Overview

Diffusion models learn a data distribution by gradually corrupting examples with noise and training a denoiser to reverse that process. Major approaches include DDPMs and score-based SDEs, latent diffusion and diffusion Transformers, flow and consistency methods, and discrete masked or continuous-time Markov diffusion; conditioning and guidance adapt them to specific tasks. They are used for images, video, audio, language, molecules, and scientific forecasting, with evaluations ranging from ImageNet FID to language perplexity and domain-specific measures of accuracy and uncertainty. A central tension is how to make sampling faster and generation more controllable without sacrificing quality, calibration, or diversity, while theory and ongoing audits address generalization, memorization, and privacy.

Latest papers

1,278 papers90 papers in the last four weeks, down 44% on the four weeks before. 0.7% of all new papers.

Nov 19, 2026

MixReact: Physics-Based Reactive Humanoid Control from Multimodal Cues via a Discrete-Continuous Mixture of Transformers

Humanoid agents in everyday environments are expected to react to diverse cues as humans do, with semantically appropriate and physically executable behaviors. Existing cue-driven reaction generation methods support only a narrow range of cue modalities and are limited to purely kinematic generation, leaving the resulting reactions susceptible to physical artifacts. To this end, we present MixReact, the first unified framework for physics-based reactive humanoid control from text, audio, and video cues, integrating a Discrete--Continuous Mixture of Transformers (DisCo-MoT) for state prediction with a History-aware Action Transformer (HAT) for physical execution. DisCo-MoT couples a discrete masked-modeling branch with a continuous autoregressive diffusion branch via one-way mixture attention within a MoT architecture. Guidance from discrete context mitigates collapse into averaged behaviors in one-to-many reaction generation, while direct prediction in continuous state space bypasses the decoding bottleneck of discrete representations. HAT then maps the predicted next state and observed execution history to executable actions, enabling robust closed-loop control in physics-based simulation. Experiments across all three modalities demonstrate consistent improvements over the baselines in fidelity, diversity, and physical consistency, while ablations validate the complementary contributions of the components. Our code will be released to facilitate future research. Project page: https://binlee26.github.io/MixReact_Page/
Nov 19, 2026

Keystone: Judging Previews at Key Steps for Test-Time Search in Diffusion Models

Text-to-image diffusion and flow models yield a preview, their clean-image estimate, at every denoising step. When they miscount, drop or misplace objects, test-time search is in a bind: its verifiers are reward models built for finished images, so judging early misreads blurry previews, and judging late, as best-of-N does, pays to fully denoise doomed candidates. We find that the image is not the problem: the outcome is settled as denoising starts, by the seed or by a stochastic sampler's first steps, and already shows in early previews; what fails on them is the judge, as reward models read them near chance on several kinds of prompts. We introduce Keystone, a training-free judge of previews that compiles a prompt into typed questions about its objects, counts, colours and positions and has an off-the-shelf multimodal decision model answer them. Asked this way, the judge reads counts and positions from early previews better than reward models read them from finished images. Searching at these key steps on GenEval with SD3.5-Medium, Keystone's adaptive search and pruning significantly outperform best-of-8 with every other verifier at under half its generator compute, and as a drop-in verifier there, the Keystone score significantly improves best-of-N, progressive seed pruning and FK steering; without retuning, both searches stay significantly above the strongest other verifier's best-of-N on average over four base models and three benchmarks. Keystone thus moves test-time judgment to the key steps where the outcome already shows.
Nov 18, 2026

CoRelate: Coordinated Human-Object Motion Synthesis through Interaction Relations

Human-object motion synthesis requires coordinated human participation as objects move toward specified goals. However, reaching object targets does not ensure that contact is established and maintained, while adjustments that improve contact may interfere with task completion. We introduce CoRelate, a diffusion framework that coordinates full-body human-object motion through explicit interaction relations and task feedback for rigid and articulated objects. We represent intended participation through hand-part relations that specify the participating hand, its part-relative reference position, and participation interval. These relations serve both as generation conditions and as references for measuring interaction realization. At each denoising step, a temporal refiner trained jointly with the generator combines relation deviations, task residuals, and motion features from the current prediction to jointly update human and object motion across the sequence. After generation, temporal motion correction addresses remaining contact failures by optimizing local motion segments while preserving object trajectories and protected human motion. Experiments demonstrate improved object-goal satisfaction, contact F1, and complete interaction success over the evaluated baselines. Ablation studies support the benefits of relation conditioning and the complementary roles of task and relation feedback.
Nov 18, 2026

