Multimodal Diffusion Transformers

Recent advances in diffusion models have shown impressive performance in controllable image generation and dense prediction tasks. However, existing approaches typically treat diffusion-based controllable generation and dense prediction as separate tasks, overlooking the potential benefits of jointly modeling the heterogeneous distributions. In this work, we introduce UniGP, a framework built upon MMDiT, which unifies controllable generation and dense prediction through simple joint training, without the need for complex task-specific designs or losses, while preserving the backbone's versatile priors. By learning controllable generation and prediction under different conditions, our model effectively captures the joint distribution of image-geometry pairs. UniGP is capable of versatile controllable generation, dense prediction, and joint generation. Specifically, the proposed UniGP consists of DUGP and a unified dataset training strategy. The former, following the principle of Occam's razor, uses only a copied image branch of MMDiT to model dense distributions beyond RGB, while the latter integrates heterogeneous datasets into a unified training framework to jointly model generation and perception tasks. Extensive experiments demonstrate that our unified model surpasses prior unified approaches and performs on par with specialized methods. Furthermore, we demonstrate that multi-task joint training provides complementary benefits: generative priors enrich perceptual details, while perceptual learning improves structural alignment in generation.

High-resolution Diffusion Transformer (DiT) inference contains substantial spatial redundancy, but many spatially adaptive implementations encode regional computation as attention masks, which can inadvertently move scaled dot-product attention (SDPA) away from FlashAttention fast paths. We identify this avoidable systems bottleneck as Mask-Induced Dispatch Tax (MIDT) and show that it grows with latent sequence length. We introduce SAFE-DiT, a training-free Semantics-Aware Fast-path Execution framework that separates exact mask elision from approximation-based spatial scheduling. SAFE-DiT removes only provenance-certified image self-attention masks that induce a row-wise constant shift in attention logits, preserves semantics-bearing masks such as text-padding masks, and realizes spatial adaptation through prompt-conditioned token partitioning, selective state updates with global context, and periodic context refresh. We call this acceleration-only configuration SAFE-Core and report sensitivity-weighted classifier-free guidance separately as SAFE-DiT+SW. On the evaluated PyTorch SDPA stack, redundant masks make long-sequence attention $4.1\times$ to $5.8\times$ slower than the mask-free path. On Lumina-Next, SAFE-DiT achieves $2.69\times$ end-to-end acceleration at $1024^2$ resolution and $5.09\times$ at $2560^2$, reduces peak memory at $2560^2$ from 94.1 to 27.9 GB, and enables $3072^2$ generation when dense inference runs out of memory. Paired metrics, component ablations, and a blinded human study support visual non-inferiority of SAFE-Core to the dense fast-path baseline, while SAFE-DiT+SW provides a separate prompt-alignment operating point without reintroducing spatial self-attention masks. Code is available at https://github.com/xuanhuayin/SAFE-DiT.

End-to-end pixel-space diffusion models bypass the lossy compression of Latent Diffusion Models (LDMs) but struggle to jointly model low-frequency semantics and high-frequency signals in high-dimensional space. Existing works heavily rely on complex pixel decoders to alleviate this issue. In this paper, we challenge this trend by revealing that these decoders primarily compensate for the optimization difficulties inherent to velocity prediction ($v$-prediction). Under the clean data paradigm ($x$-prediction), they are redundant. Motivated by this insight, we advocate for simplicity over complexity and introduce PixelU, a minimalist, single-stage U-shaped Diffusion Transformer tailored for pixel space. PixelU abandons auxiliary decoders in favor of zero-cost skip connections, which provide an "information highway" that directly routes uncorrupted high-frequency spatial details from shallow to deep layers. To further enable the backbone to focus exclusively on modeling low-frequency semantics, we introduce a constant-channel spatial down-sampling mechanism as a natural low-pass filter, which compresses deep features into a compact, low-frequency semantic manifold. Extensive experiments demonstrate that this decoupling of frequencies could outperform the strong baseline (JiT-G) with only about 1/3 of its computation cost. On ImageNet 256$\times$256 and 512$\times$512, PixelU achieves FID of 1.63 and 1.92 respectively, surpassing recent pixel-space methods and establishing a simple yet powerful new paradigm for end-to-end diffusion models.

Diffusion Transformers (DiTs) have driven substantial progress in image and video generation but suffer from prohibitive computational costs. Feature caching accelerates inference by reusing intermediate representations. Existing methods rely on historical features for implementation simplicity, yet suffer from severe error accumulation at high acceleration ratios. To address this limitation, we investigate the nature of the requisite feature correction. We demonstrate that the optimal calibration update is characterized by a shared low-rank subspace across diverse prompts. Guided by this structural insight, we propose LearniBridge, a learnable calibration mechanism for feature caching that bridges multiple timesteps through lightweight LoRA updates. This mechanism enables effective calibration requiring only 3-5 training samples. Extensive experiments on image and video generation show that LearniBridge achieves up to $5.87\times$, $5.75\times$, and $4.10\times$ acceleration on FLUX, HunyuanVideo, and WAN2.1, respectively. On WAN2.1, it improves VBench by 1.28% over the previous SOTA at $4.10\times$ acceleration. Our code is available at https://github.com/Iiiiiiirene/LearniBridge.

Diffusion transformer (DiT) research on image generation has converged to a single evaluation setup: class-conditional generation on ImageNet. While methods improve the FID and related metrics, it is increasingly unclear whether they reflect real progress in generative modeling. The natural alternative, i.e., text-to-image (T2I) generation, is perceived as too costly or inconvenient to train and evaluate and is often skipped. We argue that this perception no longer holds. We introduce NanoGen, a unified DiT training and evaluation framework. NanoGen matches state-of-the-art DiT baselines on ImageNet and, with 12 lines of configuration change, also trains competitive text-to-image models. It currently supports RAE, VAE, pixel-space, and MeanFlow diffusion methods under both ImageNet and T2I setups. Under NanoGen, training T2I requires comparable compute to ImageNet. After training 21 latent diffusion models with NanoGen, we observe that method ranking shows no strong correlation between ImageNet and T2I generation: Pearson correlation is between -0.377 and -0.580 across three metrics. This suggests that a method which improves class-conditional ImageNet FID may show no corresponding improvement on T2I, clearly indicating the necessity of evaluating DiTs on both tasks. To this end, we summarize ImageNet and text-to-image results, which yields DiffusionBench, a holistic benchmark for DiT research. We recommend reporting DiffusionBench in place of ImageNet alone: methods that improve DiffusionBench are more likely to reflect broader progress.

Sparse voxel representation has emerged as a scalable foundation for image-to-3D Gaussian Splatting (3DGS) generation, yet current methods struggle to preserve high-frequency visual details of input images due to two structural bottlenecks. First, they adopt discriminative 2D features optimized for semantic abstraction to construct sparse voxel latents, which suppress reconstructive cues and induce a representation bottleneck. Second, in the generation stage, standard diffusion transformers lack effective mechanisms to align dense 2D image tokens with sparse 3D voxel latents, resulting in a cross-modal correspondence bottleneck. To address these issues, we propose FLUX3D, a scalable image-to-3DGS framework that boosts both representation learning and cross-modal alignment during generation. We first revisit 2D feature selection for sparse-voxel-based 3D representation learning, propose Diffusion-Aligned Structured Latents (DA-SLAT) and couple it with a decoder-only architecture to improve 3DGS reconstruction fidelity. We also design a sparse-structure-aware diffusion framework, which integrates the Sparse-structure Multimodal Diffusion Transformer (SMDiT) and Modal-Aware Rotary Positional Embedding (MARoPE) to achieve geometry-agnostic 2D-3D alignment. Extensive benchmark experiments demonstrate that FLUX3D yields substantial improvements in appearance fidelity and significantly outperforms all state-of-the-art (SOTA) methods in generating high-quality 3DGS assets.

Can representations learned for image generation also support the evaluation of generated images? We study text-to-image reward prediction as a downstream task of generative representation learning. To this end, we introduce DiT-Reward, which converts a pretrained text-to-image Diffusion Transformer into a reward model by processing near-clean image latents and aggregating text-conditioned image representations across transformer layers. Under the same training data mixture as HPSv3, DiT-Reward outperforms HPSv3 on all four evaluated preference benchmarks, reaching 85.6% on HPDv2 and 77.6% on HPDv3. When the generative backbone is frozen, a lightweight learned head can still extract meaningful preference predictions from its representations. Probing across depth further reveals that downstream reward performance is strongest in the middle-to-late layers and benefits from combining representations across different stages. We also observe consistent positive scaling with generative backbone capacity. Finally, when used to optimize Stable Diffusion 3.5 Large with Flow-GRPO, DiT-Reward outperforms HPSv3 along the matched training trajectory, with particularly clear gains in realism. Direct latent scoring also achieves a 1.65x inference speedup over HPSv3 with comparable peak memory. These results show that pretrained generative DiTs provide transferable representations for reward modeling and policy optimization.

