Diffusion Model Inference Acceleration

Few-step diffusion models substantially compress temporal computation, making the spatial cost of each model evaluation an increasingly dominant source of inference latency. Progressive-resolution inference reduces this cost by performing early denoising at low resolution and reserving high-resolution computation for refinement. However, existing methods typically lift intermediate latents directly and rely on subsequent steps to absorb the induced distribution mismatch. In the few-step regime, the limited recovery budget leaves these errors as visible artifacts, constraining how late the transition can occur and, consequently, how efficiently it can be performed. We introduce SelfLift, a self-recovering progressive-resolution framework that derives both transition-repair signals and trajectory-aligned supervision from the generative model itself. SelfLift-zero proposes a training-free Artifact-Aware Consistency Lift, using disagreement between direct latent lifting and pixel-VAE re-encoding as both a localized artifact-risk signal and a model-native correction direction. It enables reliable late transitions without external super-resolution, extra denoiser evaluations, or sampling-schedule modifications. Building on this robust transition, SelfLift-rich performs On-Policy Self Recovery on student-visited states, transferring dense high-resolution guidance from an internal self-teacher while remaining aligned with the altered progressive-resolution dynamics. Across FLUX.2-Klein and Z-Image-Turbo, SelfLift reduces end-to-end latency by 41.5% and 44.1%, respectively. Combined with timestep distillation, it delivers overall speedups of 29.61x and 19.21x over the corresponding 50-step models while preserving competitive generation quality, establishing a stronger speed-quality frontier for few-step diffusion.

We propose LF-MultiDiffusion, a training-free panorama generation method that extends MultiDiffusion to support linear projections between target and reference image spaces. Our key idea is to reformulate latent aggregation as a regularized least-squares problem and solve it efficiently with a Krylov-based iterative solver inside the denoising loop. This formulation enables denser and more natural mappings than prior training-free methods, yielding more stable generation with far fewer perspective views. As a result, LF-MultiDiffusion reduces the number of image generator evaluations during denoising and significantly improves inference efficiency. Experiments show that LF-MultiDiffusion achieves better visual quality, text alignment, and panoramic consistency than the strongest training-free baseline, while providing a 15.36$\times$ speedup. Our project page is available at: https://ahykw.github.io/lfmd.

Recent advancements in low-light image enhancement have leveraged diffusion models for their strong ability to generate perceptually realistic, detailed images. Patch diffusion models further offer a promising solution to size-agnostic image restoration while improving efficiency. However, existing methods typically rely on small, fixed patches (e.g., 64$\times$64) that cannot capture image-level brightness context, whereas enlarging the receptive field improves brightness and colour estimation but substantially increases computational cost. Moreover, low-light images often exhibit uneven brightness across regions, making it necessary to ensure that locally enhanced patches remain visually coherent when combined into the full image. To address these limitations, we propose P-PatchDiff, a scalable progressive patch diffusion framework for low-light image enhancement that dynamically adjusts patch size throughout the denoising process, enabling a gradual shift from local to global views. A Multi-Patch Alignment strategy is also introduced to normalise features across varying patch scales using an estimated global brightness proxy. Rather than pursuing pixel-level reconstruction accuracy, P-PatchDiff focuses on scalability and coherent brightness across the whole image, allowing the model to perceive multi-scale information and better enhance regions with varying brightness. We empirically demonstrate that P-PatchDiff effectively enhances images ranging from 400 $\times$ 600 to 4K and is 80$\times$ faster than existing patch diffusion models while using less than 9GB of memory. The code is available at https://github.com/RuoyuGuo/P-PatchDiff.