Training Denoising Models of Any Architecture Without Stored Images

Can a pretrained denoising model replace the dataset used to train it? We investigate this question through teacher-supervised training, showing for the first time that continuous-time diffusion and flow models of any architecture and size can be trained from random initialization without stored images. Our students regress the teacher's velocity on a rolling pool of noisy samples, advanced along trajectories of random step counts to obtain independent, dense and approximately continuous noise levels. This saves considerable storage for large datasets and applies even when the data are unreleased or private. Our experiments include 12 teachers of 6 families and 6 datasets. Our method outperforms standard training of the same network on CIFAR-10 for students 4x smaller than the teacher, comes within 0.75 to 1.92 FID of it at the teacher's size and surpasses training on teacher-generated images and step reduction methods without teacher-based initialization. Our code is available at https://github.com/mever-team/ddm.
Nov 18, 2026

FDM+: A Framework for Constrained Control of the 2D Radiative Transport Equation

Applying the frozen diffusion model (FDM) [29] to the constrained control of the 2D radiative transport equation presents several challenges. These include the dimensional difference between control variables and state functions, the resolution mismatch between traditional numerical solvers and deep learning models, and solution non-uniqueness. To address these obstacles, we propose a novel approach, FDM+, which includes three key improvements. Firstly, a decoupled prediction architecture is introduced to address the dimensional heterogeneity. Secondly, an efficient parameterization strategy resolves the resolution mismatch. Meanwhile, we derive a total error bound to analyze how the parameterization error and the model estimation error jointly affect the performance. Thirdly, the conditional distribution learned by the diffusion model is leveraged to perform design space exploration for multi-solution problems. A series of numerical experiments demonstrate that FDM+ outperforms existing methods in reconstruction accuracy and design flexibility.
Nov 18, 2026

WaveView: The Novel View Is in the Noise

Generative novel view synthesis (NVS) employs a diffusion model to render a static or dynamic scene from a target camera view. Explicit warp methods complete a reprojection of the source with video diffusion models, and implicit methods encode the camera parameters or 3D cues. Both add inference overhead, and the generated layout often drifts from the ground truth. We present WaveView, which injects the warped source into the phase of the initial noise of a pretrained generator. The Dual-Tree Complex Wavelet Packet Transform (DT-CWPT) allows local and anisotropic phase injection. The generator is thus constrained less in disoccluded regions and along the epipolar line. WaveView keeps the architecture and adds negligible overhead. On the DL3DV benchmark, it outperforms all baselines in every overlap bin. In the hardest bin, it gains 0.84 dB PSNR on the covisible region over MetaView, which shares its generator, while running 2.1 times faster. For dynamic NVS, fine-tuning AnyView with the same noise raises its zero-shot PSNR on AnyViewBench from 11.58 to 12.35 dB.
Nov 18, 2026

ProxySearch: Decoupled Inference-Time Scaling for Diffusion Models via Asymmetric Noise-Rank Transfer

Noise search at inference time improves text-to-image diffusion but requires many multi-step teacher rollouts per query. Distilled few-step models are now routinely released, yet their potential as proxies inside noise search is uncharacterized. We measure proxy-teacher rank agreement across three distillation regimes and three backbones (DiT, MMDiT, UNet) and find a sharp asymmetry. On our main pool, closed-form image statistics transfer from the proxy to the teacher (Spearman ρρ up to 0.80), while learned semantic scorers do not (ρ≤ρ\le 0.12). Paired null controls trace this rank to the first denoising step for coarse statistics and to distillation for finer ones. We use the asymmetry for noise search under a joint constraint: a semantic reward specifies content, and a numeric target specifies image statistics. In PROXYSEARCH, the proxy filters candidates by statistical match in 1 step, and the teacher returns the survivor with the highest semantic reward. No weight trades one objective against the other. On FLUX-1024, PROXYSEARCH tightens the statistical target by 21-23% at the same number of teacher trajectories, with no detected loss on the optimized reward.
Nov 18, 2026

Training-Free Domain Generation for One-Shot Federated Learning with Server-Side Pre-trained Diffusion Models