Realistic integration of user-specified textures into scene images is a fundamental task in computer graphics and image editing. While existing material transfer and reference-guided inpainting methods can edit surface appearances, they often fail to address the specific requirements of texture tiling. This task necessitates precisely repeating a reference pattern according to user-defined parameters such as frequency, orientation, and scale. Furthermore, current generative approaches often struggle to maintain the structural fidelity of the reference texture, limited by either destructive pixel-level resampling or the lack of fine-grained spatial information in semantic image encoders, and they frequently fail to preserve the coherent lighting and geometry of the original scene. In this paper, we propose a novel framework for controllable and high-fidelity texture tiling based on Diffusion Transformers. Our approach introduces two key technical innovations to decouple spatial manipulation from content generation. First, we propose a Coordinate-Transformed Rotary Embedding mechanism. By applying 2D affine transformations directly to the relative positional embeddings between the target latent and the image condition, we achieve precise control over tiling patterns without explicit pixel warping, thereby utilizing the full information of the reference condition without degradation. Second, a Disjoint Attention Mask is employed to shield reference features from semantic leakage. This preserves structural integrity while seamlessly blending the synthesized texture with the scene's original lighting and geometry. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines in both control accuracy and texture fidelity.

Choreographic motion generation poses unique challenges for AI, demanding precise semantic control over complex, temporally structured, and expressive full-body dynamics. While existing models can synthesize motion from music, they remain largely black boxes. Conversely, attempting to condition generation on both text and music frequently leads to modality collapse, where dense acoustic rhythms overwhelm sparse semantic text prompts, destroying user controllability. To resolve this spatial-temporal conflict, we propose STREAM (Structural-Temporal Rhythmic Energy-based Attention for Motion), a modality-decoupled diffusion transformer. STREAM strictly separates conditioning pathways: global text semantics dictate the kinematic structure via Adaptive Layer Normalization (AdaLN), while a novel Bimodal Energy-Based Attention Module (BEAM) routes these features to the musical beat without overwriting the semantics. We further introduce Motorica++, a newly curated dataset enriched with domain-specific dance vocabulary and frame-level semantic annotations from existing Motorica dataset. Additionally, to rigorously quantify zero-shot editability, we propose the Exchange Evaluation Protocol and Editable Dance Score (EDS). Through extensive experiments, STREAM achieves state-of-the-art alignment between motion and music while perfectly preserving choreographic semantics, positioning AI not merely as a reactive synthesizer, but as a controllable, collaborative partner for artistic direction. The source code and datasets are available at https://github.com/SeongJong-Yoo/STREAM.

Localizing document tampering is extremely challenging, as manipulations are crafted to appear visually consistent and often leave only subtle traces that are nearly invisible to the human eye. In prior work, evaluation has been largely dominated by synthetic benchmarks that closely match the training distribution, and methods have shown steady progress under this setting. However, these gains often translate poorly to human-made forgeries and to cross-domain evaluation, where both the source documents and the tampering pipeline can change, leading to a distribution shift. In addition, since the introduction of the Frequency Perception Head for the discrete cosine transform (DCT) modality, it has become a standard choice, and subsequent work has largely focused on downstream modules and fusion strategies rather than revisiting the backbone itself. To help close this gap in cross-domain performance and improve the DCT backbone design, we propose \textbf{DiffNet}, a relatively simple yet effective RGB--DCT early-fusion architecture driven by two key design choices. First, to ensure that the decoder aggregates multi-scale inconsistency evidence rather than operating on raw, content-heavy activations, we apply a lightweight multi-level discrepancy transformation at the output of each backbone stage, replacing features with magnitude-only responses to learned zero-sum filters. Second, we design an efficient DCT-domain backbone that relies on a lightweight frequency-index-aware DCT--quantization joint embedding. Our approach achieves state-of-the-art performance on cross-domain and human-made document tampering localization, outperforming prior methods by around 30%, with up to $7\times$ higher throughput than the previous best model.

Audio editing aims to modify specific content in an existing audio clip according to a text instruction or description while preserving the remaining acoustic content. Despite the remarkable progress of diffusion models, existing training-based editing methods mainly rely on the local inductive biases and cross-attention interaction in convolutional U-Net backbones, which often hinder long-range semantic alignment and precise understanding and localization of instructions. In contrast, diffusion transformers provide stronger global modeling and multimodal fusion, but existing editing architectures usually adopt a simple stack of diffusion transformer blocks. Applying joint attention over concatenated audio and text tokens in all blocks results in quadratic complexity with respect to token length. To balance editing performance and efficiency, we propose a novel instruction-guided audio editing framework based on rectified flow matching (RFM), named RFM-Editing 2, built on a hybrid two-stage diffusion transformer. The proposed model performs joint attention over audio and text tokens to establish coarse semantic alignment at the low-resolution stage, then switches to alternating joint-attention and cross-attention blocks to refine editing details at the high-resolution stage. This coarse-to-fine strategy enables efficient and accurate instruction-guided audio editing. Experiments show that the proposed framework achieves notable performance gains on challenging editing tasks involving overlapping audio events and complex instructions, while substantially improving editing efficiency.

Latent Diffusion Models (LDMs) have become dominant in visual synthesis, but their quality-compute trade-off is largely constrained by the tokenizer's fixed compression ratio. Variable-length tokenizers (VLTs) promise adaptive compression by varying token counts, allowing diffusion models to flexibly balance quality and compute. However, conventional VLTs modulate length by truncating ordered token sequences, which makes token semantics depend on token position and breaks representational alignment across lengths. This leads to a cross-length shift in the latent distribution that hinders a single variable-length diffusion model from operating effectively. To address this, we propose a novel variable-length tokenizer that modulates length by merging tokens. We show that encouraging similar tokens to merge enables direct cross-length representation alignment when the diffusion transformer operates according to the merging pattern. Since conventional merging methods are data-dependent, making the merging pattern inaccessible during generation, we introduce learnable global merging, which is data-independent, to ensure compatibility with diffusion transformers. On ImageNet 256$\times$256 generation, our merging-based variable-length tokenizer integrated with a diffusion transformer achieves a superior gFID-compute trade-off compared to prior VLT methods. Code is available at this https URL

Three-dimensional (3D) brain MRI is central to clinical neurology and neuro-oncology, where generative models could augment under-represented cohorts, simulate disease trajectories, and support privacy-preserving data sharing. Latent diffusion has been the go-to solution for modeling imaging data, but it places two competing demands on the tokenizer: encoder embeddings must retain the clinical information that downstream tasks act on, and the decoder must reconstruct anatomically faithful volumes. Existing reconstruction-driven tokenizers achieve the second at the expense of the first. To address this, we introduce a fully volumetric masked-autoencoder (MAE) based tokenizer for 3D brain MRI latent diffusion, decoupling encoder and decoder: a frozen 3D MAE encoder produces clinically informative embeddings, while a dedicated CNN decoder reconstructs voxels from a linear projection of those embeddings. We pretrain the encoder on 35,309 volumes from 18 public cohorts spanning four modalities, ten disease categories, and 200+ acquisition sites, and demonstrate its dual utility in two settings. First, on a 23-task linear-probing benchmark, the encoder outperforms or matches SOTA models (i.e., BrainIAC, BrainSegFounder, and MedicalNet) on 21 of 23 tasks. Second, a conditional diffusion transformer (DiT) trained on these clinically informative embeddings supports both conditional generation across six variables and patient-specific longitudinal forecasting. Together these results establish a single 3D brain-MRI embedding space capable of both downstream clinical tasks and controllable generation.

Reinforcement learning (RL) post-training of Diffusion Transformers (DiTs) is prohibitively expensive, requiring thousands of high-end GPUs. Existing works explore two directions to reduce cost: seed exploration improves training convergence by selecting high-contrast samples, yet adds compute to the critical path; spot GPUs offer 69--77% lower cost, yet sit idle during training because DiT rollouts finish nearly simultaneously, which prevents LLM-style pipelining of rollout with training. Spot preemptions further break Sequence Parallelism (SP) groups, fragmenting GPU topology. We present Spotlight, the first system that harvests spot GPUs for DiT RL post-training. Spotlight rests on two key insights we devise: (1)~we show that exploration can tolerate stale model weights because exploration that uses the model weights from the previous iteration preserves the relative ranking of random seeds, allowing exploration to run on idle spot GPUs during training. (2)~SP reconfiguration can reuse on-node state, reducing group recovery from minutes to sub-second launches. Built on these insights, Spotlight introduces three techniques: a bandit-based exploration planner that maximizes reward variance within the training time budget, elastic sequence parallelism that reconfigures SP groups on the fly via persistent schedulers and intra-node weight copying, and a preemption-aware pull-based request scheduler that balances load and commits in-flight state upon preemption. We implement Spotlight on the open-source RL platform ROLL and evaluate it on Qwen-Image post-training. Spotlight reaches the same target validation score $4\times$ faster than baselines, reducing total cost by $1.4$-$6.4\times$ while achieving superior image quality on DeepSeek-OCR and Geneval datasets with resolution $512\times512$ and $1280\times1280$.