Diffusion models have achieved strong results in high-fidelity image synthesis, but their iterative sampling process makes large-scale generation computationally expensive. This limitation is especially relevant when generating synthetic face datasets for face recognition, where a large number of subjects with many samples in different poses, expressions, ages, etc., are required. In this work, we show that identity-conditioned face synthesis can be performed at a substantially lower computational cost by a latent Consistency Model with few iterations, without compromising image quality. For training, we distill knowledge from the foundation Diffusion Model Arc2Face (teacher) by adapting its original text-to-image pipeline to an embedding-to-face setting, replacing textual prompts with ArcFace identity embeddings. Our distilled model (student) generates identity-conditioned face images with an average inference time of 0.4819 seconds per image, compared with 2.102 seconds for Arc2Face, resulting in a 4.36$\times$ speed-up. Quantitative results, based on FID scores, show that the distilled model remains competitive with Arc2Face across all evaluation protocols. On 100k generated images, it achieves near-parity on CelebA (13.921 vs. 12.928) and outperforms the teacher on WebFace42M (9.317 vs. 9.802). Further evaluations on Synth-500 and AgeDB show a moderate performance gap for the former but comparable results for the latter. These results indicate that Arc2Face can be accelerated through task-specific latent consistency distillation while preserving high image quality for large-scale synthetic face generation. Our proposal is publicly available at https://github.com/UFPR-IPASP-PR/FaceRec-IdentityConsistency.

Speculative decoding speeds up generation with an efficient draft model (drafter) that proposes tokens for a target model to verify in one pass, preserving the target's output distribution. High-acceptance block-diffusion drafters such as DFlash and DFlare fill an entire block in one parallel pass. In many cycles, the target accepts the whole block, so the drafter exhausts its trained block horizon before verification fails. We call this unrealized acceptance stranded speed-up. A mean committed length, per prompt or per cycle, hides it, whereas the acceptance histogram exposes it as a spike in the ceiling bin, the fraction of cycles that accept the entire block. We recommend the histogram as a preflight check before spending training compute. Naively widening the block at inference does not recover the speed-up, because once the block outgrows its training size, the drafter's bidirectional attention shifts its distribution even at early positions and erodes front-of-block verification. Instead, we post-train the drafter on a longer block with a short curriculum that emphasizes the newly exposed positions, a method we call DBloom. Expanding the pretrained DFlash and DFlare drafters from block size 16 to 24 across Qwen3-8B and Qwen3-4B targets raises the per-prompt committed length on the high-ceiling benchmarks by a median of +0.8 tokens (up to +1.1). Once continuation fine-tuning precedes expansion, the increase reaches 1.37 tokens. The same expansion also lifts committed length on all seven benchmarks for Gemma-4-12B-IT, a different model family, by a median of +0.41 tokens (Arm A), and the full continuation-then-expand pipeline (Arm B) adds +0.29 to +0.98 tokens over the same B16 drafter. In a prompt-matched comparison against JetSpec, a contemporary tree-based drafter not used in our design, DBloom commits more tokens on every benchmark at tree budgets up to 64 nodes.

This work presents $\textbf{Lapis}$, a $\textbf{l}$inear-$\textbf{a}$ttention-based $\textbf{pi}$xel-$\textbf{s}$pace generative framework that achieves efficient and high-fidelity depth estimation with one-step diffusion. While generative frameworks have significantly advanced monocular depth estimation with superior detail fidelity, the $\mathcal{O}(N^2)$ complexity of standard attention and the multi-step denoising process introduce prohibitive computational costs when scaling them to high-resolution image applications. Although linear attention and one-step prediction are intuitively viable, directly applying them leads to poor structural consistency, detail loss, and noise. Lapis rectifies these limitations through a coarse-to-fine hierarchy. Specifically, a Patch-level Consistency Module restores structural coherence by integrating semantic and spatial priors. Subsequently, a Pixel-level Refinement Module recovers sharp geometric boundaries via skip-connection-based pixel correspondence. Furthermore, to mitigate sampling noise inherent in one-step diffusion, we leverage the manifold assumption and adopt a direct $\mathbf{x}$-prediction strategy to target the clean data manifold. Extensive evaluations on multiple benchmarks demonstrate that Lapis consistently achieves state-of-the-art (SOTA) accuracy and boundary sharpness across various resolutions, reducing inference latency by up to 7.6$\times$ at 1080P and 10.9$\times$ at 1440P resolution compared to previous SOTA generative models.

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.