One-shot federated learning (OSFL) has attracted increasing attention for its ability to substantially reduce communication overhead by limiting client-server interaction to a single communication round. Nevertheless, this limited interaction makes learning from heterogeneous client distributions challenging. Diffusion-based OSFL approaches seek to address this challenge by synthesizing data that approximate client distributions. However, existing diffusion-based approaches often impose substantial computational costs on clients and lack effective visual guidance to generate data aligned with client distributions. To address these limitations, we propose FedDoG, a training-free Domain Generation framework that leverages server-side pre-trained diffusion models to synthesize data aligned with client domains. FedDoG extracts compact domain representations using a frozen image encoder, substantially reducing client-side computational and communication costs. These representations are incorporated into the diffusion model through a pre-trained image-conditioning adapter, thereby providing explicit visual guidance for domain generation. To further improve the quality of synthetic data, FedDoG adapts the strength of visual conditioning according to domain concentration and filters generated samples based on class consistency and predictive entropy. Experiments on three benchmarks demonstrate that FedDoG generates higher-quality synthetic images and outperforms existing OSFL baselines in classification accuracy. Meanwhile, FedDoG reduces client upload and computation time by 99.9% and 91.5%, respectively. The code is available at https://github.com/YXTong17/FedDoG.
Nov 18, 2026

Variable Length Diffuse Everything: Arbitrary Length Generative Modeling on Arbitrary State Spaces

Diffusion and flow-based generative models have achieved remarkable results across diverse applications, but most assume data of fixed-dimension/length. Meanwhile, real-world problems such as layout generation often involve variable-length, multimodal data, such as continuous bounding box coordinates paired with discrete types. Existing methods typically address variable length by padding data to a fixed size or by introducing specialized objectives to model lengths explicitly, adding complexity and limiting generality. We propose Variable Length Diffuse Everything (VL-DE), a simple and general framework for variable-length multimodal generation on arbitrary state spaces. Instead of padding, VL-DE handles variable length through an insertion-based mechanism grounded in a principled theoretical framework, yielding a tractable training objective. In addition, by utilizing the fact that VL-DE is based on a generator with an additive form, we develop a higher-order sampler by appropriately handling the time dependence. Extensive experiments on layout and molecule generation demonstrate the effectiveness of our approach.
Nov 18, 2026

Image-Condition-Controlled Diffusion Network with Large-Scale CT Pretraining for CT-Free Attenuation Correction in Dynamic Cardiac PET

Positron emission tomography (PET) provides quantitative information on myocardial perfusion. In dynamic PET, tracer kinetics are tracked through a series of reconstructed time frames. Reliable attenuation correction is essential for accurate measurements. Clinical attenuation maps (μμ-maps) are typically derived from computed tomography (CT), which adds radiation exposure and may introduce PET-CT misregistration due to respiratory and cardiac motion. Moreover, supervised CT-free methods rely on paired non-attenuation-corrected (NAC) PET and CT data, which are scarce, whereas large-scale unpaired CT data are more readily available. To address these challenges, we propose Large-scale Unpaired CT-Informed Diffusion (LUCID), a framework that transfers anatomical priors learned from 10,000 chest CT volumes in the public CT-RATE dataset to μμ-map synthesis for dynamic cardiac 82^{82}Rb PET. First, a convolutional neural network (CNN) predicts an initial μμ-map from each NAC PET frame. Second, the CT-pretrained diffusion backbone refines a low-noise perturbation of the CNN prediction through a single denoising step, conditioned on the corresponding NAC PET frame via ControlNet. Finally, frame-wise predictions are averaged voxel-wise to obtain the final μμ-map. We further introduce the voxel-wise temporal standard deviation (TSD) to quantify inter-frame variability without requiring a reference μμ-map. We evaluated LUCID on an in-house dataset of 102 patients and found that it achieved the best quantitative accuracy while maintaining competitive inter-frame consistency. Ablation studies showed that refinement with a diffusion backbone pretrained on in-house CT data from the training patients was less accurate than the CNN prediction alone, highlighting the importance of large-scale CT pretraining.
Nov 17, 2026

Error accumulation in reverse diffusion with an imperfect score: beyond path-space analyses

Diffusion modeling is a popular method of sampling from a complex, high-dimensional distribution pp using learned approximations of its (noised) score functions. Generating a single sample requires simulating the reverse diffusion process, which requires many queries to the learned scores. Thus, a priori, the learning errors can easily compound. Indeed, in existing theoretical analyses, the sampling error (as measured in total variation distance to pp) accumulates polynomially with the horizon of the stochastic process. This is a consequence of bounding the error of the output law by the error of the law of the entire reverse diffusion path. In this paper, we demonstrate that by circumventing the "path-space" analysis, the error accumulation of the diffusion sampler can be mitigated when pp satisfies a Poincaré inequality. Moreover, we achieve this result as a special case of a broader framework that applies to any distribution pp, but relaxes the error metric from total variation distance to a weaker metric capturing all sample statistics that do not "witness" the multimodality of pp. We then zoom out from analyzing the diffusion sampler to the information-theoretic question: given a learned score with some error bound, how accurately can we identify pp? We show that, for any given bound on the Poincaré constant of pp, the diffusion sampler already nearly achieves optimal total variation error.
Nov 17, 2026

