the-shelf U-Net-based diffusion models to high resolutions, with limited progress on adapting off-the-shelf Diffusion Transformer (DiT) models despite their strong text-to-image generation capabilities at limited resolutions. In this work, we find two key challenges particularly hindering the application of off-the-shelf DiT models for high-resolution image synthesis in a training-free manner, namely, spatial disorder and long generation time. To address these challenges, we propose a novel method tailored to adapt off-the-shelf DiT models for high-resolution image synthesis. Extensive experiments show the efficacy of our method. Our code is available at: https://github.com/zylwithxy/HRDiT.

Layout-guided multi-instance generation is essential for controllable image synthesis in Multi-Modal Diffusion Transformers (MM-DiTs). However, integrating this capability into unified architectures remains challenging. Prior frameworks rely on redundant full-resolution canvas padding and Shifted-RoPE to manage multiple reference images. This mechanism drastically inflates computational overhead for sparse layouts and disrupts critical low-frequency RoPE features, creating a severe spatial-frequency compromise that blurs absolute spatial correspondence. To overcome these limitations, we propose ControlRef, a highly efficient and precise multi-instance synthesis framework. ControlRef utilizes a Unified Instance-Layout Control (UILC) attention mask to strictly decouple inter-instance semantic interactions and enforce precise regional binding. To further promote region-level spatial alignment, we introduce Anchored 4D-RoPE, a novel positional encoding mechanism that directly anchors tokens to their absolute geometric centers. By pre-aligning reference images to their corresponding bounding box resolutions, physically anchoring both layout and reference tokens to their absolute geometric centers, and stacking the references along the z-axis, Anchored 4D-RoPE natively preserves spatial priors and mitigates the spatial-frequency compromise without lossy shifting. Extensive experiments demonstrate that ControlRef achieves state-of-the-art visual fidelity and localization accuracy, while concurrently slashing inference latency by over 80% in sparse layouts and reducing memory overhead by 50% in dense scenarios.

All-in-one image restoration seeks a single model that can recover images degraded by diverse and spatially non-uniform corruptions. However, many unified Transformers rely on fixed patch partitioning: task/degradation condition is injected only into the backbone blocks after tokenization, leaving the embedding and reconstruction stages insensitive to local degradation variations. In contrast to previous approaches, we present Flexible Image Transformer (FIT) that explicitly models degradation awareness across the entire pipeline, from patch sampling to pixel reconstruction. Specifically, FIT employs a lightweight Degradation Encoder to predict a global degradation vector $\mathbf{g}$ and a spatial degradation map $\mathbf{M}$ from local degradation severity, which jointly condition the patch embedding and unembedding through adaptive deformation. Moreover, to improve robustness across degradation types, we introduce a task-token dropout strategy that regularizes task conditioning during training. On five standard benchmarks (BSD68, Rain100L, SOTS, GoPro, and LOLv1), FIT achieves state-of-the-art performance with 30.72 dB average PSNR on the five-degradation setting and 32.83 dB on the three-degradation setting, outperforming recent unified restoration methods by +0.5$\sim$1.1 dB. Moreover, the learned offsets provide a direct handle for visualizing degradation-aware spatial adaptation.

Character image animation remains a foundational yet challenging task in computer vision. Existing approaches can be broadly categorized into three paradigms: methods based on explicit motion representations suffer from extraction errors and identity drift; methods based on implicit motion features lose fine-grained dynamics through compression; and in-context learning approaches avoid intermediate representations but incur prohibitive computational costs. Furthermore, all current systems are designed for offline synthesis, unable to meet the real-time requirements of interactive applications such as digital avatars and live-streaming hosts. To address these limitations, we present Wan-Animate-2, an end-to-end character animation framework that directly consumes the driving video within a redesigned Diffusion Transformer. Our architecture achieves superior motion fidelity and identity preservation by eliminating intermediate motion extractors entirely. We further introduce text driven viewpoint control that decouples the output camera perspective from the driving video--a capability rarely supported by prior character animation methods that rely on explicit motion representations. Beyond generation quality, we present Wan-Animate-2-Lite, an efficient variant that reduces inference latency to real-time thresholds through a three-stage training paradigm: teacher forcing pretraining with error buffer mechanism, and Self-Forcing distillation with chunk-wise backpropagation. This enables streaming character animation for interactive applications, opening new deployment scenarios that were previously infeasible. Qualitative evaluations and user studies demonstrate that Wan-Animate-2 achieves high-fidelity animation results across diverse characters and motion patterns. To foster further research and community development, we will release the Wan-Animate-2-Base model weights to the public.

Text-to-image diffusion transformers learn about objects and scenes by learning to generate them, making them strong candidates for training-free zero-shot open-vocabulary semantic segmentation. State-of-the-art attribution methods score each pixel independently, comparing its features against a fixed text-derived class representation, whether as an output-space similarity or as a cross-attention weight. This discards structured signals the model itself exposes: the temporal structure of the generative trajectory, the visual appearance statistics of each concept, and the image's own pairwise feature geometry. We present MAVISEG, a training-free refinement layer that recovers these signals. Because its operators consume only a pixel-by-concept score field and a pixel feature space, MAVISEG is capture-agnostic rather than tied to one attribution method. Across six benchmarks it achieves the strongest overall results among training-free methods, including the best mIoU on every benchmark. Interestingly, gains are largest where the initial capture is weakest, and individual operators contribute depending on the noise in the field they refine. Our results indicate that diffusion transformers carry more concept-level information than current attribution methods recover, and that much of it is lost on the way to the mask rather than absent from the model.