Unified multimodal models jointly support understanding and generation, but incur substantial redundant computation across tokens, layers, and generation timesteps. Through token-importance probing, we identify an asymmetric core-expansion structure: understanding exhibits a stable importance component, while generation largely shares this component but requires progress-dependent corrections. We therefore propose CE-Router, which uses a task-shared core scorer and progress-conditioned generation expansions, optimized through generation decomposition and cross-task core alignment. At inference, CE-Router compacts token computation and supplies a learned routing signal to Unified Computation Scheduling, which coordinates layer skipping, FFN pruning, diffusion-head cache reuse, and denoising-step early exit. Experiments on two representative UMM architectures demonstrate consistent quality--efficiency improvements across both tasks, retaining 98.03% of dense understanding performance with a 1.93$\times$ end-to-end inference speedup.

Joint text-to-video-audio generation produces synchronized visual and acoustic content, but the long sampling trajectories and heterogeneous multimodal computation of large models make inference prohibitively expensive. We present TurboT2VA, a distillation and inference framework for accelerating a 19B-parameter joint video-audio model. Large-scale T2VA distillation is challenged by modality-imbalanced optimization, the difficulty of continuous-time consistency training at scale, and the quality--diversity trade-off. TurboT2VA addresses these issues with per-modality normalization and a progressive curriculum comprising discrete consistency warm-up, continuous consistency refinement, and joint consistency--distribution matching. The curriculum first establishes a stable, diverse generation trajectory and only then introduces distribution-level refinement. On LTX-2, four-step distillation reduces generator latency from 50.52s to 2.51s at the standard evaluation resolution of 512$\times$768, achieving a 20.1$\times$ speedup while maintaining strong visual quality, audio fidelity, diversity, and video-audio synchronization. We further develop an architecture-aware inference stack that combines guarded W8A8 and fused operators, padded-text compaction, and modality-aware sparse attention while preserving dense cross-modal and text-conditioning paths. Under the high-resolution deployment setting at 1024$\times$1792, the complete stack reduces generator latency from 318.74s to 5.83s on one NVIDIA H20, achieving a 54.67$\times$ generator-only speedup. Inference code and generation demos are available at https://github.com/thu-ml/TurboDiffusion/tree/main/turbot2va.

In-context diffusion transformers concatenate instruction, target, and reference tokens into a single sequence for joint attention. Reference-side computation must therefore be repeated at every denoising step, with the cost growing rapidly as more references are added. Decoupling reference tokens from the target enables exact key-value reuse across denoising steps, but prevents the references from attending to the instruction, degrading instruction following and reference fidelity. This trade-off cannot be resolved through attention-mask design alone. We introduce AnchorCache, a parameter-free token-layout and attention-mask co-design that inserts static text anchors. These anchors condition the reference representations on the instruction during cache construction, after which the resulting reference keys and values can be reused exactly across denoising steps. To recover the quality initially lost through this structural conversion, we apply teacher-forced velocity distillation followed by a short on-policy stage that queries the teacher at student-visited states. To our knowledge, this is the first use of on-policy distillation for architectural recovery in diffusion models. Across benchmarks spanning image, speech, and video generation, AnchorCache matches full-attention quality. Its efficiency gains increase with the reference-context size, reaching a 6.40x speedup in diffusion transformer inference.

Speculative decoding losslessly accelerates autoregressive language models by verifying multiple draft tokens in parallel. Diffusion-based drafters further reduce proposal latency by predicting an entire token block in parallel, but their position-wise distributions are marginal rather than conditioned on tokens selected along each draft path. Existing recurrent correction incorporates causal information along a single draft chain, whereas diffusion-based tree construction broadens candidate coverage without carrying this correction along individual branches. We introduce DARTree, a training-free speculative decoding method that extends a pretrained AR correction head from chains to trees. DARTree first constructs a fixed-width candidate tree by expanding and scoring all nodes at each depth in a single batch, and then only applies best-first pruning to select the verification tree, decoupling AR-head inference from sequential heap operations. Across seven math, code, and chat benchmarks, DARTree achieves the highest average acceptance length and speedup in all four model--temperature configurations, accepting up to 12.97 tokens per verification round, 98.6% more than DFlash and 27.9% more than Domino in the same setting, and reaching up to 9.73$\times$ lossless speedup over locally measured autoregressive decoding.