Diffusion transformers have demonstrated remarkable generative capabilities, yet the rich perceptual representations computed across their denoising trajectory are discarded once the content is rendered. We present MMDiff, a framework that transforms a frozen diffusion transformer into a multi-modal generative system that jointly produces images alongside any combination of dense perceptual modalities using lightweight decoder heads. Our central finding is that perceptual information is temporally distributed along the denoising trajectory, and that multi-timestep feature fusion with spatially varying aggregation weights is essential, improving semantic segmentation results by up to 28.7% mIoU over single-timestep extraction. We further adopt concept-driven attention extraction for interpretable spatial guidance, and show that frozen diffusion features are competitive with and complementary to state-of-the-art encoders such as DINOv3. By training only lightweight decoder heads on a frozen backbone, we achieve strong performance in semantic segmentation, salient object detection, and depth estimation, and demonstrate that this framework enables effective synthetic data generation at scale.

Unified Multimodal Models (UMMs) have emerged as a critical direction for general-purpose multimodal intelligence, integrating understanding and generation into a single framework. However, existing UMMs face prominent challenges: (1) the inherent learning conflicts between visual understanding and generation tasks, leading to suboptimal modeling in both tasks; (2) different understanding and generation visual spaces impeding scalability; (3) over-reliance on task-specific data that neglects the duality of text-image understanding and generation. To address these challenges, we propose UniDDT, which leverages a Noisy ViT encoder along with an LLM to unify semantic encoding for visual generation and understanding tasks, while employing a separate diffusion decoder to decouple diffusion decoding from text decoding. With this Noisy ViT encoder, UniDDT is able to leverage the latent space as a unified visual representation, enabling seamless compatibility between understanding and generation tasks. Thus, the scalability within the generation tasks and the semantic expressiveness within understanding tasks can be balanced. Also, we construct dual data structures from the same image-text pairs, fostering interdependence between the generation and understanding data to exploit their inherent duality. Extensive experiments demonstrate that UniDDT achieves effective unification of multimodal understanding and generation with enhanced semantic consistency and scalability. For visual generation tasks, our UniDDT achieves 0.87 GenEval score and 86.9 DPG overall score. For multimodal understanding tasks, our UniDDT achieves 1699.5 score on MME benchmark and 76.5 overall score on SEEDbench.

Diffusion models excel in Real-World Image Super-Resolution (Real-ISR) due to their powerful generative priors but suffer from slow iterative sampling. Although existing one-step distillation methods accelerate inference, they typically require auxiliary teacher models that inflate training memory and restrict scalability to large-scale architectures. Furthermore, these fixed-step models lack the flexibility to trade off speed for quality. In this paper, we propose TEASR, a training-efficient any-step diffusion framework for Real-ISR that enables both one-step and multi-step restoration within a unified model. Our key idea is to perform self-adversarial distillation within a single diffusion model, eliminating the need for auxiliary teachers or discriminators. Specifically, we propose a timestep-aware rectification strategy that stabilizes one-step generation across noise levels. These two designs further enables the distillation of 20B-parameter diffusion models on a single GPU, significantly improving training efficiency. Moreover, we introduce a dual-branch diffusion transformer with decoupled timestep condition to separate the current noise state and the denoising target to enhance sampling quality. Extensive experiments demonstrate that TEASR supports seamless any-step sampling and consistently outperforms state-of-the-art methods across multiple datasets.

Mechanistic interpretability seeks to explain neural network behavior by decomposing model computations into interpretable features and circuits. While transcoder-based circuit tracing has recently enabled detailed causal analyses of large language models, multimodal diffusion transformers for image generation remain comparatively opaque. We still lack tools for understanding how semantic information propagates across denoising steps and how text and image representations interact within double-stream MM-DiT architectures. Existing methods provide only partial insight: attention maps expose a limited view of token interactions, while sparse autoencoders can discover interpretable features but do not directly reveal how these features are transformed and composed through nonlinear MLP layers. In this work, we extend transcoder-based circuit tracing to multimodal diffusion transformers. We train timestep-conditioned transcoders that faithfully approximate the input-output behavior of MLP sublayers in FLUX.1[schnell]. By replacing MLPs with transcoders and linearizing the remaining computation, we obtain exact feature-to-feature attribution and recover compact, interpretable circuits. Empirically, our transcoders match or slightly outperform sparse autoencoders on the sparsity-faithfulness tradeoff. The resulting circuits reveal mechanisms underlying attribute binding and cross-stream semantic propagation, and provide causal explanations for systematic generation errors. Moreover, circuit-guided interventions are substantially more precise and effective than standard SAE-based steering. Our results demonstrate that transcoder-based circuit analysis is feasible for state-of-the-art diffusion transformers and provides a powerful framework for understanding and controlling multimodal generative models. The code is available at https://github.com/Artalmaz31/DifFRACT

Diffusion Transformers with Mixture-of-Experts (DiT-MoE) improve model capacity under sparse activation, but diffusion inference is still bottlenecked by redundant computation across timesteps. Existing caching methods mainly operate at the token level, which becomes suboptimal in DiT-MoE because each token update is internally decomposed into multiple routed expert branches. Our analysis shows that cross-timestep redundancy in DiT-MoE is better characterized at the expert-branch level than at the whole-token level. Based on this observation, we propose MoECa, a fine-grained caching framework that performs branch-level feature reuse across timesteps. MoECa further introduces expert-aware adaptive control and synchronized cache updates across MoE and attention paths to maintain stable intermediate states. Experiments on multiple DiT-MoE models show that MoECa consistently achieves a better speed-quality trade-off than prior caching methods, with up to 2.83$\times$ inference speedup and minimal quality degradation.

The success of generative models in language and visual generation has inspired extensive applications to generative robot planning. However, most existing works either focus on single-robot planning, or generate multi-robot trajectories in a sequential manner with iterative post-processing to resolve inter-robot conflicts. In this work, we investigate whether coordinated multi-robot trajectories, as a special spatiotemporal distribution, can be learned and generated with a generative model in a feed-forward manner. We propose Robots as Tokens (Roken), a unified diffusion transformer that directly generates multi-robot trajectories that satisfy both (individual) safety and (global) connectivity constraints. The core design of Roken is to represent each robot as a discrete token, allowing them to naturally interact with each other through self-attention, and cross-attend to map tokens for environment layouts. We further introduce several auxiliary tasks based on Bayes' theorem to provide multi-scale spatial-temporal supervision for efficient learning of the conditional distribution. In training, Roken absorbs diverse expert trajectories from different team sizes. During inference, Roken behaves as a versatile multi-robot planner that can handle single-robot planning, coordinated multi-robot trajectory generation, and conditional trajectory generation by fixing some robot tokens as conditions. Experiments in diverse cluttered environments show that Roken can generate coordinated multi-robot trajectories to perform connectivity-constrained goal navigation tasks with high success rates, outperforming the baseline method used to generate the training dataset. Roken also demonstrates good scalability after training with mixed team sizes, and shows generalization to unseen or partially observed environments, verifying its potential to learn from diverse data and perform versatile tasks.