Diffusion transformers deliver strong image generation, but their training cost grows superlinearly with resolution. Recent work justifies training or sampling at reduced resolution on a spectral premise: at high noise, a downscaled latent preserves almost the full surviving signal. Whether a downscaled step also preserves the native training gradient signal, however, has remained unresolved. We reduce how that signal changes under downscaling to two terms: a noise-dependent term governed by the downscale ratio, which decays at high noise as the spectral premise predicts, and a σ-independent floor governed by the target grid's absolute token count, carried by the compute graph itself and removed by no noise level. The measured (route, σ) map corroborates the account and uncovers structure the spectral picture cannot express: on the 1024->768 route, a window (0.65 < σ< 0.95), predicted by no spectral criterion at any tolerance, where the downscaled gradient stays within a small margin of the native one. Training LoRA adapters with downscaled steps restricted to the routes and noise windows the map validates reduces training time by 14.6% at a fixed step budget while remaining near-native in weight space. Code is available at https://github.com/sorryhyun/anima_lora.

Transfer-based adversarial attacks often transfer poorly across heterogeneous architectures because CNNs favor local textures while Vision Transformers (ViTs) rely on global shapes. We propose Season, a spectrum-aware orthogonal gradient refinement framework for L-infinity transfer attacks against black-box target models on ImageNet, using a white-box surrogate. Season decomposes each update into a low-frequency branch capturing structural cues and a high-frequency branch capturing textures. A low-saliency guidance scheme reallocates high-frequency energy to background regions, preserving foreground structures that ViTs depend on. An orthogonal projection then forces the textural update to lie in the orthogonal complement of the structural direction, mitigating feature interference. As a training-free plug-and-play wrapper, Season enhances eight gradient-stabilization and input-enhancement attacks without modifying their cores. Across eight CNN, ViT, and MLP targets, Season improves transfer success rate by 6.6 percentage points on average and up to 16.0 points over strong baselines under a unified protocol.

Diffusion Transformers (DiTs) have demonstrated exceptional performance in high-fidelity image and video generation. To alleviate their massive computational overhead, temporal feature caching has been proposed to bypass redundant computations. However, existing cache-then-forecast methods driven by derivative-based polynomials often cause severe quality degradation under high acceleration due to unstable long-step predictions. To address this bottleneck, we propose Barycentric Rational Forecasting with Chebyshev Enhancement (BRACE). Motivated by the observation that DiT feature trajectories are globally smooth yet frequently exhibit sharp irregularities and local non-smoothness, BRACE shifts the paradigm from derivative-driven polynomial extrapolation to feature-driven rational forecasting. Specifically, it maintains a local sliding window to cache sparse historical features and leverages adapted Chebyshev weights to formulate a barycentric rational function, directly aggregating these raw features to ensure numerical stability. Extensive experiments demonstrate that BRACE achieves state-of-the-art quality-efficiency trade-offs across various DiT architectures with negligible computational overhead.

Recent advances in Diffusion Transformers (DiTs) have enabled remarkable progress in visual synthesis, benefiting from their superior scalability. To facilitate DiTs' capability of capturing meaningful internal representations, recent works such as REPA incorporate external pretrained encoders for representation alignment. However, the underlying mechanisms governing representation learning within DiTs remain poorly understood in the community. To this end, this paper first presents a systematic analysis of the representation dynamics of DiTs via quantifying the diversity of block-wise representations. Specifically, we introduce a novel metric, termed the Weighted Diversity Score (WDS), to measure the representational discrepancies across different blocks. Through extensive investigations on the evolution and influence of internal representations under various settings, we reveal that representation diversity across blocks is a critical factor for effective representation learning in DiTs. More importantly, WDS exhibits a strong correlation with synthesis quality across diverse settings, model scales, and training stages (Pearson's $r=-0.869$ with $\log(\text{FID})$), suggesting its potential as an indicator to reflect model performance and a principled guide for model optimization. Based on this key finding, we propose DiverseDiT++, a novel framework that explicitly promotes diverse representation learning. Concretely, our method incorporates long residual connections to diversify input representations across blocks and a representation diversity loss to encourage blocks to learn distinct features. Extensive experiments on ImageNet $256\times256$ and $512\times512$ demonstrate that our DiverseDiT++ yields consistent performance gains and convergence acceleration when applied to different backbones with various sizes,...

Text-to-video (T2V) diffusion transformers (DiTs) are trained with detailed video captions, whereas inference often relies on user prompts rewritten by a prompt enhancer (PE). Prior work has improved generation by optimizing the PE, the DiT, or both; some methods have also sought to narrow the training-inference mismatch through shared schemas. Yet even within a shared schema, inference-time PE outputs and DiT training captions may still differ in detail selection, information organization, descriptive granularity, and phrasing. We refer to this residual mismatch as the PE-Caption gap and introduce CAPE-T2V, a two-step Captioner-Anchored Prompt Enhancement framework toward two-sided conditioning alignment in T2V generation. First, CAPE-T2V constructs three types of PE training examples, pairing captioner-generated targets with concise source captions, detailed source captions, or pseudo user prompts derived from those targets. It then fine-tunes the PE to map each input to its paired target. Second, CAPE-T2V fine-tunes the DiT on video-derived captions rewritten by the Anchored PE; the same PE rewrites user prompts at inference. Relative to a baseline using the same caption schema, CAPE-T2V achieves higher aggregate scores on StoryEval, VBench-2.0, and T2V-CompBench across Wan2.2 and LTX-2.3. Further, CAPE-T2V exhibits a smaller PE-Caption gap than the baseline: its DiT fine-tuning captions are closer in distribution to inference-time PE outputs, as measured by squared maximum mean discrepancy in a fixed embedding space. Overall, these results support CAPE-T2V as an effective approach to mitigating the PE-Caption gap. The project is available at https://github.com/yizzz927/CAPE-T2V.

Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. Existing continuous language models either inherit embedding spaces not designed for joint generation and decoding, or compress autoencoded latents to ease diffusion, sacrificing token-level fidelity. Instead of simplifying the representation to suit the generative model, we preserve a high-capacity, decodable text latent and design the diffusion model to learn its distribution directly. We introduce AURORA-LM, a continuous-latent diffusion language model that separates the construction of a decodable text representation from the modeling of its distribution. A Query-based Encoder-Decoder organizes text into a high-capacity, prefix-aligned latent sequence, and a Block-causal Diffusion Transformer learns its distribution through flow matching, generating blocks left to right while denoising positions within each block in parallel. Because such a latent is harder for diffusion to model, AURORA-LM restricts only the noisy-input pathway while retaining the full clean-latent prediction target, accommodating full-width latents without reducing decoder-facing capacity. We further calibrate the noise-level distribution to the latent width, and introduce self-trajectory consistency to bridge independently sampled training noise and iterative denoising at inference. AURORA-LM achieves the strongest performance among evaluated continuous and diffusion-based language models on OpenWebText free generation and XSum summarization. Scaling to 1B parameters with about 1500 EFLOPs of total compute yields further gains, surpassing a larger publicly released latent-diffusion language model under a matched evaluation protocol. All experiments are conducted on Ascend NPUs.

Diffusion Transformers (DiTs) have achieved state-of-the-art (SOTA) performance in visual generative modeling, yet their training remains computationally prohibitive. While the recently proposed Momentum Orthogonalization (Muon) optimizer offers a promising alternative to AdamW, its direct application to DiTs yields suboptimal late-stage convergence. In this paper, we identify the root cause of this bottleneck: standard DiT architectures fuse functionally distinct weights (e.g., within AdaLN and QKV layers) into unified tensors for computational efficiency. Applying Muon to these fused tensors inadvertently induces implicit subspace coupling, which distorts update directions and degrades global optimization. To address this, we introduce Chunked Muon (CMuon), a simple yet highly effective strategy that partitions these matrices into independent sub-components prior to orthogonalization. Extensive experiments demonstrate that a 675M-parameter DiT trained with CMuon achieves a FID of 1.18 on ImageNet 256 in just 200 epochs. This represents more than a 2x training speedup over AdamW, while effectively overcoming the late-stage convergence plateaus of vanilla Muon.

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

Diffusion Transformers (DiTs) have emerged as a core architecture in generative modeling due to their scalability and adaptability to multimodal tasks. DiTs comprise isotropic transformer blocks, and learn representations progressively across depth, where the denoising objective drives later layers to focus on fine-detail reconstruction. This results in degraded representation quality and an imbalanced encoder-decoder behavior. Prior approaches such as representation alignment (REPA) mitigate this by encouraging stronger early representations via training regularization. Alternatively, U-Net-style DiT architectures introduce explicit multi-scale encoder-decoder structures for improved convergence. But they build on standard U-Net wisdom via learnable operators for spatial downsampling, which are not well-suited to transformer architectures, introducing inefficiencies and compatibility issues with components such as cross-attention and representation regularization. In this work, we propose UDT, a U-Net diffusion transformer that combines the representation power of DiTs with the encoding-decoding benefits of U-Nets, through data-adaptive token merging for downsampling and upsampling, while preserving the DiT token dimension. Our baseline UDT architecture outperforms existing U-Net DiTs and achieves performance comparable to REPA across all model sizes. Furthermore, using architectural optimization and REPA, UDT outperforms SiT's 7.9 FID at 1400 epochs (w/o CFG) within 40 epochs (~ 40x faster convergence) for XL model size on 256x256 ImageNet. Finally, it achieves strong image generation performance with CFG, reaching FID of 1.38 (320 epochs) with SD-VAE and 1.35 (500 epochs) with VA-VAE, providing a new backbone for DiTs with strong empirical benefits.

Background and Objective: Generating realistic medical images with anatomically accurate segmentation masks helps address the shortage of annotated data in medical imaging, particularly in optical coherence tomography (OCT) of mouse eyes, where manual retinal layer delineation is labour-intensive due to tiny structures and required expertise, resulting in scarce datasets. While diffusion models perform well in medical image synthesis, joint image-mask generation has relied mainly on U-Net-based denoisers, leaving diffusion transformers largely unexplored. Methods: We propose a conditional dual-output Diffusion Transformer (DualDiT) for joint synthesis of OCT B-scans and segmentation masks of the upper retinal cell layers in ex vivo mouse retina. DualDiT encodes both modalities into a shared latent space via a pretrained VAE, concatenates their latent representations, and performs conditional diffusion over the joint tensor. We compared DualDiT against two adapted diffusion baselines: DDPM and LDM. Generative quality was assessed via Fréchet Inception Distance (FID) and spatial FID (sFID); practical utility via synthetic data augmentation for downstream U-Net segmentation; and perceptual realism via evaluation by three domain experts. Results: DualDiT achieved the best generative quality (FID 56.14, sFID 114.35), outperforming DDPM and LDM. Expert panels misclassified 46% of synthetic samples as real and 42% of real samples as synthetic. Adding DualDiT-generated images and masks improved Dice and IoU scores on a held-out segmentation test set. Conclusions: DualDiT shows that transformer-based diffusion models can effectively learn the joint distribution of OCT images and segmentation masks, surpassing DDPM- and LDM-based baselines in generative fidelity, downstream utility, and perceptual realism, highlighting its potential for data augmentation in annotation-scarce medical imaging.