WoRD: Diffusion-Based Posterior Inference for Online Goal Recognition

Goal recognition is the task of inferring an agent's intended goal from its observed behavior. Most existing work focuses on high-level symbolic recognition, abstracting away the continuous motion through which agents reach their goal. Methods that do reason over continuous trajectories rely on an explicit dynamics model that is hard to acquire and rarely transfers across settings. Neither approach explicitly learns how agents move toward each candidate goal, limiting accuracy on ambiguous trajectory prefixes where such knowledge is most needed. We present WoRD (Waypoint Recognition with Diffusion), which uses a learned trajectory prior to infer the agent's goal from partial observations. At each step, WoRD runs one warm-started denoising chain per candidate goal, steers each chain toward the observed trajectory through a closed-form log-likelihood gradient, and reweights the resulting samples by Bayes' rule into a calibrated posterior over goals. We evaluate WoRD on simulated robot settings, pedestrian trajectories, and vehicle trajectories using four metrics targeting distinct failure modes. Across all three settings, the same inference procedure preserves calibrated multimodal beliefs through ambiguous prefixes where filter- and classifier-based observers collapse to a single mode or grow overconfident.
Nov 17, 2026

Capability Composition in Heterogeneous Teams via Barrier-Guided Diffusion Sampling

The agents in a heterogeneous team can cooperate to accomplish tasks that exceed their individual capabilities. The team trajectories for a cooperative task can be generated by a generative model, provided that the model has been trained on demonstrations of the agents cooperating. Yet, demonstrations of cooperation are costly and may not exist for tasks that the team has never performed. The conditions under which the agents can combine their capabilities, however, are physical properties of the team and are known before any demonstration is collected. We encode the conditions as constraints on the team trajectories, and we enforce the constraints while sampling from generative models trained on demonstrations of each agent alone. We prove that the proposed method can generate every team trajectory that satisfies the constraints, although the demonstrations contain no cooperation. Hence the constraints determine which team trajectories can be generated, and the demonstrations determine how likely each team trajectory is. Numerical results on two cooperative tasks show the effectiveness of the method, which generates cooperative team trajectories that satisfy the constraints and requires no retraining when agents join or leave the team. Finally, we implement the method on a multi-robot testbed.
Nov 17, 2026

Hierarchical Diffusion Transformer for End-to-End Image Generation

Pixel-space diffusion learns visual representations and image generation jointly in a single end-to-end training stage, without a separately pretrained visual encoder. However, plain pixel-space Transformers still trail leading two-stage latent diffusion models in generation quality and training efficiency. We hypothesize that part of this gap arises from retaining one token per patch throughout the backbone, which does not explicitly encourage a visual hierarchy that summarizes related content at progressively coarser levels. To address this, we introduce Hierarchical Diffusion Transformers (HiT), which progressively group related tokens and represent each group with a single token. The main Transformer bottleneck processes these compact group tokens, and ungrouping restores features to their original image positions. A compression budget regulates the average compression ratio across the batch, allowing group memberships and counts to vary across images and diffusion timesteps during both training and generation. The model exhibits emergent semantic structure, learning to group coherent objects and object parts without segmentation labels. In text-to-image generation, HiT further learns to group caption words with image patches depicting the same concepts, without region annotations. HiT achieves 1.48 FID on ImageNet 256×256256\times256 and reaches comparable generation quality to JiT using approximately 40% of its cumulative training FLOPs. For text-to-image generation, HiT exceeds UniLiP's average score across five benchmarks with roughly 6×6\times fewer denoiser parameters and 3×3\times fewer forward FLOPs. These results suggest that learning a visual hierarchy within the generator breaks down the barrier to stronger image generation without external visual pretraining.
Nov 17, 2026