Post-training INT8 (W8A8) quantization of diffusion transformers is widely deployed as a speed optimization, yet on consumer Ampere GPUs it is frequently slower than the FP8 and NF4 alternatives it is meant to beat. We trace this to a software artifact: the production "INT8" forward quantizes weights and activations only to immediately dequantize them back to bf16 and run a bf16 matrix multiply, never engaging the GPU's INT8 tensor cores, so the hardware's compute advantage is left entirely unrealized. We close this gap with a single fused Triton INT8 GEMM (int8xint8->int32 on Ampere tensor cores, with per-token x per-channel dequantization and bias folded into the epilogue, autotuned per GEMM shape) dropped into the Ideogram 4.0 diffusion transformer's linear layers in place of the dequantize-to-bf16 path. In the kernel, the int8xint8->int32 accumulation is bit-exact against torch._int_mm and the dequantized output matches the reference at cosine similarity 1.0 with no NaNs, running 2.8-4.2x faster than bf16 per GEMM. End to end it delivers a ~1.1x (~9-10%) speedup at 768px, and at 1024px it generates an image in 156.5 s on a single RTX 3090, faster than the single-card NF4 (164.5 s) and FP8 (172.9 s) baselines, at no measurable quality cost on these point estimates (PickScore/CLIPScore). INT8 thus goes from the slowest variant to the fastest, and 1024px becomes single-GPU feasible. The primary speed criterion (beat FP8, by ~9.5%) is comfortably met; the NF4 margin (~4.9%, single-run n=4) is within run-to-run variance we did not quantify and is best read as consistent with meeting the stretch target. We close with an honest deployment map: the win is specific to consumer Ampere, and on A100 and B200 the same kernel loses to those cards' fast native bf16/FP8 paths.

Image-to-3D methods often trade off faithfulness and completeness: depth estimators are anchored to input pixels but stop at the visible surface, while image-to-3D models generate complete shapes that are often misaligned with the input. We introduce World Tracing, a generative pixel-aligned geometry representation that predicts 3D points aligned with observed pixels while completing geometry beyond the visible surface. For each input pixel, World Tracing predicts an ordered stack of camera-space 3D points, where the first layer represents the visible surface and subsequent layers represent front-to-back intersections with occluded surfaces. We instantiate this representation with a world-tracing diffusion transformer, WT-DiT, which treats multiple geometry layers as separate denoising tokens coupled through factorized and global attention. WT-DiT is trained with pixel-space flow matching and a mixed noise schedule that balances visible-surface reconstruction with occluded-geometry generation. World Tracing achieves strong performance on visible-surface reconstruction and complete geometry generation across object, scene, and dynamic benchmarks, outperforming both depth predictors and image-to-3D generators. It also preserves 2D-to-3D correspondence, enabling text-driven 3D scene editing, geometry-conditioned novel-view video synthesis, and training-free integration with textured-mesh generators.

Recent progress in video generation has shown impressive visual synthesis capabilities. However, open-domain customized video generation remains limited by the lack of large-scale, annotated datasets capturing diverse identity-specific attributes. To address this, we introduce PexelsCustom-1M, the first publicly available million-scale dataset for identity-preserving video generation, containing one million curated <identity, text, video> triplets across 8,000+ categories. Leveraging this, we propose CustoMDiT, a parameter-efficient framework that adapts a pretrained multimodal Diffusion Transformer into a customized video generator with only 8% additional learnable parameters. Our method surpasses prior state-of-the-art. However, benchmarks such as DreamBooth cover only 100 classes, which is insufficient for real-world applications. To overcome this, we construct OpenCustom, a new benchmark with 1,000+ categories, created via cross-dataset knowledge fusion from ImageNet and MS-COCO. Extensive experiments confirm the advantages of both our dataset and model. We will open-source the entire ecosystem--including dataset, pipeline, benchmark, and implementations--to support further research.

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

Routine full-disk EUV imaging has been available only since the modern era, such as SOHO and SDO. To extend EUV coronal context into earlier periods, we leverage the multi-decade availability of full-disk \HeI{} observations, whose absorption is modulated by coronal irradiance and magnetic topology and is widely used as a proxy for open-field regions. We present a diffusion-based conditional image translation framework, Coronal Hole-aware Diffusion Model Translator (CH-aware DMT), to reconstruct synthetic SDO/AIA 193 Å EUV images from \HeI{} inputs. The model is trained on temporally co-aligned SOLIS \HeI{} and AIA 193 Å pairs spanning 2011--2015 using a month-based split, where January--October are used for training, November is used for validation, and December for testing. On the held-out test set, the reconstructions preserve dominant full-disk EUV morphology (CC=0.92) and recover CH-related low-intensity structure (CC=0.84). We further assess historical applicability by (1) comparing reconstructed AIA 193 Å morphology with SOHO/EIT 195 Å over 2005--2015; (2) comparing reconstructed AIA 193 Å images generated from KPVT \HeI{} inputs against Yohkoh/SXT soft X-ray observations; and (3) evaluating long-term reconstructed disk-integrated emission statistics against observational EUV series and independent solar activity proxies (sunspot number and F10.7 radio flux over 1974--2015). These results indicate that CH-aware DMT conditioned on \HeI{} can provide a physically plausible synthetic AIA 193 Å coronal proxy for historical studies, supporting multi-decade analyses of large-scale coronal evolution before the direct EUV imaging was available.

Pose-guided text-to-image generation often suffers from limb distortions and feature crosstalk in complex multi-person scenarios. While existing UNet-based adapters struggle with long-range spatial dependencies, emerging Multimodal Diffusion Transformers (MM-DiTs) offer superior global modeling. However, naive signal concatenation in MM-DiTs severely disrupts pre-trained latent distributions. To address this, we propose TrioPose, a native pose-driven framework built upon the SD3.5M architecture. Specifically, we introduce a Triple-Stream Pose-Aware DiT (TSPA-DiT) that treats pose as an independent modality. It employs layer-wise activation and zero-initialized dual-residual injection to smoothly enforce geometric constraints while preserving pre-trained latent stability. To resolve severe multi-instance occlusions, we design a Learnable Relational Bias Mask that categorizes topological connectivity into fine-grained physical states, mapping them into continuous attention soft constraints to effectively decouple inter-instance interference. Furthermore, a Pose-Guided Spatial Loss Weighting strategy modulates the native diffusion objective using heatmap-derived error maps, focusing anatomical supervision strictly on distortion-prone regions. Extensive experiments demonstrate that TrioPose achieves state-of-the-art performance across challenging benchmarks, including Human-Art, CrowdPose, and OCHuman. Notably, it attains an AP of $64.33$ on Human-Art, representing a $30\%$ improvement over prior arts, while setting new standards for visual fidelity and text-image semantic alignment in complex multi-human generation.

Diffusion transformers (DiTs) equipped with multimodal attention (MM-Attn) have become a dominant paradigm for image generation. However, preventing the generation of harmful content remains a critical challenge, particularly in image-to-image (I2I) editing tasks. Existing safety mechanisms are primarily designed for text-to-image (T2I) synthesis or U-Net-based architectures, which limits their effectiveness for unified safety mitigation in DiT-based frameworks. To bridge this gap, we propose Unified Visual Safety Regulator (UVR), a training-free safe generation framework that regulates unsafe semantics in generated images. UVR is grounded in an analysis of attention dynamics from the perspective of information flow in MM-Attn. We identify a task-independent start-up stage, during which unsafe semantics in output patches rapidly emerge and can be accurately localized, followed by task-specific semantic amplification and interference stages, where harmful signals are further propagated and entangled with benign content. Based on these observations, UVR mitigates unsafe generation through unified, targeted attention modulation and explicit restriction of harmful information flow over the identified unsafe output patches. Experiments across various concepts show that UVR achieves state-of-the-art safety performance by achieving 91% and 77% erase rate in image synthesis and editing tasks, while preserving visual quality and fidelity with minimal degradation. Code is available at https://github.com/deng12yx/UVR.

Content-Preserving Style transfer, given content and style references, remains challenging for Diffusion Transformers (DiTs) due to entangled content and style features. With a reverse triplet synthesis pipeline to build a million-scale training set and a dual-branch Style-Content DiT (SC-DiT) that decouples style and content via separate ROPE embeddings and causal masking, we observe that such a one-stage training paradigm on mixed style categories causes semantic styles to dominate, hindering texture style learning, and harming content preservation. To address these issues, we propose Style-CCL, a Multi-Stage Curriculum Continual Learning framework that trains SC-DiT from semantic (easy) to texture (hard) styles, and from clean to synthetic data, with Random Memory Rehearsal across stages to avoid catastrophic forgetting. Extensive experiments demonstrate that our Style-CCL achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.

Modern video diffusion models generate increasingly realistic and temporally coherent videos, motivating their use as candidate world simulators. Yet it remains unclear whether these models internally encode physical structure, or merely reproduce motion patterns seen during training. We study this question by probing video diffusion models along latent trajectories corresponding to real videos with known physical plausibility. To obtain such trajectories, we approximately invert the deterministic sampling process by integrating the learned velocity field backward from a clean video latent to noise, giving access to the model's intermediate states and attention maps. Using these recovered trajectories, we show that physical plausibility is linearly decodable from diffusion transformer states across IntPhys and InfLevel, reaching around 81.27% average accuracy and outperforming dedicated representation-learning baselines such as V-JEPA and VideoMAE. Surprisingly, this signal is absent from the VAE latent input and emerges inside the denoising transformer itself, despite the model not being trained with a self-supervised predictive objective. These findings suggest that physically meaningful representations can arise as a byproduct of generative denoising.