As image generation models scale to ever higher resolutions, global coherence, local detail, and texture fidelity become critical axes for generation quality. However, standard flow matching treats all spatial frequencies uniformly, ignoring the natural frequency hierarchy where high-frequency bands become indistinguishable from pure noise far earlier than coarse structures. We introduce WaiT, a Wavelet-aware image Transformer that decomposes generation into coarse and fine bands via lossless wavelets. True to its name, the high-frequency bands wait for the signal: staying pure noise until coarse structure has emerged, then joining the flow for joint refinement. Since standard FID discards fine-grained detail through aggressive downsampling, we introduce a more stringent three-axis evaluation protocol to assess quality at native resolution. On ImageNet 512x512, WaiT achieves a pixel-space FID of 1.43 and is Pareto-optimal across all three axes, reducing sampling compute by up to 50%. With our largest 2B model, we set a new state-of-the-art FID of 1.3 for pixel-space models on ImageNet 512 resolution. Our formulation outperforms even the strongest latent-space models on texture fidelity, and scales seamlessly to high-resolution OpenImages and to video generation, achieving a state-of-the-art FVD of 0.84 on Kinetics-600 with no algorithmic modifications.

Medical image fusion aims to integrate complementary information from diverse imaging modalities to support clinical diagnosis. Existing methods typically apply uniform fusion rules globally, lacking a deep understanding of diagnostic intents and pathological structures. To address these limitations, we propose MIND, a Multimodal Intent-Driven Network via Diffusion Transformers (DiTs) for medical image fusion. Specifically, we utilize BioMedGPT to generate intent-driven fusion texts from source images, guiding the fusion process with pathology-aware diagnostic intents. To combat the loss of 2D spatial continuity caused by 1D sequence flattening in DiTs, we design a Multi-scale Latent Adapter. This module explicitly extracts source image features before serialization, injecting them into the network via strict dimensional alignment to effectively supplement image features. To resolve the semantic shift caused by decoupling image outputs from diagnostic intents, we design a medical semantic consistency loss. This loss ensures deep semantic locking between fused images and fusion texts while maintaining the stability of the underlying physical manifold reconstruction. Comprehensive experiments on the Harvard, BraTS, and GFP datasets reveal that MIND delivers superior fusion quality, significantly improves downstream brain tumor segmentation accuracy, and enables flexible interactive fusion, holding significant promise for intent-driven intelligent clinical decision support systems.

We present ReGenVC, an end-to-end generative video codec that compresses talking-head video to an ultra-low bitrate and decodes it in real time. The encoder reduces a source clip to a compact bitstream -- a neurally compressed first frame, per-frame pose keypoints, and metadata -- totaling about 26 kB for a 77-frame sequence. The decoder is a four-step distilled diffusion transformer that reconstructs the video conditioned on the transmitted pose and reference frame. Compared with x264/x265, ReGenVC reduces the bitrate to roughly one tenth of that required by traditional codecs (about 26 kB vs. 250--280 kB for essentially artifact-free reconstruction); at a matched ultra-low bitrate, conventional codecs collapse into blocking artifacts while ReGenVC stays sharp by exploiting a strong generative prior. The central obstacle to deploying such a codec is decoder latency: multi-step sampling with transformer and VAE components is too slow for interactive use. We make the decoder real-time through four-step distillation and three model-preserving system techniques: (i) eight-GPU unified sequence parallelism (Ulysses & Ring), (ii) a spatially-split VAE, and (iii) a three-stage overlapped pipeline; an analytical timing model characterizes the real-time feasibility region. On an 8-GPU node, the system sustains 24 fps output (972 ms per 25-frame window, within the 1000 ms budget), enabling a live browser stream without observed frame underruns. A hybrid CPU-GPU deployment further runs the encoder on the CPU at 24 fps and offloads the decoder-side one-shot conditioning encoders to the CPU, reducing the per-GPU memory peak from 21.1 GB to about 7.7 GB. To our knowledge, ReGenVC is the first end-to-end generative video codec to combine ultra-low-bitrate encoding with real-time decoding on an 8-GPU system.

Diffusion models are widely used to generate high-quality images and videos, but their iterative denoising process remains computationally intensive. A growing class of training-free accelerators reduces this cost by reusing cached intermediate features or forecasting future ones. To control draft drift, these methods sometimes compute an exact block feature for verification. Yet the resulting exact feature is typically used only to measure discrepancy or guide a later decision and is then discarded. We find that this previously computed feature can instead be reused for correction. Forwarding it at the verification site resets the local draft residual and reduces downstream feature error. Based on this observation, we introduce FeatFix, a local exact-feature correction method for cached diffusion inference. FeatFix operates at a fixed sparse set of layer--timestep sites. At each selected site, it replaces the complete draft block output with the exact output computed from the same incoming state, avoiding token- or channel-level partial replacement and full-timestep recomputation. Experiments across four image and video backbones show that FeatFix consistently accelerates generation, achieving a speedup of up to $6.70\times$ over Vanilla while maintaining competitive output quality.