Geometry-conditioned multi-view diffusion enables high-quality 3D texture generation, but its repeated per-view denoiser evaluations introduce substantial computational cost. Existing training-free accelerators primarily exploit temporal redundancy by reusing computation across denoising steps. In multi-view texturing, however, skipping a step also removes the cross-view interaction that continually aligns different observations of the same surface, leading to rapidly degraded consistency and fidelity. Our analysis identifies a complementary source of redundancy: although intermediate features remain view-specific, geometrically corresponding surface points exhibit transferable evolution in their predicted clean signals. Based on this observation, we introduce \gc{}, a training-free plugin that evaluates a rotating subset of anchor views and transports their geometry-aligned per-step $\xz$ updates to the remaining views. Periodic full-view computation controls accumulated error, while sampler-consistent reconstruction preserves the denoising trajectory. \gc{} requires neither retraining nor architectural modification and uses the position maps already available in geometry-conditioned texturing pipelines. Across Hunyuan3D-2.1, SyncMVD, and MVPainter, \gc{} achieves a stronger speed--fidelity trade-off than temporal caches and step reduction at operating points above $2\times$. On Hunyuan3D-2.1, it delivers a $2.21\times$ denoiser-loop speedup with an MV-LPIPS of 0.0293 and an MV-PSNR of 33.60 dB, providing the best fidelity among all tested methods above $2\times$. The same transferred configuration reaches the highest speedup and lowest FLOPs on SyncMVD, while \gc{} achieves the lowest FLOPs and best fidelity among the accelerated methods on MVPainter. These results establish cross-view geometry as an effective acceleration axis for multi-view texture diffusion.

Diffusion models have achieved dominant performance in visual generation but suffer from substantial inference overhead. While cache-based acceleration has emerged as a promising solution, existing policies rely on local similarity heuristics, which we identify as being significantly misaligned with final generation quality. This discrepancy stems from the non-uniform propagation and accumulation of errors along the denoising trajectory. To address this, we propose Global-Impact Cache (GCache). We first establish a rigorous theoretical characterization of the error propagation upper bound. Recognizing that this bound can be overly conservative for complex, highly non-convex diffusion models, we further reparameterize the propagation exponent with a Bernstein form and reformulate cache policy search as a bilevel optimization problem. In detail, GCache identifies an optimal reuse policy in the inner objective while aligning the error-weighting function with generation quality loss in the outer objective. This framework effectively reconciles theoretical rigor with empirical performance, learning to prioritize computation where it most impacts visual fidelity. Extensive experiments demonstrate that GCache consistently outperforms prior caching strategies on both video and image generation. Notably, on the state-of-the-art Wan2.1 video diffusion model, GCache maintains a 2.17x speedup while significantly enhancing generation quality, reducing LPIPS from 0.1095 to 0.0316.

Vision-Language-Action (VLA) models promise to bring end-to-end reasoning to autonomous driving, but their computational cost remains far too high for real-time control. The core challenge is structural: VLA inference is not a single bottleneck but a cascade of four. Visual encoding wastes compute on overlapping video frames; language-model prefill recomputes context that could be carried over from the previous timestep; reasoning tokens are generated serially despite low entropy; and flow-matching denoising applies uniform compute to a non-uniform velocity field. Addressing any one stage in isolation leaves the others untouched. We propose FlashDrive, an algorithm-system co-design framework that targets all four stages simultaneously. Our key insight is that each bottleneck admits a distinct, lightweight algorithmic shortcut: temporal overlap enables streaming KV-cache reuse across frames; the low per-token entropy and strong intra-block correlations of driving-domain reasoning make a non-autoregressive diffusion drafter highly effective for speculative decoding; and the velocity field's structure---sharp at the endpoints, flat in the middle---permits adaptive step caching that concentrates compute where it matters. Layered on system-level CUDA Graph compilation and kernel fusion, these techniques compound. Applied to Alpamayo 1.5-10B with W4A8 quantization, FlashDrive reduces end-to-end latency from 717ms to 151ms (4.7x) while leaving accuracy essentially unchanged: [email protected] shifts by only 0.08m, minADE1 improves, and closed-loop collision and off-road rates improve in simulation. By raising a 10B-parameter reasoning VLA from 1.4Hz to 6.6Hz on a single GPU, FlashDrive moves end-to-end autonomous driving substantially closer to real-time deployment.