Harnessing Image Diffusion Prior for Photo-Realistic Video Restoration

Recent image generation models produce photo-realistic results with rich details, but transferring their powerful priors to video restoration remains difficult due to temporal inconsistency from stochastic detail synthesis. Existing video restoration methods improve temporal coherence but often sacrifice visual fidelity, leaving a gap between image-level quality and video-level stability. We propose a generative video restoration method that bridges this gap by enabling temporally consistent detail synthesis from image generation priors. Our method combines noise energy rebalancing to suppress structure-disturbing low-frequency stochasticity, motion-aligned noise warping to propagate details along object motion, a self-supervised denoising encoder to mitigate encoder-induced temporal uncertainty, and decoupled temporal attention to model cross-frame dependencies. Together, these components preserve fidelity, enhance fine-grained realism, and maintain temporal consistency. Extensive experiments show that our method substantially outperforms existing video restoration approaches in visual quality, detail richness, and temporal stability.
Nov 17, 2026

Temporal Diffusion Models for Joint Signal-Channel Estimation under Gaussian and αα-Stable MIMO Channels

Diffusion models have emerged as powerful tools for learning complex data distributions and solving inverse problems. However, most diffusion models rely on the Gaussian distribution, which does not accurately represent heavy-tailed data. In this work, we study diffusion-based reconstruction under Gaussian and symmetric αα-stable fading channels, while also exploring the benefits of using possible temporal dependence in the channel. We consider two variants: pilot-based channel-only estimation and joint image-channel recovery for MIMO systems. Our approach is based on several standard and score-based diffusion models for both Gaussian and αα-stable distributions. We refine these models by guiding the reverse diffusion process using the observation likelihood and show how a single trained diffusion model can be reused for the location-scale family of distributions without retraining. In blind MIMO recovery, separate diffusion models are used for the image and channel, and their estimates are combined through likelihood guidance. For each setting, we compare independent and temporal reconstruction, where information from previous observations is used to improve the estimation. We show that, under a time-dependent channel model, such a temporal reconstruction improves channel estimation and the quality of the reconstructed image. Finally, we show that, for the same heavy-tailed channel data, using an αα-stable diffusion prior provides better channel estimation and image reconstruction than using a Gaussian diffusion prior.
Nov 17, 2026

Towards Data-Efficient, Real-Time World Models

State-of-the-art world models have been used to produce sequences of gameplay that accord with provided user-input actions. However, high quality, consistent gameplay generation often comes at the cost of inference speed, making real-time interactive play and ideation challenging. The substantial quantities of data used to train state-of-the-art models have also made custom training of these models appear impractical, limiting many of their potential uses in creative fields. In this work we demonstrate a major step towards efficient world models: achieving both fast high-quality inference, and improved data efficiency. This is done by replacing the next-token prediction paradigm with a two-stage discrete diffusion-based approach, introducing a lightweight refinement transformer which carries out iterative masked predictions. Building on an autoregressive world model capable of generating highly consistent and complex sequences over minutes (Kanervisto et al., 2025), we achieve substantial model speed-up (79x) with a modest quality trade-off. Additionally, we demonstrate that this model can be trained to reproduce high-quality, consistent gameplay using a dataset containing approximately one day (28 hours) of gameplay data. Our results demonstrate progress towards removing barriers to the use of world models in settings where real-time inference and data efficiency are essential.
Nov 17, 2026

Do Diffusion Models Learn to Generalize Basic Visual Skills?

Diffusion models generate high-quality images, but how well they generalize is hard to measure. Current evaluations rely on human preference and prompt-following benchmarks for text-to-image models and on distribution-level metrics such as Fréchet Inception Distance (FID) and Inception Score (IS) for class-conditional models, but these cannot tell whether models learn basic visual skills or largely match training patterns. To address this gap, we introduce the Visual Generative Lab (VGL), a controlled experimental framework for studying how diffusion models learn and generalize basic visual skills, including size, position, rotation, count, color, shape, and their compositions. By generating synthetic training data for each skill from explicit rules, VGL allows precise measurement of generalization with task-specific, rule-based metrics. Training models from scratch on these data reveals three patterns in this controlled setting: (1) with our default design, extrapolation is limited: size and position degrade within a few pixels of the training range, count reaches only about 50% accuracy on out-of-range counts, and rotation extrapolation accuracy varies widely across seeds; (2) for compositional generalization, the number of distinct basic visual skill combinations matters more than dataset size; (3) skills mostly fail independently, except position. Code and data-generation scripts: https://github.com/AmishSethi/visual-generative-lab.
Nov 17, 2026

Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers

Latent diffusion models rely on a decoder to map denoised latents to pixels, adding complexity and inference cost. We introduce Decoding Without a Decoder (DWD), which lets a finetuned diffusion transformer produce final pixels directly through a linear head. Unlike previous methods, where the denoising output leads to final pixels, DWD separates the two roles: the original denoising head continues to drive sampling, then the new DWD head predicts the pixels at the final step. We extend pretrained latent models with DWD efficiently by finetuning only the last few transformer blocks, training the DWD head with pixel quality losses while anchoring the denoising head to the pretrained model to preserve the sample trajectory. Across latent image and video generators, DWD removes the decoder while maintaining or even improving pixel quality. On Wan video generation, it visibly reduces detail distortion and flickering, lowering warping error by 23%. Remarkably, the same separation also improves pixel diffusion models that do not even use a decoder: a final-step DWD head produces substantially better texture detail in text-to-image generation, and improves ImageNet FID from 1.55 to 1.43, a new state of the art among pixel flow models.
Nov 17, 2026

What Does Weight-Space Programming Add to Diffusion Editing? Program Construction versus Score-Guided Refinement

Weight-space programming (WSP) lets a language model construct executable weight updates for another network. We study diffusion concept editing through an interface in which programs define training contexts, preservation constraints and update directions, while a shared numerical fitter selects their coefficients. We compare generated programs with published editors, a shared-panel hand-written baseline and budget-matched search over conventional edits on 24 TIMED test edits for Stable Diffusion 1.4. The preregistered rewrite pipeline shows no advantage over the strong baseline. In an exploratory comparison, one-shot Astra programs improve fresh-scene VLM joint success, which requires subject presence together with whole-image target matching, by +10.0 points over the shared-panel baseline, while CLIP relative match shows no detectable change. These programs organise adjustable directions around protected text-state subspaces. Probing and rewriting the Astra programs then raises CLIP by +6.2 points while joint success changes by -9.7. On a fixed 144-image sample, five additional evaluators give point estimates in the same directions under whole-image target matching; checks of subject-local change give less decisive differences. Together, these results identify useful behaviour in complete generated programs and show that score-guided revision need not preserve it: program construction and revision should be evaluated as distinct stages of a WSP pipeline.
Nov 17, 2026

An Optimization Framework for Denoising Generation

Noise-unconditioned denoising learns a single score, so generation can be viewed as gradient ascent on a fixed log-density landscape. This optimization interpretation leaves two questions: do trajectories approach meaningful data, and do different initializations retain diverse outcomes? We establish both phenomena for empirical and population multi-scale Gaussian-mixture objectives. Under explicit posterior and landscape conditions, ascent converges almost surely within O(mσmin⁡2/τ)O(mσ_{\min}^2/τ) of an mm-dimensional reference manifold with reach at least ττ. We derive a source--sink identity for the same fixed field and use it to prove finite-resolution regional coverage. In an affine construction, we further identify a positive-bandwidth regime where even empirical-score optimization can avoid training-sample memorization while preserving diversity at the training resolution. The same fixed objective naturally supports penalty-guided controllable generation. Experiments support our predicted mechanisms, show improved robustness under large-step guidance, and demonstrate downstream gains consistent with finite-data generative generalization.
Nov 17, 2026

DSADiff: Data Structure-Aware Diffusion for Crystal Structure Prediction via Mitigating Complexity- and Imbalance-Induced Biases

Crystal Structure Prediction (CSP) is an important step in accelerating the generation and discovery of new materials. In recent years, diffusion-based CSP methods have made significant progress. However, existing models still suffer from generation bias when facing diverse materials spaces, while systematic approaches for tracing the sources of such performance variations remain limited. To address this issue, we propose LAS (Layered Analysis Strategy), a mechanism analysis framework that identifies crystal complexity and elemental composition imbalance as two key factors affecting generation difficulty and further reveals their relationships with model generation performance. Building on these findings, we further propose DSADiff (Data Structure-Aware Diffusion), a data structure-aware diffusion model that transforms the key factors identified by LAS into a unified generation difficulty representation and uses it to jointly drive feature modeling and training strategy optimization. This enables the model to adapt its representation and generation capability to samples with different data structures and difficulty levels, thereby mitigating learning bias across samples. Experiments on three CSP benchmark datasets, including Carbon-24, demonstrate that DSADiff consistently improves both Match Rate and RMSE, with a maximum Match Rate improvement of 9%, surpassing existing state-of-the-art methods and achieving new state-of-the-art performance. This work deepens the understanding of diffusion-based CSP models from the perspective of generation bias and provides a new direction for designing less biased generative models for complex materials spaces.
Nov 14, 2026