Looped Transformers create a useful train- and test-time compute axis by reusing the same Transformer block over recurrent depth, increasing effective depth at a fixed parameter count. However, that shared block must then govern an entire trajectory of varying hidden states over trained and extrapolated depths. Furthermore, in additive-injection looped Transformers, an input-conditioned signal is reintroduced at every recurrent step, so applying the shared transition at an input-conditioned reference can still move the hidden state. In this paper, we propose Source-Centered State Evolution (SCSE), which is designed to reconcile input conditioning with reference-preserving shared recurrence. Specifically, SCSE retains input dependence through its learned anchor and initial deviation, allows nonzero deviations to drive recurrent computation while mapping zero deviation to zero, and guarantees exact anchor invariance through its zero-deviation mask. The designated anchor is thereby a one-step fixed point by construction. The zero-deviation forcing bias is the next deviation produced from the anchor itself and vanishes in SCSE, while nonzero deviations remain active and support state-dependent recurrent computation. Our theory shows that the zero-deviation forcing bias is a design degree of freedom whose task effect can be harmful, neutral, or beneficial; SCSE resolves this choice in favor of exact anchor invariance by setting the bias to zero. Across WikiText-2, WikiText-103, direct web-corpus pretraining, held-out web-text transfer, and LAMBADA completion, SCSE improves the controlled recurrent quality frontier. Ablation studies identify the learned anchor and the anchor-coordinate deviation recurrence as the primary contributors to the gain, and a trained-model case study grounds the anchor-response diagnostic in observed recurrent motion.

Audio deepfake detectors often degrade when generators, corpora, or recording conditions change. We use a Diffusion Transformer (DiT), trained only on bona fide speech, as a frozen reconstruction probe. Reconstructions at masking ratios 0.5, 0.75, and 0.9 yield explicit multi-ratio residual maps. Because these residuals are domain sensitive, our audio-anchored detector passes the projected frozen-WavLM auditory representation into the fusion sum without gate-based attenuation and uses residuals only as a scalar-gated additive correction. The pre-specified seed-42 run obtains 6.5442% EER / 0.18456 min-DCF on ASVspoof 5 Eval and 13.8372% / 0.36921 on ITW Full; three-seed means are 6.8885 (0.3308)% and 15.3328 (2.0719)%. The latter is below a separately optimized WavLM-ResNet18 reference under both supervision settings. Auxiliary supervision raises dynamic competitive fusion from 18.4007% to 25.2968% mean ITW EER, worsening all three seeds. The results support reconstruction residuals as complementary evidence and motivate a non-competitive auditory path for ASVspoof 5-to-ITW transfer, without claiming a componentwise causal ablation of anchoring alone.

An often overlooked factor of robot manipulation performance is the embodiment of the robot itself. Motivated by this problem, we study motion-conditioned robot co-design, where the goal is to generate complete robot designs that track target end-effector trajectories (from human demonstrations) while optimizing user-defined rewards. We introduce Transformer Transformer, a diffusion transformer trained on RoboTokens, a unified tokenization of robot embodiments, states, and actions. The same architecture can be used across embodiment spaces (e.g., wheeled bimanual, quadrupeds, humanoids) and use cases (embodiment generation, cross embodiment controller). Rather than overfitting to one reward function, Transformer Transformer is a dynamics model, whose reward-agnostic state and action predictions can be converted into reward-specific value predictions. These value predictions are used to steer embodiment diffusion towards high value robot designs, through a procedure we call Dynamics Self-Guidance. Experiments across multiple design spaces show zero-shot optimization of unseen rewards and trajectories, improving performance and runtime over the evolutionary baseline. Finally, we fabricated an optimized ALOHA design, which reduced tracking error by over 70% compared to the original design.

The rapid advancement of video generation models has led to the increasing misuse of image-to-video (I2V) models. Although substantial progress has been made in detecting AI-generated videos, proactive defenses against I2V models remain underexplored. In particular, current proactive defenses against I2V models predominantly rely on gradient-based adversarial attacks, which require defenders to possess GPUs with substantial memory resources (VRAM) to generate adversarial examples. To address this issue, we propose I2VShield, a privacy protection method based on generative adversarial attacks tailored to Diffusion Transformer (DiT)-based I2V models. The proposed method primarily consists of two components: (1) a text-adaptive perturbation generation framework integrating adversarial learning to mitigate computational overhead while maintaining visual imperceptibility; and (2) an untargeted Multimodal Attention Disruption (MAD) attack that exploits the inherent vulnerabilities of DiT-based I2V models, maximizing the deviation of the internal attention features from their clean states. Extensive experiments demonstrate that our approach achieves highly competitive protection performance across various datasets and mainstream DiT-based I2V models, particularly in disrupting spatiotemporal coherence, while substantially reducing computational costs.

Stereo depth estimation for driving, robotics and augmented reality must run at high resolution under tight latency budgets, yet in transformer-based matchers the global self-attention that aggregates scene context grows quadratically with the number of pixels and comes to dominate runtime. We show that the joint self-attention stage of a stereo transformer, whose role is to spread context across both views, can be replaced by a data-independent Walsh-Hadamard token mixer that mixes tokens globally in the transform domain at log-linear cost, while the data-dependent cross-attention that performs left-right correspondence is retained. On synthetic driving data the mixer matches the attention baseline in end-point error while reducing model compute by a factor of 2.46 and single-image inference latency by a factor of 2.65. A complexity analysis shows the benefit is governed by the ratio of sequence length to channel width, which explains why high-resolution stereo matching is a particularly favorable setting and why classification transformers are not; we confirm this token-to-channel scaling on non-stereo long-sequence benchmarks. Furthermore, we introduce a hybrid log-disparity loss function designed to up-weight small-disparity pixels corresponding to long-range objects. This approach reduces the error on distant objects without incurring any additional computational overhead.