Data-Free Quantization (DFQ) addresses data security concerns by synthesizing samples, without accessing real data. It has garnered increasing attention in the context of Vision Transformers (ViTs), owing to the superiority of the self-attention mechanism compared to classical convolutional operation. However, previous DFQ arts for ViTs often suffer from a distribution mismatch between synthetic samples and input distribution expected by quantized models Q, resulting in the suboptimal performance. In this paper, we propose a novel Masked Attention Alignment approach for Data-Free Quantization of ViTs, named MaskAQ, revealing that: 1) the semantics in the self-attention mechanism is predominantly localized to a sparse subset of patches, called informative regions; 2) the informative regions dominate the mutual information between synthetic samples and Q's outputs. To these ends, we incorporate differential entropy maximum over patch similarity of synthetic samples, to decouple informative regions from noisy background. To couple with varied Q, the informative regions are selected to align full-precision models with Q via a masked attention alignment objective, thus yielding high-quality synthetic samples. Furthermore, a periodic sample refreshing strategy comes up to endow MaskAQ with the capacity to continually adapt to the evolving state of Q throughout the training process, to preserve desirable mutual information with synthetic samples. Extensive experiments verify the merits of MaskAQ over state-of-the-art approaches across multiple backbones and downstream tasks. Our code is available at https://github.com/hfutqian/MaskAQ.

The graphical representation of the brain offers critical insights into diagnosing and prognosing neurodegenerative disease via relationships between regions of interest (ROIs). Despite recent emergence of various Graph Neural Networks (GNNs) to effectively capture the relational information, there remain inherent limitations in interpreting the brain networks. Specifically, convolutional approaches ineffectively aggregate information from distant neighborhoods, while attention-based methods exhibit deficiencies in capturing node-centric information, particularly in retaining critical characteristics from pivotal nodes. These shortcomings reveal challenges for identifying disease-specific variation from diverse features from different modalities. In this regard, we propose an integrated framework guiding diffusion process at each node by a downstream transformer where both short- and long-range properties of graphs are aggregated via diffusion-kernel and multi-head attention respectively. We demonstrate the superiority of our model by improving performance of pre-clinical Alzheimer's disease (AD) classification with various modalities. Also, our model adeptly identifies key ROIs that are closely associated with the preclinical stages of AD, marking a significant potential for early diagnosis and prevision of the disease.

Diffusion transformer (DiT) has been widely adopted in the generative diffusion field, advancing the denoising of query tokens through attention and Feed-Forward (\text{FFN}) layers. FFN actually acts as the key-value vocabulary for decoding visual contents where the value embeds the visual semantical knowledge. We present that focusing on critical query tokens corresponding to more complex details and encouraging the model to improve these tokens is essential for fine-grained visual generation. To this end, we propose FocusDiT, which applies a Masking scheme to focus on critical query tokens that are exclusively fed into FFN. The masked queries can retrieve visual tokens from the FFN vocabularies, and use them to decode their visual details. Extensive text-to-image experiments validate the effectiveness of token masking in enhancing generative performance.

Recent advancements in text-guided audio generation have yielded promising results in diverse domains, including sound effects, speech, and music. However, jointly generating speech with environmental audio remains challenging due to the inherent disparities in their acoustic patterns and temporal dynamics. We propose ImmersiveTTS, an environment-aware text-to-speech (TTS) model that generates natural speech seamlessly integrated within environmental contexts by explicitly modeling cross-modal interactions. Our model builds on a multimodal diffusion transformer and fuses transcript-aligned speech latent with text-conditioned environmental context via joint attention. To enhance semantic consistency, we introduce a domain-specific representation alignment objective tailored to environment-aware TTS, leveraging complementary self-supervised representations from speech and audio encoders. Experimental results show that ImmersiveTTS achieves higher naturalness, intelligibility, and audio fidelity than existing approaches across objective metrics and human listening tests.

Real-time streaming video-to-video editing (V2V) is critical for interactive applications such as live broadcasting and gaming, yet it remains a formidable challenge due to the stringent requirements for temporal consistency and inference throughput. In this paper, we present SANA-Streaming, a system-algorithm co-designed framework for high-resolution, real-time streaming video editing on consumer GPUs, with the following three core designs: (1) Hybrid Diffusion Transformer architecture introduces softmax attention in part of the blocks to improve local modeling capabilities while preserving the efficiency of linear layers. (2) Cycle-Reverse Regularization is a novel training strategy that enforces semantic consistency by predicting source frames from generated content via flow matching, improving temporal consistency without requiring paired long edited videos. (3) Efficient System Co-design combines fused GDN kernels and Mixed-Precision Quantization (MPQ) optimized for the NVIDIA Blackwell (RTX 5090) architecture. By profiling real-world throughput, our MPQ maximizes Tensor Core utilization while maintaining generation quality. The resulting system achieves real-time 1280 x 704 resolution editing at 24 end-to-end FPS on a single RTX 5090 GPU, with the DiT core running at 58 FPS. Experimental results demonstrate that our co-design approach significantly outperforms existing SOTA methods in both temporal coherence and system throughput.

Joint audio-video generation aims to synthesize temporally synchronized and semantically coherent visual-acoustic content. However, existing open-source methods mainly rely on either dual-tower designs with posterior alignment or fully unified tri-modal designs that mix textual context, audio and video in one shared space. The former weakens fine-grained audio-video co-evolution, while the latter couples semantic conditioning with low-level synchronization. To address these limitations, we propose NAVA, a Native Audio-Visual Alignment framework for joint audio-video generation. NAVA is built upon context-conditioned native audio-visual alignment: it first establishes audio-video correspondence in a dedicated interaction space, and then uses external context to condition the joint denoising process. Specifically, NAVA is instantiated with an Align-then-Fuse MMDiT architecture, which transitions from modality-aware audio-video alignment to modality-shared joint denoising. Furthermore, we introduce Timbre-in-Context Conditioning to associate reference timbre cues with corresponding speech spans to achieve controllable speech timbre. Experiments on Verse-Bench and Seed-TTS, together with a user study, demonstrate that NAVA achieves superior video quality, precise audio-visual synchronization, competitive audio quality, and stronger reference-timbre controllability using only 6.3B parameters.

Diffusion Transformers have become a powerful backbone for text-to-image generation, but their layered and cross-modal generation process makes safety control fundamentally different from prompt-level filtering or output-level detection. Harmful semantics may be weakly expressed in text representations, progressively bound to visual latents, and finally entangled with rendering dynamics. As a result, safety steering at a fixed layer can be unstable, and a steering mechanism learned from known risks may not transfer reliably to a shifted target risk domain. We propose SafeDIG, a safety steering framework that formulates DiT safety adaptation as position-aware sparse feature transfer. SafeDIG first constructs Sparse Autoencoders over functionally distinct DiT intervention positions and uses robustness-aware pre-training routing to prioritize intervention sites that are expected to remain stable under source-target risk shift. It then separates transferable safety features from domain-specific activation geometry by freezing the SAE encoder as a reusable sparse safety dictionary and adapting only the decoder to the target-domain activation manifold. During inference, SafeDIG combines Blend and Repel operations to steer unsafe activations toward transferred safety manifolds or away from harmful sparse directions. Experiments on FLUX.1 Dev and Stable Diffusion 3.5 Large show that SafeDIG consistently reduces target-domain and overall unsafe generation rates while preserving source-domain safety and image quality.

Time series foundation models (TSFMs) are transforming the forecasting paradigm through large-scale cross-domain pretraining. However, most existing TSFMs remain univariate, and recent efforts to enable cross-variate modeling still operate directly within the raw variate space. This design introduces fundamental limitations in semantic alignment and relational expressivity. Specifically, raw-space group mixing lacks a dedicated mechanism to align heterogeneous physical quantities, while standard non-negative attention fails to capture the complex synergistic and antagonistic interactions ubiquitous in real-world systems. To address these challenges, we propose Falcon-X, decouples variates from the raw space and maps them into a unified latent prototype space. Falcon-X employs a Unified Prototype Diff-Attention mechanism that explicitly evaluates both positive and negative semantic affinities to explicitly align heterogeneous variates. Cross-variate interactions are then efficiently performed within this shared space via Latent Entity Attention, naturally facilitating zero-shot structural transfer. Finally, a Variate Reassembly Router robustly reconstructs variate-specific trajectories via a request-and-dispatch mechanism. Extensive evaluations on the GIFT-Eval and fev-bench benchmarks demonstrate that Falcon-X achieves excellent forecasting performance, offering a principled and scalable paradigm for complex multivariate environments. Falcon-X is publicly released to support future research.