Memorization Lives in a Few Attention Heads: Mitigating Memorization in Diffusion Models via Projection-Based Suppression

Diffusion models can memorize and reproduce training samples during inference, raising serious privacy and copyright concerns. Existing inference-time mitigation methods often suppress conditional information globally, which may reduce memorization but can also harm semantic alignment and image quality. In this paper, we revisit memorization from the perspective of cross-attention heads. Through systematic analysis and targeted ablation, we find that memorization is mainly driven by a small subset of shallow-layer attention heads with abnormally strong semantic control, rather than by the entire cross-attention mechanism. Motivated by this finding, we propose Projection-based Top-k Attention Head Suppression, a targeted inference-time mitigation method. Our method first identifies the top-k attention heads most related to memorization using a semantic-control score, and then removes their condition-specific components through directional projection between conditional and unconditional representations. This design selectively weakens memorization-related features while preserving useful semantic information. Extensive experiments on different diffusion models and datasets show that our method effectively reduces memorization and achieves a better trade-off between memorization mitigation and image quality than existing approaches.
Nov 14, 2026

dit.xpu : Taming DiT Sparse Inference on Mobile SoCs

Diffusion Transformers (DiTs) enable on-device image generation and editing but incur high attention costs at high resolutions. Existing sparse attention reduces computation, yet dynamic sparsity introduces preprocessing and irregular data accesses that poorly fit mobile NPUs. We present DiT.XPU, a sparse DiT inference framework for heterogeneous mobile SoCs. DiT.XPU combines tile-aligned sparse attention with anchor-based error compensation by exploiting slowly varying dense-sparse output residuals across denoising steps. Heterogeneous XPU collaboration overlaps NPU dense attention with GPU sparse calibration, while mask-guided dataflow reduces data transfers and layout conversion by overlapping data movement with computation. Across five DiT models on two Snapdragon smartphones, DiT.XPU achieves up to 1.56x end-to-end denoising speedup over the Hexagon-NPU backend and 11.5x attention-kernel speedup over the fastest alternative, while preserving generation quality close to dense attention. Code is available at https://github.com/IF-LAB-PKU/dit.xpu.
Nov 13, 2026

Role-Separated Diffusion Transformers: Global Rule Isolation and Deterministic Sampling Manifold Compression

We present Role-Separated Diffusion Transformers (RoSeDiT), a new diffusion transformer architecture that isolates the global rule from the image tokens. RoSeViT showed in visual reasoning that a vanilla ViT spreads a shared rule over all LL tokens at an entropy of order LdLd, whereas an enforced controller route reduces it to order dd. In class-conditional DiT and SiT, the class and the noise level are explicitly conditioned by adaLN, so a RoSeViT-style controller channel on the ViT side is insufficient to establish the full role separation. By introducing the adaLN listening mechanism, which makes adaLN compute under the controller's command, RoSeDiT brings the entropy theorem and role separation into the latent diffusion domain. Under the same SiT training and evaluation protocol at the same level of parameter count, RoSeDiT surpasses SiT at all four model scales at 400K steps, and RoSeDiT-L/2 even surpasses SiT-XL/2's 17.2 FID with 15.34 and 32% fewer parameters. RoSeDiT-XL/2 surpasses SiT-XL/2's 7M-step FID of 8.3 at 1.4M steps with 7.98, using one-fifth as many training steps. We further observe that models with the same denoising loss can have a large FID gap, which indicates a misalignment between the denoising objective and the image manifold. We formalize this phenomenon into Deterministic Sampling Manifold Compression (DSMC), a two-stage training schedule with which RoSeDiT-XL/2 surpasses SiT + REPA's 4M-step FID of 5.9 at 2.4M steps with 5.84.
Nov 13, 2026

Restoring Diversity in Distilled Image and Video Generators by Probing Activations

Distilling diffusion models into few-step student models substantially reduces sampling latency, but induces severe diversity collapse: samples generated from different random seeds show minimal variation. In this work, we investigate the root cause of this failure by probing the internal representations of student models relative to their teachers. We find that distilled students maintain total activation variance comparable to their teachers, but concentrate it into substantially fewer directions. This directional collapse is localized to an identifiable intermediate block at the initial denoising step. Consequently, diversity collapse stems from a misallocation of activation variance rather than an inherent loss of generative capacity. Guided by this finding, we propose ReDiverse to reinstate the missing variance directions. At training time, ReDiverse-FT fine-tunes a lightweight adapter that regresses student activations onto the teacher's, reviving the teacher's feature variance characteristic. At inference time, ReDiverse-TF introduces a training-free mechanism that steers activations along precomputed teacher principal directions. Across image and video distillation benchmarks, ReDiverse consistently restores sample diversity while preserving generation quality, further supported by the user study.
Nov 13, 2026