Modern large language models scale successfully by pairing capacity growth with efficiency, keeping per-token and deployment costs under control as capacity grows. AIGC Foundation Models (AFMs), especially diffusion-transformer backbones, have begun to adopt sparse experts, but recent efforts mostly enlarge total parameter counts and sparsity ratios without importing the efficiency mechanisms that made LLM scaling practical, so generation quality is seldom balanced against training and deployment cost. This raises a natural question: can the architectural principles behind efficient LLM scaling be adapted to AFMs in a more balanced way? We introduce ModernMOE (MMOE), a modernization of SiT-style diffusion transformers that systematically adapts routed experts, shared and lightweight experts, gate-residual routing, and attention-residual information reuse to AIGC generation. Rather than treating MoE as a single plug-in replacement, MMOE studies how different modern expert components affect convergence, efficiency, and generation quality when composed inside a diffusion transformer. Every experiment in this paper is trained on a single eight-GPU H100 node with batch size 256 for 400k steps, an accessible single-machine budget. Under matched training and sampling protocols and at this budget, MMOE reaches lower FID at every recorded checkpoint, that is, it converges faster per training step, than dense and intermediate sparse-expert baselines, and among the sparse variants it attains the best quality-cost balance. Routing analysis further shows stable expert specialization across depth, substantial use of lightweight routes, and modest step-to-step routing changes during denoising. These results suggest that AFMs can follow the balanced scaling path of LLMs by importing proven efficiency designs, rather than by simply increasing total parameters and sparsity ratios.

We introduce EgoPlay, an event-triggered video-to-video editor for egocentric streams, obtained by fine-tuning a pretrained V2V diffusion transformer on event-conditioned data built primarily from Ego4D. Given a monocular video and an event-triggered prompt of the form "when X happens, do Y," EgoPlay infers whether and when event X occurs, preserves pre-event frames, and applies edit Y only to the post-event continuation. Rather than cascading a separate event detector with an editor, EgoPlay learns event recognition, temporal restraint, and pixel-level editing jointly in a single end-to-end model, while also handling negative and multi-event prompts. To support this, we construct a large-scale dataset of 106K event-triggered clip-prompt pairs spanning positive triggers, fabricated-trigger negatives, and multi-event prompts. We then train a bidirectional video diffusion editor with event-triggered supervision and derive a causal variant for chunk-by-chunk streamable inference. We further introduce an event-aware evaluation protocol that separately measures post-trigger editing quality, pre-trigger preservation, and false-trigger robustness. On the Ego4D benchmark, EgoPlay substantially outperforms EgoEdit, the state-of-the-art instruction-based egocentric video editing baseline, with relative gains of 17.7%, 16.9%, and 16.4% in editing quality, visual quality, and background consistency. It also surpasses a VLM-guided detector-editor baseline by 15.7%, 14.5%, and 13.5% on the same metrics, while using less than half the GPU memory.

The vast chemical design space and complex, interdependent design variables make catalyst discovery for targeted properties highly labor- and resource-intensive. Although generative models have emerged as a promising solution, existing approaches are generally limited to single-property conditioning or narrow chemical spaces. Here, we present Catalyst Diffusion Transformer (CatDiT), a unified framework for inverse catalyst design that generates valid and novel structures ranging from intermetallic alloys to oxide surfaces. By learning compressed latent representations, CatDiT enables efficient training and rapid sampling while supporting simultaneous conditioning on adsorbate type, binding energy, and catalyst class. The model provides reliable control of discrete properties and directional control of continuous properties, enriching candidate pools for reaction-specific catalyst discovery. As a representative application, multi-conditional generation for the nitrogen reduction reaction (NRR) yields 28 density functional theory (DFT)-relaxed alloy candidates that satisfy the target activity window and lie above the pure-metal *N-*H scaling line, corresponding to a ~1.5-fold enrichment over the source distribution. These results establish CatDiT as a practical and scalable approach for property-directed catalyst inverse design and targeted catalyst generation.

Many recent robot policies pursue stronger control by using large pretrained vision-language models (VLMs) as the action backbone. We introduce WorldDiT, a unified diffusion transformer architecture that couples action generation with visual world modeling and achieves strong performance without a large pretrained VLM action backbone. During training, a single diffusion transformer generates continuous action chunks and predicts normalized RGB patch targets from future camera frames. Across four LIBERO simulation suites, WorldDiT lies on the reported Pareto frontier for total model parameters and mean success among methods reporting all four suites. These results provide a strong sub-billion-parameter baseline for future scaling studies.

Fine-grained, device-initiated communication allows fused GPU kernels to issue remote stores directly from their compute pipelines, a pattern increasingly used in expert parallelism (EP), tensor parallelism (TP), and Ulysses-style sequence parallelism (UP). Existing designs reason about where communication is issued and when remote data becomes ready, but lack a quantitative model of the sender-side interval after a remote store is accepted and before it becomes visible at the destination. This interval determines whether communication remains decoupled from computation or backpressures it. We identify X-Stage, a software-visible post-issue stage with finite decoupling. Downstream pressure can dissipate while the issuer resumes useful work, whereas sustained injection consumes X-Stage headroom and eventually stalls the compute pipeline. We characterize this behavior and build a calibrated model that predicts whether remote-store arrivals accumulate backpressure or recover during intervening computation. Guided by the model, we reshape bursty arrivals when they would exhaust X-Stage headroom and exploit natural compute windows when headroom can recover concurrently. Evaluation across representative EP, TP, and UP workloads shows up to 1.62x fused-kernel, 1.75x end-to-end, and 1.43x sender-visible speedup, respectively. Microbenchmarks further validate the model's predictions of backlog accumulation, recovery, and sender-side backpressure.