Flow matching with clean-data prediction has shown that regressing the clean point can exploit low-dimensional structure more effectively than predicting an ambient noised quantity. We ask whether this principle remains useful after images are mapped into a learned latent space, where compression has already removed much of the raw pixel variability. We introduce JLT, a 130M latent diffusion Transformer over frozen FLUX.2 VAE codes, and compare clean-latent prediction with a matched velocity-prediction DiT under the same representation, backbone, and training settings. Although the three variables x, epsilon, and v are linearly convertible for a fixed corruption time, a local Gaussian analysis shows that velocity regression inherits an isotropic target-covariance floor and amplifies low-variance latent directions, while clean prediction damps them. On ImageNet 256 x 256, JLT-B/1 obtains FID-50K 2.50 with classifier-free guidance, with a large matched-target gap over velocity prediction. These results suggest that prediction targets in latent diffusion are representation-dependent geometric choices, rather than interchangeable algebraic parameterizations.

Diffusion Transformers (DiTs) achieve strong visual quality, but their iterative denoising process requires many costly Transformer evaluations. Training-free acceleration methods reduce this cost by caching, forecasting, or verifying intermediate features, yet the runtime decision of when to execute a Full step is often driven by fixed schedules or hand-tuned thresholds. We propose \textbf{SoftCap}, a training-free control layer for cache-based DiT inference. SoftCap couples a Trajectory Drift Observer, which estimates local cache risk from lightweight hidden-state statistics, with a Soft-Budget PI Controller, which adjusts the Full-triggering threshold from realized compute relative to a fixed reference profile. The budget is a soft ceiling: it shapes the threshold but does not require a run to spend a prescribed number of Full evaluations. On FLUX.1-dev, SoftCap improves over SpeCa at a comparable middle-compute operating point, raising ImageReward from 0.967 to 0.981 and reducing LPIPS-Full from 0.518 to 0.498 at nearly identical FLOPs, while target-sweep diagnostics show the intended soft-ceiling behavior as the budget is relaxed.

Diffusion Transformers (DiT) achieve strong performance in image generation but incur substantial inference costs. While prior work has reduced this cost via quantization and distillation, semi-structured sparsity, which can nearly halve FLOPs, remains underexplored. A key reason is that most existing approaches focus on weight sparsification, and pruning 50% of the weights can remove critical model capacity and degrade generation quality. Our study, however, shows that DiT activations are intrinsically sparse and significantly more robust to N:M semi-structured sparsification than weights. Motivated by this observation, we advocate a paradigm shift from weight sparsification to activation sparsification. We propose RT-Lynx, which applies N:M sparsification to activations and incorporates error-compensation techniques to mitigate accuracy loss. We further implement highly optimized CUDA kernels tailored to this setting, achieving up to a 1.55x speedup on average in linear layers. Extensive experiments across multiple diffusion models demonstrate that our method preserves the generation quality of the original models while substantially accelerating inference.

Multi-Modal Diffusion Transformers (MM-DiTs) encode rich representations for training-free concept grounding, but existing attention-based methods often produce overlapping activations on visually confusable concepts, a failure mode we call concept leakage, where target responses spill over to non-target objects. To address this issue, we propose AnchorDiff, a training-free grounding method that decouples semantic localization from structural refinement. AnchorDiff selects a high-confidence anchor from concept-to-image attention map and propagates it as a one-hot seed over a hybrid graph derived from image-to-image self-attention. The graph uses output-space similarity for dense within-object propagation and a row-wise attention gate to suppress cross-object connections. Additionally, we introduce the Multi-Concept Confusion Dataset, which contains images with multiple visually similar concepts and separate masks, enabling explicit evaluation of concept leakage. Experiments show that AnchorDiff achieves strong grounding performance on ImageNet-Segmentation and PascalVOC, while substantially reducing concept leakage on our Multi-Concept Confusion Dataset.

Modern GANs often introduce adversarial supervision on intermediate generator outputs and interpret the resulting multi-stage synthesis as coarse-to-fine hierarchical generation. In this work, we challenge this interpretation. We argue that standard scale-wise adversarial supervision does not construct a proper coarse-to-fine hierarchy: each intermediate image is independently pushed toward the real distribution at its own resolution, but this scale-wise realism does not ensure that outputs across stages represent the identical generated sample. Moreover, the scale-specific image produced at each stage is not used as an explicit refinement target for the subsequent stage. Therefore, its adversarial loss can improve a scale-specific output without constraining later stages to preserve the same sample trajectory, allowing them to move toward a different sample rather than refine the previous output. We refer to this problem as a cross-scale trajectory misalignment problem. To resolve it, we propose CAT, a Cross-scale Aligned Transformer for multi-scale adversarial generation. CAT keeps the discriminator scale-wise, so each intermediate output is evaluated at its own resolution, while adding a simple generator-side consistency regularization that aligns intermediate outputs with the final output. On class-conditional ImageNet-256, CAT-H/2 achieves an FID-50K of 1.56 with one-step inference after only 60 training epochs, outperforming strong one-step GAN and diffusion/flow baselines.

Causal Transformer language models suffer from strictly sequential decoding and a quadratic per-step attention cost. While linear-time causal models and discrete diffusion models each address these weaknesses, their integration remains inherently inconsistent: diffusion requires bidirectional attention, while causal models are unidirectional. To unify these architectures, we propose $B^3D-RWKV$, a diffusion RWKV variant that integrates the model's $O(L)$ inference efficiency with parallel, bidirectional discrete-diffusion through a \emph{triplet-block layout} method. $B^3D-RWKV-7.2B$ reaches comparable accuracy on an 8-task suite versus existing models while significantly outperforming baselines in decoding throughput with an average of $\mathbf{1.6\times}$ speedup.

We introduce ERNIE-Image, an open-source text-to-image generation model built upon an 8B single-stream DiT architecture. ERNIE-Image aims to bridge the gap between current open-source models and leading closed-source systems through more effective mining of large-scale pre-training data and improved supervision quality throughout training. During pre-training, we adopt a bottom-up data construction pipeline that combines fine-grained image categorization, rich caption annotation, aesthetic assessment, and hierarchical sampling. This strategy reduces data noise while preserving long-tail concepts and detailed real-world knowledge, providing a stronger foundation for complex generation tasks. In the post-training stage, we use a top-down data construction pipeline for high-demand scenarios, diversify prompt annotations to better match real user inputs, and apply a stabilized DPO strategy to align the model with human aesthetic preferences. We further train ERNIE-Image-Turbo for efficient 8-NFE generation and propose MT-DMD to mitigate capability drift during distillation. To make the model easier to use in practical scenarios, we equip it with a lightweight Prompt Enhancer that expands concise user intents into structured visual descriptions. In addition, we develop ERNIE-Image-Aes, an industrial-grade aesthetic model, together with ERNIE-Image-Aes-1K, a human-annotated benchmark for realistic aesthetic evaluation. Extensive qualitative and quantitative experiments show that ERNIE-Image achieves leading performance among open-source models and approaches top-tier commercial models in instruction following, text rendering, and aesthetic quality. We release the trained models and aesthetic resources to facilitate further academic research and technical progress in the AIGC community.

Diffusion models have shown promise in learning to solve constraint optimization problems. However, they are mostly restricted to problems with binary variables and rely on graph neural networks, hindering their application to a broader range of problems such as those with general discrete variables or constraint structures that necessitate global rather than local reasoning. We investigate the use of Diffusion Transformers to address the aforementioned limitations. A naive implementation performs poorly due to a fundamental mismatch between the standard diffusion process and constraint solving: while the former applies small, incremental denoising across all variables, the latter requires substantially altering specific subsets of variables to attain feasibility or optimality. Our method, Blocked Gibbs Diffusion Transformer (BloGDiT), is the first to address this limitation by replacing standard joint Gaussian denoising with blocked Gaussian denoising. BloGDiT uses iterative block resampling and anneals the block size over time to facilitate large, targeted edits within a block of variables. Across Sudoku, Graph Coloring, Maximum Independent Set, and MaxCut, BloGDiT matches or outperforms existing methods, demonstrating that blocked Gibbs-style diffusion provides a highly effective inductive bias for Transformer-based constraint satisfaction and optimization.