Diffusion Transformers (DiTs) incur quadratic self-attention cost over spatiotemporal tokens. Existing training-free sparse attention methods often construct sparse masks from block-level or cluster-level proxy scores, which can obscure fine-grained differences among keys and miss high contribution keys under aggressive sparsity. Moreover, such proxy scores may yield overly concentrated softmax distributions, causing Top-$p$ to retain too few keys for some query clusters. Although a fixed Top-$k$ minimum alleviates this failure mode, a shared value cannot adapt to variations across heads and inputs. To address both limitations, we propose SCOPE, a training-free sparse attention framework that combines 3D-RoPE-aligned key subspace clustering with online per-head Top-$k$ estimation for efficient video-DiT inference. SCOPE partitions post-RoPE keys into temporal, height, and width subspaces, clusters them independently, and aggregates the corresponding centroid scores through lookup tables to obtain per key proxy scores for each query cluster. Building on existing hybrid Top-$p$/fixed Top-$k$ selection, SCOPE derives a head-specific Top-$k$ value online by averaging the initial retained key counts within each head, weighted by query cluster size, and selects additional keys only for query clusters whose initial retained key counts fall below this value. Sparse attention is then computed over the selected original keys and values. Across six model--task configurations, SCOPE consistently outperforms existing training-free baselines in both fidelity and latency, achieving up to a $1.99\times$ end-to-end speedup on 720p HunyuanVideo with $28.46$ dB PSNR relative to dense attention.

Generative models like Diffusion Models and Flow Matching have demonstrated remarkable capabilities in synthesizing high-fidelity driving videos, but are severely constrained by high inference latency due to the requirement of extensive sampling steps. We argue that this inefficiency stems from the prevailing reliance on a standard Gaussian source distribution, where consecutive frames are initialized as independent Gaussian noise. This paradigm disregards the rich spatiotemporal correlations inherent in driving videos, compelling the model to regenerate deterministic scene structures existing in previous frames from noise, which is both computationally redundant and prone to geometric inconsistency. To address this problem, we propose GeoFlow, a novel framework designed to achieve efficient driving video generation by harnessing explicit geometric priors. Instead of sampling from standard Gaussian noise, we leverage multi-view geometry and spatially-adaptive noise injection to construct a Geometry-Aligned Prior (GAP) distribution as starting point. This initialization bridges the gap between source distribution and data distribution, yielding a significantly straighter and shorter sampling trajectory. Extensive experiments demonstrate that GeoFlow can achieve remarkable efficiency of both training and inference: merely several hours of fine-tuning on baseline models can significantly boost few-step generation quality, while fully converged training drastically reduces number of inference steps required for state-of-the-art video generation.

Video diffusion transformers are costly to sample: every denoising step applies self-attention over a long 3D token sequence, a quadratic cost that dominates as resolution and duration grow. Sparse attention reduces this cost without retraining, but existing methods pursue aggressive sparsity, where further speedup costs disproportionately more attention fidelity. We target the opposite end of this trade-off: fix near-lossless fidelity by construction, and remove as much computation as this constraint permits. Two observations make this regime practical: roughly 40% of block interactions can be removed while retaining 99% of the attention mass, and the high-mass support remains stable across denoising steps. We propose LoSA, a training-free sparse-attention method that fixes a retained-mass threshold of 99% rather than a sparsity ratio: it measures exact block attention masses at one early dense step, keeps, for each head and query block, the smallest key/value block set meeting the threshold, and reuses the frozen block indices for all remaining steps. On Wan2.1-1.3B, LoSA alone gives a $1.36\times$ speedup with a 0.06-point VBench Overall drop. The benefit is largest under composition: combined with feature caching, LoSA reaches a $3.2\times$ speedup on HunyuanVideo at a 0.02-point drop, versus 0.32 points for the strongest sparse baseline at comparable speed. Across three video diffusion transformers and speedups up to $3.2\times$, LoSA consistently achieves the best training-free speed-quality trade-off.