Video super-resolution (VSR) using large-scale Diffusion Transformer (DiT) priors achieves exceptional perceptual quality but is often impractical due to the quadratic computational cost of processing dense spatio-temporal token sequences. Existing efficiency-oriented methods risk irreversible detail loss and temporal flickering, a vulnerability especially pronounced in one-step diffusion models. To address this, we propose TRaM-VSR, a Token Routing and Merging framework for adaptive token allocation, leveraging both context-aware video priors and network-level priors. First, token importance is estimated by fusing motion-sensitive temporal cues with semantic text similarity, isolating dynamic objects and structural boundaries. Next, this importance is further calibrated and adjusted by an offline planner to guide routing across optimally grouped network blocks. Technically, within each routed group, structurally critical tokens are processed in a high-fidelity local stream, while less informative tokens are aggregated into a compact global stream, both modulated by network depth and aligned with the multigranular nature of diffusion models. Extensive experiments show that TRaM-VSR accelerates inference significantly while preserving state-of-the-art reconstruction quality and robust temporal consistency. The code is available at https://github.com/Ree1s/TRaM-VSR.

Post-training quantization (PTQ) of diffusion transformers (DiTs) to W4A4 severely degrades output quality, because activations entering each linear layer contain outliers that 4-bit formats cannot represent. The standard fix applies an invertible linear transform to the activations and its inverse to the weights before quantizing both. Normalization layers between blocks force this transform to run online at every denoising step, making its inference computation cost the binding design constraint. Existing options trade quantization quality for inference cost: per-channel scaling (SmoothQuant) is computationally cheap but impacts the magnitude of the channels, which can harm quantization accuracy; fixed Hadamard transforms yield better quantization accuracy but require large block sizes that incur a high online cost; learned full-$d$ invertible transforms calibrate best but entail a prohibitive dense $d \times d$ matrix multiplication (GEMM) per layer per step. We propose KroQuant, a PTQ method that applies a learned Kronecker-structured invertible transform to each 32-element block of the activation, storing less than half the parameters of per-channel scaling. The block-local structure runs as small tensor-core GEMMs, and on an MI350 GPU the KroQuant quantizer kernel is up to $14\%$ faster than the SmoothQuant kernel. Offline LoRaQ weight calibration then absorbs the residual per-weight quantization error. On PixArt-$Σ$, SANA, and FLUX.1-schnell at W4A4 (MXFP4e2), KroQuant produces outputs closer to the FP reference than SVDQuant and LoRaQ on MJHQ-30K and SDCI, while preserving or improving image quality.

Rectified-flow-based diffusion transformers, particularly FLUX, have demonstrated outstanding performance in high-quality image generation. However, achieving fast and accurate inversion--transforming images back to latent noise for faithful reconstruction and editing--remains a challenging bottleneck due to the discretization errors of linear solvers. This paper introduces SlerpFlow, a straightforward yet highly effective zero-shot approach that unlocks the full potential of FLUX for high-fidelity inversion and editing. Unlike existing approaches (e.g., RF-Solver) that rely on complex numerical approximations such as high-order Taylor expansions to correct trajectory errors, we present a geometric view based on the Manifold Hypothesis: the empirically observed trajectory curvature is not a numerical artifact, but rather serves as a necessary "centripetal force" that constrains the flow to remain on the data manifold. Guided by this insight, SlerpFlow integrates Spherical Linear Interpolation (Slerp) to rectify flow velocity directions on the hypersphere, strictly adhering to the intrinsic curvature of the latent space. Crucially, by caching the corrected velocity for subsequent steps, SlerpFlow achieves high-precision inversion while maintaining the computational efficiency of a first-order Euler solver. Extensive experiments on FLUX-based reconstruction and editing tasks demonstrate that SlerpFlow improves reconstruction fidelity and achieves stronger semantic alignment in editing without requiring additional training. Code is available at https://github.com/0answer0/SlerpFlow.

The modeling of hydrometeorological time series with limited observations is a key challenge in the monitoring of hydro-systems and water resources, as well as for flood or drought risk assessment. Due to the high variability of the underlying processes and the sparsity of available measurements, traditional statistical approaches often struggle to accurately represent their dynamics. In this context, recent advances in deep learning offer a promising direction for improving the representation and generation of complex temporal processes sampled at several observation sites. This study investigates the application of transformer-based diffusion models to the simulation and reconstruction of hydrological time series. The proposed framework is applied to the joint modeling of water quantity and quality at six sites spread across three adjacent headwater catchments located in North-East France on a limestone plateau covered by forests and field crops. The model is calibrated and validated using available observational data, which has been quality controlled and corrected for sensor drift and malfunction through collaborative efforts by LNE metrology expertise and Andra monthly quality control over more than 15 years. Its performance is compared with several established baseline approaches commonly used for time series modeling. Quantitative evaluation metrics are employed to assess the ability of the proposed method to reproduce key temporal characteristics of the observed signals in two settings: the imputation of incomplete time series and the forecasting of upcoming hydrological conditions. Results support the effectiveness of the transformer-based approach and highlight its capacity to capture and simulate the complex patterns present in hydrological data. In particular, the results indicate that diffusion models can efficiently sample realistic time series distributions under observation settings with variable missing data for both forecasting and imputation.