JaRS: Contrast-Free Localization of Memorized Regions in Diffusion Models via Jacobian Row Sums

Diffusion models sometimes memorize and reproduce regions of their training images, raising privacy and copyright concerns. We aim to localize these regions. Existing contrast-based methods use the diagonal of the denoiser Jacobian, which reflects the posterior variance of each pixel. However, a small diagonal entry is ambiguous: it signals either memorization or averaging over many neighboring pixels in smooth or predictable regions. To resolve this ambiguity, these methods subtract a reference map, which requires a second evaluation with a null prompt or an earlier checkpoint. Unconditional models have no null prompt, and earlier checkpoints are unavailable in general. We show that Jacobian row sums (JaRS), unlike the diagonal, separate memorization from local averaging without a reference map. Pairing JaRS with an image-level detection score that uses the same statistic at an early denoising step yields JaRS+. Across Stable Diffusion v1.4 and v2.1, Realistic Vision v5.1, and an unconditional diffusion Transformer (DiT), JaRS+ outperforms contrast-based localization methods.
Nov 13, 2026

Analyzing Learning-Dynamics Across Loss and Prediction Choices for Diffusion Models

Denoising generative models predict clean data (xx), noise (εε), or velocity (vv), while training losses can be measured in any of these spaces. Although the resulting nine combinations are algebraically related, they can yield substantially different denoising accuracy and sample quality. We disentangle the effects of prediction and loss spaces. For a fixed prediction target, we show that changing the loss space is exactly equivalent to timestep reweighting and provide a heuristic analysis of how this reweighting affects generation quality. We then show that a rank-mm linear bottleneck compressing a DD-dimensional noisy input imposes an irreducible squared-error floor of D−mD-m on noise and velocity prediction, whereas clean-data prediction can exploit low-dimensional data structure. Combining this bound with prediction-dependent error conversion yields distinct necessary conditions on bottleneck capacity. Experiments with U-Net and JiT across CIFAR-10, Imagenette, and ImageNet-256 support these theoretical predictions. Under compression, prediction choice strongly affects both denoising accuracy and sample quality, with noise-prediction errors closely matching the theoretical lower bound. Our results provide a quantitative understanding of how architectural bottlenecks interact with prediction and loss choices in denoising generative models.
Nov 13, 2026

MASTARS: Multi-Agent Sequential Trajectory Augmentation with Return-Conditioned Subgoals

The performance of offline Multi-Agent Reinforcement Learning (MARL) is limited by the quality of the fixed offline dataset, motivating trajectory augmentation. However, augmenting multi-agent trajectories presents a fundamental challenge of the diversity-coordination trade-off: independent generation improves diversity but lacks coordination, while joint generation induces coordination but lacks diversity, reproducing the training distribution. To address this challenge in multi-agent data augmentation, we introduce MASTARS, a novel diffusion-based framework that generates multi-agent trajectories sequentially across agents while enforcing cross-agent consistent coordination via the method of inpainting. MASTARS produces trajectories that are diverse, coordinated and globally coherent. Experiments on various benchmarks such as MPE, SMAC, and SMAC-v2 demonstrate significant performance gains across multiple offline MARL methods.
Nov 13, 2026

Complex Schrödinger Bridges

Diffusion bridges have emerged as versatile frameworks for facilitating transformations between arbitrary data distributions. However, these approaches are hindered by an immense optimization space and its underlying high-order structures. This paper explores a variant of diffusion bridges for solving real-world problems in complex vector spaces governed by scaling and rotation laws, such as Hamiltonian structures. To address this, we introduce conformal optimal transport (COT), a complex-valued transportation framework designed for physical systems with geometric and dissipative dynamics under a geometric lifting. Building on this, we propose a new class of dynamical diffusion bridges, called complex Schrödinger bridges (CSBs), whose derived algorithms can handle diffusion-based tasks using a conformal geometric prior. For scalability, we propose image-to-image CSB (I2^2CSB), a tractable diffusion-bridge technique with an analytic spiral posterior. Finally, we implement a physics-informed CSB algorithm within simulations of quantum electrodynamics, mimicking the logical behavior of quantum logic gates. Our extensive results validate the underlying theory while highlighting advantages of complex-valued diffusion.