Diffusion-based super-resolution (SR) achieves strong perceptual quality but requires costly iterative denoising. Existing one-step distillation methods reduce inference time but depend on expensive pretrained teachers, whereas CTMSR avoids distillation through PF-ODE consistency training yet does not explicitly model the restoration dynamics from low-resolution (LR) inputs to high-resolution (HR) images. We propose MeanSR, a one-step perceptual SR method that learns an LR-conditioned average velocity field to directly capture the finite-time transition from degraded or noisy inputs to plausible HR outputs. We further reformulate distribution trajectory matching for average-velocity generation and introduce a Stage-Aware Temporal Sampling strategy to improve trajectory learning. Experiments on synthetic and real-world benchmarks show that MeanSR outperforms CTMSR on CLIPIQA, MUSIQ, and MANIQA while substantially reducing FLOPs and inference latency. MeanSR also reconstructs sharper structures and more realistic textures with fewer perceptual artifacts.

Diffusion language models (DLMs) iteratively refine a sequence, allowing earlier predictions to be revised as context evolves. This rollback capability distinguishes them from irreversible autoregressive generation, but makes inference costly. Every denoising update alters the global context, forcing both prompt and response states to be recomputed even though only response tokens are revisable. Key-value (KV) caching could reduce this cost, yet conventional caching assumes immutable historical states and is therefore difficult to reconcile with rollback. In this paper, we introduce Adaptive Reuse of Cached Hidden States for Efficient Rollback (Archer), a training-free KV caching method for rollback-capable DLMs. Archer asymmetrically keeps the mutable response synchronized with the current hypothesis while reusing prompt K/V within a bounded state neighborhood. Although prompt representations also change under bidirectional attention, their token identities remain fixed; bounded reuse therefore amortizes repeated prompt computation without caching mutable response states. It also delays feedback from tentative tokens, reducing premature reinforcement of transient high-confidence errors and giving rollback more opportunity to correct them. Our analysis characterizes prompt reuse as a reversibility-aligned cache boundary, bounds its state-dependent approximation error, and gives a decoder-margin condition for preserving full-refresh decisions. Existing DLM acceleration often trades quality for speed. Archer shifts this frontier, attaining the best mean performance of 33.63% together with a 2.57x mean speedup on the main suite. Across evaluated settings, it improves Pass@1 by up to 3.05 points and reaches up to 2.95x speedup. Controlled analyses connect the quality gain to delayed prompt feedback and validate state-aware refresh. Our code is available at https://github.com/Hxnng/Archer.

Diffusion models have demonstrated remarkable effectiveness in image restoration tasks. However, when guiding image reconstruction, existing Diffusion Model-based Image Restoration (DMIR) methods typically rely on fixed data constraints and uniform step sizes, thereby overlooking the dynamic nature of the generative process. Such rigid designs render the models vulnerable to spatially non-uniform degradations, thus resulting in structural distortions and loss of fine details. Meanwhile, uniform step sizes introduce computational redundancy, whereas naïve step reduction strategies tend to accumulate approximation errors. To address these limitations, we propose a Local Epistemic Uncertainty Guided Active Sampling framework (LEADer). In the spatial domain, LEADer leverages pixel-wise uncertainty to dynamically modulate the prior strength within the null space, which effectively balances detail preservation and artifact suppression. In the temporal domain, it quantifies sampling stability via the uncertainty trace to enable adaptive trajectory pruning, thereby accelerating convergence. Theoretical proofs demonstrate that our framework achieves strict data consistency, while the trajectory pruning strategy admits a deterministic error bound, thereby guaranteeing stable convergence under skip sampling. Notably, our plug-and-play method can be seamlessly integrated into various DMIR baselines. Extensive experiments show that LEADer improves the performance of multiple state-of-the-art DMIR methods, while significantly reducing sampling time with negligible memory overhead. Code is available at https://github.com/JiaqiZhang-Sengoku/LEADer.

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.