Radio map (RM) estimation aims to reconstruct the spatial distribution of wireless signal characteristics, such as received signal strength (RSS), from sparse measurements, a task that is critical for spectrum management, interference mitigation, and localization in modern wireless networks. Traditional approaches, including interpolation and deep learning, either struggle to capture complex propagation effects or require large-scale retraining for each new sampling pattern, which limits their generalization. More recently, prior-based methods have combined pre-trained generative models with measurements to reduce the need for deployment-time model fine-tuning, but they typically treat the prior as a simple regularizer and lack explicit transmitter-aware integration. In this paper, we propose RadioTrace, a novel RM estimation framework without deployment-time fine-tuning that tightly integrates sparse RSS measurements with a frozen pre-trained diffusion prior. RadioTrace incorporates transmitter (Tx) location estimation directly into the denoising loop, iteratively refining Tx coordinates based on reconstruction quality to guide the generative process. To further enhance robustness, we introduce a propagation-guided K-means initialization that mitigates poor local minima in the Tx update and provides a geometry-consistent starting point. Moreover, we provide a stochastic stability analysis for the Tx-coordinate refinement component, showing that the Tx update remains stable under perturbations induced by diffusion sampling and Tx-map relaxation. Extensive experiments demonstrate that RadioTrace achieves competitive performance with state-of-the-art learning-based methods under random sampling, and maintains strong reconstruction quality under restricted-area sampling, highlighting its adaptability, robustness, and practical relevance.

Controllable image generation remains challenging for creative professionals, who often require precise regional control over materials, object identities, and spatial arrangements that cannot be reliably achieved through text prompting alone. Diffusion Transformers (DiTs) can natively ingest heterogeneous tokens stemming from texts and images, but they lack mechanisms for determining where and how these tokens should influence the output. We introduce appearance pointers, compact tokens that guide DiTs toward the correct appearance cues at the correct spatial locations by aligning text or image inputs with user-specified masks. Appearance pointers are produced by a region correspondence network and refined through a spatial aggregation mechanism, enabling the model to handle multiple regional descriptions without significantly increasing token load. Our approach introduces the first modality-agnostic interface for localized multimodal control in a DiT without retraining the base model from scratch. Across a range of metrics, our single model reaches or surpasses the performance of modality-specific state of the art methods, offering a simple and extensible path toward precise, region-aware, multimodal guidance in generative image synthesis.

Modern text-to-image diffusion transformers (DiTs) generate images through joint attention, in which text and image tokens interact directly within a single sequence. In large-scale DiTs, the conditioning input contains not only the user prompt but also chat-template tokens introduced by LLM-based text encoders. Yet how these tokens participate in the denoising computation remains poorly understood. To probe this, we introduce a causal interpretability framework. Using it to separate prompt-content tokens from chat-template tokens, we find that the template tokens carry little prompt-specific information at the encoder output. Yet surprisingly, they emerge as dominant image-to-text attention sinks and causally maintain object identity inside the DiT, acting as implicit semantic registers. We show that they acquire this identity indirectly. Rather than reading the prompt tokens, they draw the identity from the image latents into which the prompt semantics have already been injected at the very first layer. We further reveal a division of labor across heads and depth in DiTs, where distinct heads route semantics or render visual structure, and identity is committed in early blocks, carried by middle blocks, and refined in late ones. As a practical payoff, this analysis yields a training-free pruning rule that removes the causally inert prompt-reading heads and cuts $20\%$ of joint-attention FLOPs at a $1.4$-point cost in GenEval accuracy. Overall, our work not only reveals that the tokens encoding semantics at the input need not be those that maintain them during generation, but also provides a causal view of internal mechanisms in diffusion transformers.

Large-scale visual generators are increasingly capable but costly to train, fine-tune, and deploy. We introduce Mage-Flow, a compact 4B-scale generative stack for efficient text-to-image generation and instruction-based image editing. The stack is built from two co-designed components: Mage-VAE, a lightweight high-fidelity latent tokenizer, and a Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching. Mage-VAE uses one-step diffusion-style encoding and decoding with anchor-latent regularization, preserving the reconstruction quality of strong public VAEs while reducing tokenization cost by more than an order of magnitude. Together with native-resolution packing and stack-level CUDA kernel fusion, the stack supports flexible-resolution training and improves end-to-end training throughput by about $2.5\times$. Built on this foundation, we develop a complete model family with Base, RL-aligned, and Turbo variants for both generation and editing. Diffusion-NFT improves prompt following, text rendering, aesthetic quality, and editing fidelity, while few-step distillation with adversarial perceptual guidance produces 4-step Turbo models for low-latency inference. Despite its compact scale, Mage-Flow and Mage-Flow-Edit achieves competitive performance across standard generation and editing benchmarks. More importantly, the Turbo variants make high-resolution generation and editing practical for interactive use: at $1024^2$ resolution on a single NVIDIA A100 GPU, Mage-Flow-Turbo generates an image in 0.59s, and Mage-Flow-Edit-Turbo edits an image in 1.02s, while maintaining a small memory footprint. These results show that careful tokenizer--backbone--system co-design can deliver strong high-resolution generation and editing within an efficient 4B model family.