We present CoMoGen, a controllable video generation framework that generates realistic interactive dynamics from a single binary mask sequence conditioned on an input image. CoMoGen introduces a lightweight MaskAdapter that encodes binary mask sequences into a latent residual signal, injected into the Multi Modal Diffusion Transformer (MMDiT) model through a cosine-weighted schedule. Unlike the hierarchical coarse-to-fine design of UNet architectures, MMDiT operates as a sequence of uniform transformer blocks, making it difficult to identify which layers are responsible for the motion generation. Therefore, we propose a novel way to determine "Motion Layers" operating in the attention space of MMDiT. We fine-tune the model by using Low-Rank Adaptation (LoRA) to the Motion Layers, without requiring any architecture change in the MMDiT. This selective adaptation enables our method to focus on motion-critical components, yielding reduced computational cost. Despite its simplicity, CoMoGen enables precise subject motion and plausible interactions with surrounding humans, objects, and scenes. Comprehensive experiments on different datasets show that CoMoGen consistently outperforms prior controllable video generation methods and achieves state-of-the-art performance in motion fidelity and perceptual realism. Project page: mericadil.github.io/CoMoGen.

Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce SEGA, a training-free method that dynamically scales attention across RoPE components according to the latent's spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.

Controlling physics-based humanoids from natural-language instructions is a critical step toward general-purpose embodied agents. However, existing methods remain constrained by a tension between semantic expressiveness and physical feasibility, often failing to jointly achieve faithful instruction following, high-quality motion, and stable long-horizon control. We propose SCRIPT, a scalable diffusion policy with a multi-stage training framework for language-driven physics-based humanoid control. The core of SCRIPT is a Joint Action-State-Text Diffusion Transformer (JAST-DiT), which represents actions, physical states, and text as dedicated token streams and couples them through joint attention, enabling direct interaction between language semantics and control dynamics. To stabilize autoregressive control, we introduce a nonlinear history conditioning mechanism, which preserves the dense recent context and samples increasingly sparse cues from long-term history. Beyond supervised imitation pre-training, we propose a post-training stage, further improving the performance using Reinforcement Learning with Hybrid Rewards (RLHR). By injecting learnable noise into the flow-sampling process, RLHR effectively improves motion quality and instruction following within closed-loop simulations using hybrid physical feedback and text rewards. Quantitative evaluations demonstrate that SCRIPT outperforms prior state-of-the-art methods, with gains across text alignment, motion quality, and physical realism metrics. Furthermore, scaling studies on the 1200-hour MotionMillion dataset demonstrate consistent performance gains with model scaling, highlighting SCRIPT's robust scalability for large-scale pre-training. Our code will be publicly available for future research.

Diffusion Transformers (DiTs) achieve superior image generation quality but suffer from quadratic computational complexity relative to token count. While various token reduction (TR) methods have been proposed to mitigate this cost, they overlook the primary objective of generative models: minimizing recovery error, which requires reflecting output token similarity. They rely solely on input token similarity inherited from reduction-only ViT paradigms, leading to a fundamental misalignment with this objective. To bridge this gap, we propose DiTo, a novel TR paradigm that shifts the focus toward output-centric token reduction. Based on the observation that output token similarity is consistently preserved across adjacent timesteps, DiTo utilizes prior-step similarities as an effective proxy to establish token correspondences at a Matching timestep, which are then reused across multiple subsequent Reduction timesteps. To optimize this interleaved scheduling, we propose Pair Match Ratio (PMR)-guided Interval Scheduling to determine the optimal matching frequency. Furthermore, to mitigate localized approximation errors and resulting blocking artifacts caused by repeated reuse, we propose Frequency-aware Token Matching by incorporating a selection-frequency penalty. Extensive experiments demonstrate that DiTo consistently outperforms existing TR methods with 1.6-3.9 dB higher PSNR at comparable speedups, achieving a superior Pareto frontier.

Flow matching with $x$-prediction -- regressing the clean data point rather than the ambient velocity -- is known to exploit low-dimensional manifold structure effectively in pixel space \cite{li2025back}. We ask whether a pretrained representation space, while containing a low-dimensional data manifold of comparable intrinsic dimensionality, offers a distribution more favorable for flow-matching learning. Comparing pixel, SD-VAE, and DINOv2 features along four geometric axes, we find that pixel and DINOv2 share nearly identical intrinsic dimensionalities (both $\hat{d}\!\approx\!33$) yet DINOv2 exhibits $7.3\times$ higher effective rank, $35\times$ better covariance conditioning, $11.5\times$ lower excess kurtosis, and $1.7\times$ lower on-manifold interpolation error; SD-VAE latents are consistently intermediate, indicating that the advantage stems from representation-learning objectives rather than mere compression. These statistical properties render the flow-matching regression well-conditioned and remove the need for the specialized prediction heads or Riemannian transport used by prior DINOv2 diffusion methods. We propose the \emph{Representation Image Transformer} (RiT): a vanilla Diffusion Transformer trained by $x$-prediction on frozen DINOv2 features, augmented only by a dimension-aware noise schedule and joint \texttt{[CLS]}-patch modeling. On ImageNet $256{\times}256$, RiT attains FID 1.45 without guidance and 1.14 with classifier-free guidance, outperforming DiT$^\text{DH}$-XL with $19\%$ fewer parameters (676M vs.\ 839M). The resulting ODE is efficiently solvable at coarse discretizations: with classifier-free guidance, $5$ Heun steps already reach FID 2.0 and $10$ steps reach 1.25, without distillation or consistency training. Code at https://github.com/lezhang7/RiT.

Diffusion Transformers (DiTs) have become a de facto backbone of modern visual generation, and nearly every major axis of their design -- tokenization, attention, conditioning, objectives, and latent autoencoders -- has been extensively revisited. The residual stream that governs how information accumulates across layers, however, has been directly inherited from the original Transformer. In this paper, we present a systematic empirical analysis of cross-layer information flow in DiTs, jointly along depth and denoising timestep, and identify three concrete symptoms of traditional residual addition, namely monotonic forward magnitude inflation, sharp backward gradient decay, and pronounced block-wise redundancy. Motivated by this diagnosis, we propose Diffusion-Adaptive Routing (\textsc{DAR}), a drop-in residual replacement that performs \emph{learnable, timestep-adaptive, and non-incremental} aggregation over the history of sublayer outputs. Moreover, the proposed \textsc{DAR} is compatible with many modern Transformer enhancement methods, such as REPA. On ImageNet $256\times256$, \textsc{DAR} improves SiT-XL/2 by $2.11$ FID ($7.56$ vs.\ $9.67$) and matches the baseline's converged quality with $8.75\times$ fewer training iterations. Stacked on top of REPA, it yields a $2\times$ training acceleration in the early stage, suggesting cross-layer information routing as an underexplored design axis in diffusion modeling, one that operates orthogonally to existing representation-alignment objectives. Beyond pretraining, \textsc{DAR} can also be applied during the fine-tuning stage of large-scale T2I models and preserves high-frequency details during Distribution Matching Distillation.

Text-to-image diffusion models often face a severe trilemma in human portrait generation: text-image alignment, photorealism, and human-perceived aesthetics inherently inhibit one another. Supervised Fine-Tuning (SFT) is an effective method for enhancing the photorealism of image generation. However, it often leads to overfitting to the training dataset, corrupts pre-trained image priors, and degrades alignment or aesthetics. To break this bottleneck, we propose a feature supervision paradigm for Multimodal Diffusion Transformers (MM-DiT). Specifically, we introduce a lightweight cross-modal alignment mechanism that implicitly extracts multi-granularity vision-aligned text representations from SigLIP 2 and applies supervision to the image branch of MM-DiT during the training stage, with zero extra inference overhead. Our method injects vision-aligned text guidance while preserving the base model's original generalization, avoiding degradation caused by SFT. Furthermore, our method directly mines implicit multi-granularity aesthetic signals from pre-trained vision foundation models to optimize human-perceived aesthetics. Extensive experiments on MM-DiTs show that our method pushes the Pareto frontier and achieves synergistic improvements across text-image alignment, photorealism, and human-perceived aesthetics.

Creating and editing high-quality 3D content remains a central challenge in computer graphics. We address this challenge by introducing CompoSE, a novel method for Compositional Synthesis and Editing of 3D shapes via part-aware control. Our method takes as input a set of coarse geometric primitives (e.g., bounding boxes) that represent distinct object parts arranged in a particular spatial configuration, and synthesizes as output part-separated 3D objects that support localized granular (i.e., compositional) editing of individual parts. The key insight that enables our method is our use of a diffusion transformer architecture that alternates between processing each part locally and aggregating contextual information across parts globally, and features a novel conditioning technique that ensures strong adherence to the user's input. Importantly, our method learns to infer part semantics and symmetries directly from the user's coarse layout guidance, and does not require part-level text prompts. We demonstrate that our method enables powerful part-level editing capabilities, including context-aware substitution, addition, deletion, and style-preserving resizing operations. We show through extensive experiments that our method significantly outperforms existing approaches on guided synthesis, as measured by objective metrics and LLM-based evaluations.