Large video diffusion models provide rich spatiotemporal priors for autonomous driving, but existing world-action models often inherit the cost of iterative future-video generation even though deployment only requires an ego trajectory. We ask a more basic question: how much of a video diffusion model must be executed to make a reliable driving decision? Through a controlled study of video denoising timesteps and Diffusion Transformer (DiT) depth, we find that planning performance is largely insensitive to the tested video-noise levels, whereas strong trajectories can already be decoded from intermediate layers. Based on this observation, we introduce Adaptive-WAM, a quality-aware multi-exit planner built on a Wan2.2-5B backbone. Trajectory diffusion heads are attached to selected DiT blocks, and a lightweight trajectory-quality scorer terminates inference once the best trajectory decoded so far satisfies a quality threshold; otherwise, computation continues from the cached hidden state to a deeper exit. The deployed planner therefore avoids the iterative classifier-free denoising loop and VAE decoding required for future-video synthesis, while dynamically allocating backbone depth according to trajectory quality. On NAVSIM, the adaptive single-trajectory planner achieves 90.8 PDMS; a separate fixed-exit variant reaches 92.6 PDMS with 64 proposals. It further obtains 89.9 EPDMS on NAVSIM v2, yielding the best reported results among the compared front-view video world-model planners. Without target-domain fine-tuning, Adaptive-WAM transfers to nuScenes with 0.88 m average L2 error and a 0.08% collision rate. On an A100, adaptive routing improves PDMS from 90.62 to 90.79 while averaging 170 ms end-to-end planning latency, approximately 10% below the 190 ms fixed block-15 planner and 47% below the 320 ms fixed full-depth planner. Code will be released.

Pixel-space diffusion models avoid the reconstruction ceiling of latent diffusion models by generating directly in image space. However, their substantially higher token count makes generation expensive due to the quadratic complexity of self-attention. Several existing efficiency methods reduce this cost by using larger patches at selected denoising steps, thereby representing the image with fewer tokens. Yet, each step still uses a single patch size uniformly across the entire image, overlooking that different regions suffer different fidelity losses when coarsened. We introduce MOSAIK, a damage-guided framework that varies patch size across regions and denoising steps. MOSAIK adapts the PixelDiT backbone to generate arbitrary heterogeneous patch layouts, and a lightweight predictor uses intermediate denoising features to estimate the fidelity loss caused by coarsening each region. Given a token budget, our damage-guided layout predictor assigns fine patches to sensitive regions and coarse patches elsewhere. Remarkably, while reducing FLOPs by 70% and token count by 83%, MOSAIK matches the full-compute PixelDiT on GenEval and its DPG-Bench score drops by only 1.0 point. Compared to diverse efficiency paradigms, including temporal patch scheduling and feature caching, our approach delivers highly competitive performance at moderate budgets and consistently outperforms these baselines in highly constrained compute regimes.

Current few-step autoregressive video diffusion models depend on previous fully denoised clean frames as context for all denoising steps of the current frame. However, these clean frames leak excessive local details, which causes the model to take shortcuts, resulting in compromised temporal semantics and dynamics. Inspired by the perspective of diffusion as masking, we explore the impact of noisy contexts on few-step autoregressive generation. Yet, simply applying contexts with the same noise levels provides insufficient guidance, leading to poor temporal consistency. To resolve this dilemma, we introduce In-Context Forcing, a progressive autoregressive paradigm that utilizes contexts with decreasing noise levels. By applying less masking to distant frames and more masking to adjacent ones, this approach provides adaptive guidance, effectively ensuring both robust temporal consistency and high inter-frame dynamics. Furthermore, by decoupling the strict dependence on previous clean frames, our paradigm enables cross-frame parallel denoising, achieving substantial inference acceleration without sacrificing performance. Extensive experiments on VBench demonstrate that our method significantly outperforms state-of-the-art approaches in both visual fidelity and inference speed.

Vision-language-action (VLA) models have emerged as a key component in embodied AI. Among existing approaches, diffusion-based VLA models achieve superior motion quality and generalization. However, diffusion-based VLA models are compute-intensive and must run at high control frequency, e.g., 50-200 Hz. Thus, it imposes strict latency and energy constraints on edge devices. In this work, we present Deltoris, an algorithm-hardware co-design framework for efficient diffusion-based VLA inference. First, we exploit the temporal similarity of consecutive inputs and propose a \textit{temporal-aware bit-sparsity} algorithm that computes only the differences between consecutive inputs, eliminating redundant bit-level operations. To further address the extra off-chip traffic introduced by our algorithm, we propose a \textit{speculative inference} technique, which amortizes data loading across multiple control steps. Lastly, to support these techniques, we co-design a dedicated accelerator with customized 1D systolic bit-serial PE arrays that eliminate PE workload imbalance. Our evaluation shows that Deltoris achieves up to 34.2$\times$ speedup over mobile GPUs and 6.1$\times$ over prior accelerators, while maintaining comparable accuracy.