Recent video editing models have converged on a unified conditioning design: a single diffusion transformer jointly consumes text, source video, and reference images, and one set of weights covers replacement, removal, style transfer, and reference-driven insertion. The design is flexible, but it assumes that the user already provides model-ready text, reference images, and spatial grounding for local edits, which real requests often omit. We present Aurora, an agentic video editing framework that pairs a tool-augmented vision-language model (VLM) agent with a unified video diffusion transformer. The VLM agent maps a raw user request to a structured edit plan aligned with the transformer's conditioning channels, thereby resolving textual and visual underspecification before generation. We train the VLM agent with supervised data for complete edit planning and reference-image selection, together with preference pairs for robust tool use and instruction refinement. We introduce AgentEdit-Bench to evaluate agent-enhanced video editing under textual and visual underspecification. Experiments on AgentEdit-Bench and two existing video editing benchmarks show that Aurora improves over instruction-only baselines and that the VLM agent transfers to compatible frozen video editing models. Project page: https://yeates.github.io/Aurora-Page

Diffusion-based methods demonstrate significant potential for remote sensing image super-resolution at large scaling factors, particularly in reference-based super-resolution (RefSR), where high-resolution reference images provide critical fine-grained texture priors. However, existing methods often suffer from a trade-off between over-reliance on reference information, which leads to texture artifacts, and under-utilization of such information, which results in insufficient detail recovery. To address these issues, we propose DS-DiT, a Decoupled Siamese Diffusion Transformer that decouples the interaction between low-resolution (LR) and reference (Ref) conditions within the attention mechanism. By allowing LR structural priors and Ref texture information to independently interact with the noisy latent, the framework effectively mitigates competition between the two conditional sources. To further compensate for the limited local modeling ability of global attention, we introduce a Patch-Level Weighting (PLW) module that adaptively modulates the fusion of conditional sources. In addition, the siamese architecture enables an inference-time autoguidance strategy that exploits the prediction discrepancy between strong and weak Ref conditions to improve generation quality without additional training. Experimental results across multiple datasets and scaling factors show that DS-DiT outperforms existing methods in both quantitative metrics and visual fidelity.

Recent advances in Diffusion Transformers (DiTs) demonstrate that aligning noisy latent states with well-trained semantic features-as pioneered by Representation Alignment (REPA)-can substantially accelerate training and improve generation fidelity. Subsequent analysis(e.g., iREPA) suggests that these gains arise primarily from transferring spatial structure contained in pre-trained vision representations. However, mostly existing alignment methods employ point-wise matching objectives or rely on implicit architectural tweaks, which fail to explicitly model the spatial relational geometry inherent in vision foundation models. We argue that such element-wise supervision is insufficient to capture the rich spatial topology of visual representations, and that effective alignment for generation should instead be formulated as an explicit structural constraint. To this end, we propose sREPA, a structural REPresentation Alignment framework to enforce consistency in the relational geometry of feature maps, rather than merely matching individual feature points. By encouraging the model to internalize holistic spatial layouts and structural correlations from pre-trained features, sREPA achieves faster and more stable convergence, along with improved sample quality, compared to state-of-the-art alignment strategies. Our code and models will be released.

Diffusion Transformers (DiTs) achieve state-of-the-art image generation quality but incur substantial memory and computational costs at inference. While aggressive Post-Training Quantization (PTQ) to 4-bit precision offers significant efficiency gains, it typically results in severe quality degradation. Existing approaches, including smoothing-based methods, mixed-precision schemes, rotation techniques, and low-rank residual methods, partially mitigate this issue but still leave a noticeable gap to FP16/BF16 performance. In this work, we introduce DiRotQ, a W4A4 PTQ framework that mitigates this degradation through rotation-aware activation quantization. DiRotQ identifies a low-rank subspace capturing dominant activation variance via Principal Component Analysis (PCA), preserving coefficients in this subspace at higher precision while quantizing the remaining components to 4-bit. Activations are rotated into the PCA basis at inference time using calibration-derived orthogonal transformations, while the inverse rotation is fused into the layer weights offline. Combined with GPTQ-based weight quantization, DiRotQ achieves an FID (lower is better) of 15.9 and PSNR (higher is better) of 19.1 dB on PixArt-Σ over the MJHQ-30K dataset, outperforming the prior state-of-the-art SVDQuant (FID 18.9, PSNR 17.6) under the same INT W4A4 setting. Beyond standard metrics, we introduce a VLM-as-a-Judge evaluation protocol for diffusion model quantization, the first such evaluation in this setting, providing a more holistic assessment of perceptual quality and prompt alignment under aggressive compression. On the systems side, we implement a Triton-based custom kernel to enable efficient end-to-end inference, reducing memory usage of the 12B FLUX.1-dev model by 2.1x and delivering 2.3x speedup over the BF16 baseline, on a 24 GB RTX 4090 GPU.

Vision Transformers (ViTs) are known to exhibit high-norm patch-token outliers that degrade feature map quality, a problem effectively mitigated by register tokens. As diffusion models increasingly adopt transformer architectures and move toward pixel-space training, they become closer in form to ViTs, raising the question of whether register tokens are also useful for Diffusion Transformers (DiTs). In this work, we show that DiTs differ from ViTs in a key respect: they do not exhibit patch-token outliers but still benefit from registers. Interestingly, registers are more effective in pixel-space DiTs than in latent-space DiTs. By analyzing intermediate representations, we find that register tokens produce cleaner feature maps at high noise levels, which may contribute to their effectiveness in pixel-space generation. We further observe that recent pixel-space DiT architectures implicitly incorporate register-like mechanisms, which may partially account for their strong empirical performance. Motivated by these observations, we propose Register Guidance, a technique that amplifies the contribution of register tokens responsible for improving visual structure and coherence.

Pixel-space diffusion models bypass the reconstruction bottleneck of Variational Autoencoders (VAEs) but face a fundamental "granularity dilemma": capturing global semantics favors large patch scales, while generating high-fidelity details demands fine-grained inputs. To address this issue, we propose HyperDiT, a unified framework establishing Hyper-Connected Cross-Scale Interactions to bridge the semantic and pixel manifold. Diverging from injecting semantics by AdaLN, HyperDiT utilizes Cross-Attention mechanisms, enabling fine-grained tokens to query multi-level semantic anchors globally. To resolve the spatial mismatch during multi-scale interactions, we introduce Scale-Aware Rotary Position Embedding (SA-RoPE) to ensure precise geometric alignment among tokens of varying patch sizes. Furthermore, we incorporate Registers to learn the dense semantics from a pretrained Visual Foundation Model (VFM), effectively reducing generation hallucination and artifacts. Extensive experiments demonstrate that HyperDiT achieves state-of-the-art (SoTA) FID of $\mathbf{1.56}$ on ImageNet $256\times256$ directly within the pixel space. By combining the fine-grained stream with semantic guidance, HyperDiT offers a superior paradigm for high-fidelity pixel generation.

The Diffusion Transformer (DiT) architecture is the state-of-the-art paradigm for high-fidelity image generation, underpinning models like Stable Diffusion-3 and FLUX.1. However, deploying these models on resource-constrained mobile devices entails prohibitive computational and memory overhead. While efficiency-driven approaches like Linear-DiT and static pruning alleviate bottlenecks, they often incur quality degradation. Unlike cloud environments, mobile constraints require a single-model paradigm that dynamically balances fidelity and latency. We introduce ElasticDiT, which achieves this dynamic trade-off by adjusting spatial compression ratios and DiT block depths. By integrating Shift Sparse Block Attention (SSBA) and a Tiny DWT-Distilled VAE (T-DVAE), ElasticDiT reduces inference latency and memory footprint while maintaining image quality. Experiments confirm that ElasticDiT effectively covers a wide range of fidelity-latency trade-offs within a single set of parameters. By jointly adjusting compression and depth, a single ElasticDiT model can be reconfigured on-the-fly to outperform task-specific baselines. Specifically, our flex lite variant achieves an HPS of 32.87, surpassing the Flux model, while maintaining competitive quality at 84.16 percent average sparsity through SSBA. Furthermore, the plug-and-play T-DVAE provides SD3-level reconstruction with only 1/8x the computational cost of standard VAEs, and Flow-GRPO boosts semantic alignment (GenEval: 66.93 to 73.62). These results demonstrate that ElasticDiT offers a versatile, hardware-adaptive solution that eliminates the need for multiple specialized models, providing a promising path for future high-resolution image generation on mobile devices.