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.

Video diffusion transformers (vDiTs) generate high quality but pay quadratic self-attention cost, making inference prohibitive at video-token scales. The challenge is input-adaptive sparsity: selecting critical Q/K/V tokens with negligible overhead and executing them for end-to-end gains. We present SPADE, a training-free sparse-attention engine of three parts: (i) vDiT-SSR, a specification defining 3D blocking candidates and formalizing dynamic masks via Summarizer/Estimator expressions; (ii) runtime scheme generation using SICS and a head-wise policy; and (iii) an executor with low-overhead index search, flash block-sparse attention, and kernel grouping. Across Hunyuan-Video and Wan 2.1/2.2 for text-to-video and image-to-video generation, SPADE raises sparsity and speed while preserving quality, accelerating attention by 2.26x-3.40x and end-to-end inference by 1.49x-1.80x. Our code is open-sourced at https://github.com/6somehow/DAC-SPADE.

Video Diffusion Transformers (VDiTs) enable high-fidelity generation but incur quadratic cost from dense 3D self-attention. Existing head- and block-level sparse methods share computation budgets across queries, overlooking token-specific attention demand. We observe that retained density varies across queries yet correlates log-linearly with attention entropy, while dominant interactions form query-centered neighborhoods with token-dependent radii. Based on these findings, we propose Token Radius Attention (TRA), a training-free framework that maps query entropy to an analytic token budget and converts it into a temporally decayed radius without explicit key ranking. Fused entropy extraction, warm-up reuse, and block-sparse mask construction further reduce overhead. Across seven Wan2.1, Wan2.2, and HunyuanVideo T2V/I2V configurations, TRA retains only 9-19% of attention interactions and achieves 1.56x-2.05x speedup with competitive generation quality. Code is available at https://github.com/IF-LAB-PKU/Token-Radius-Attention.

Block-diffusion drafters like dFlash generate an entire block of draft tokens in a single forward pass, drastically reducing the overhead of multiple-token drafting in speculative decoding. The crucial final step of the single-pass discrete denoising process involves using the logit distribution at each position to sample conditionally independent tokens. The resulting draft is thus a set of per-position marginals, rather than a joint distribution: no draft token is guaranteed to depend on its predecessors. Such independently sampled marginals tend to produce sequences with tokens that are individually likely, but jointly improbable under the target model's distribution, which verifies each token conditionally. This can cause early rejection and limits acceptance length. To address this, we propose xPress as a means to restore the missing causality in diffusion drafters. xPress is a lightweight causal refiner that reconciles the whole diffusion block at once through parallel refinement, restoring and propagating causal dependencies across the draft without a token-by-token loop. On Qwen3-8B, across seven math, code, and chat benchmarks, xPress raises acceptance length by about 30% on average (up to +56%) and its end-to-end decoding throughput by about 1.3 on average (up to 1.7) compared to the original dFlash diffusion drafter.

Diffusion world models generate high-quality futures, but re- peated transformer evaluations make inference prohibitively slow. Existing caches reuse intermediate features, selectively update tokens, or reuse and extrapolate denoising outputs ac- cording to local drift or short native-space histories. These criteria can miss both approximation-induced latent transition defects that accumulate across skipped steps and phase- or condition-dependent changes in the direction of latent evo- lution. We propose WorldDynCache, a risk-controlled latent dynamics approximation framework with two core compo- nents. First, a lightweight latent-transition risk estimator tracks the accumulated future impact of approximation defects and calibrates its predictions against counterfactual defects ob- served at exact anchors. Second, a condition- and phase- aware lifted latent surrogate approximates latent evolution without extra transformer evaluations. On HunyuanVoyager- 13B and Aether-5B, WorldDynCache achieves 4.92 times and 2.15 times speedups, respectively, while attaining the best gen- eration quality among the compared caching methods across WorldScore, PSNR, SSIM, and LPIPS.