Diffusion Models: From Crafting Realistic Videos and 3D Worlds to Deciphering Ancient Scripts

Latest 50 papers on diffusion models: Sep. 19, 2026

Diffusion models continue to redefine the landscape of generative AI, pushing boundaries in realism, efficiency, and real-world applicability. This digest dives into recent breakthroughs that leverage and extend these powerful models, showcasing their versatility from generating high-fidelity videos and complex 3D structures to assisting in historical decipherment and ensuring AI safety.

The Big Idea(s) & Core Innovations

At the heart of these advancements is a shared drive to make diffusion models more controllable, efficient, and robust across diverse data types and applications. A significant theme revolves around efficiently handling complex, high-dimensional data like video and 3D scenes. For instance, researchers from the University of California, Berkeley and Impossible, Inc. introduce Video DeltaNet, a video-native hybrid attention architecture that achieves a remarkable 14.5x speedup in livestream video generation by combining local Softmax attention with linear memory for distant context. This innovation is crucial for real-time applications, preserving fine-grained details while scaling efficiently.

Extending into the 3D realm, DimExDAM by researchers from SCK CEN (Belgian Nuclear Research Centre) and UCLouvain presents a hybrid diffusion-GAN framework that generates heterogeneous 3D geological microstructures from 2D images without needing 3D ground truth. This is a game-changer for fields like subsurface exploration, where 3D data is scarce, by using adversarial loss instead of denoising MSE loss for stable training. Similarly, Meta and Georgia Institute of Technology’s GazeDiT generates gaze-accurate eye images by spatially conditioning diffusion models, significantly improving downstream eye tracking performance. Their key insight is that label fidelity (how well generated data aligns with its labels) is more important than mere visual realism for synthetic data utility.

Controllability and safety are also paramount. CleanVideo from Mohamed bin Zayed University of Artificial Intelligence introduces an adaptive concept erasure framework for text-to-video models using a tri-modal gating mechanism. This allows precise intervention—determining where, when, and whether to erase undesired concepts—without compromising temporal coherence. Building on this, Southwest Jiaotong University’s DSS: Dynamic Semantic Steering and GRACE: Adaptive Concept Erasure offer training-free, inference-time concept erasure for text-to-image models. DSS enforces geometry-aware constraints via density-based semantic anchors, while GRACE uses semantically weighted sensitive subspace estimation and energy-driven dynamic gating for localized intervention. These methods ensure robust removal of sensitive content while preserving artistic intent and benign content.

Efficiency in inference is a persistent challenge. Imperial College London’s Accelerating Diffusion Sampling via Speculative Draft Trees proposes using draft trees to generalize speculative diffusion sampling, offering up to 8.3% acceleration over linear chains. Similarly, SAP and National University of Singapore’s ADSC (Adaptive Diffusion Step Controller) dynamically adjusts denoising steps based on prompt complexity, leading to up to 50% inference time reduction without retraining. For medical imaging, Zhejiang University’s Efficient 3D Whole-Body PET Image Denoising via Conditional Rectified Flow achieves ultra-low-dose PET denoising in roughly 30 seconds by using rectified flow and an optimized non-uniform sampling strategy, a drastic improvement over multi-hour DDPM baselines.

Beyond generation, diffusion models are proving adept at analytical and inverse problems. Waseda University and KDDI Research Inc.’s MyoFlow recasts HD-sEMG gesture recognition as anchor-tied rectified flow transport, enabling robust cross-session and cross-subject generalization with zero-shot recognition. In medical image analysis, Eurecat and Universitat de Barcelona’s MUMINS forecasts anatomical changes (e.g., tumor growth) and predicts per-voxel uncertainty in a single pass, addressing a critical need for prognostic tools.

Even human-computer interaction and historical preservation benefit. Wuhan University of Technology’s FROD assists oracle bone script decipherment by generating modern Chinese character candidates, using feature-matching-gated training and a Residual Denoising Diffusion Model to bridge ancient and modern scripts.

Under the Hood: Models, Datasets, & Benchmarks

The papers introduce or heavily rely on specialized models, datasets, and benchmarks to push the envelope:

• Video DeltaNet: Built on the MiniMax H3 backbone model, trained on 10,015 video clips, achieving 119.3x speedup for 14.3-second 768p video generation. Public code available at https://github.com/OpenVDN/vdn-minimax-h3.
• DimExDAM: A hybrid diffusion-GAN model, validated on the Savonnières Carbonate dataset and generalized across diverse microstructures (sandstone, illite, carbonate, Boom Clay). Code available at https://github.com/40uf411/DimExDAM.
• GazeDiT: Utilizes SegFormer for mask extraction during training and a physical eye renderer for diverse mask sampling at inference. Improves downstream performance on custom eye tracking tasks.
• CleanVideo: Evaluated on CogVideoX-2B, CogVideoX-5B, and HunyuanVideo models, using benchmarks like SafeSora and Ring-A-Bell. No public code mentioned.
• Zarya: A hybrid autoregressive-masked diffusion language model built on Qwen3, available in 0.6B, 1.7B, and 4B sizes. Code at https://github.com/ai-forever/zarya.
• Accelerating Diffusion Sampling via Speculative Draft Trees: Tested on Stable Diffusion 3 and Gaussian mixture targets. Code at https://github.com/marcellobullo/tree-specdiff.
• ADSC: Evaluated on Stable Diffusion v1-5 pipeline with DDIM and PNDM schedulers across COCO and DiffusionDB datasets. Leverages Hugging Face Diffusers codebase https://github.com/huggingface/diffusers.
• Efficient 3D Whole-Body PET Image Denoising: Conditional 3D rectified flow model, validated on UDPET and FAHZU datasets. Code at https://anonymous.4open.science/r/PET-Rectified-Flow/.
• FROD: Adapts Residual Denoising Diffusion Model with LightGlue for feature matching, trained on OBC-V, EVOBC, and HUST-OBC datasets.
• MUMINS: Uses a 3D U-Net-based diffusion framework, validated on PNG (Pulmonary Nodule Growth) and OASIS-3 brain MRI datasets. Code at https://github.com/aolivtous/MUMINS.
• CompArt: Introduces the CompArt dataset (80,032 WikiArt images with PoA annotations) and ArtDapter, a parameter-efficient adapter for diffusion models like Stable Diffusion. Code at https://github.com/jin-zhe/ArtDapter.
• LynnReal-Omni: A 32B shared multimodal diffusion transformer with a 27B Flash variant, using a dedicated VAE decoder distillation. Available on Hugging Face: https://huggingface.co/stdstu123/LynnReal-Onmi-flash-beta-0.1 and others. Code at https://github.com/LynnReal-AI/LynnReal-Omni.
• DNF-SR: Leverages Flux-Kontext image-editing model and evaluates with DINOv3. Code: https://github.com/SHH-Han/DNF-SR.
• SceneHI: A training-free framework for 3D texture synthesis, evaluated with the 3D-FRONT dataset. Code: https://github.com/Thanos-DB/SceneHI.

Impact & The Road Ahead

These advancements herald a new era of highly controllable, efficient, and robust generative AI. The ability to generate complex, multimodal outputs like consistent videos and 3D scenes with fine-grained control will transform fields from media production and virtual reality to scientific simulation and robotics. The focus on training-free methods and inference-time optimizations means these innovations can be rapidly integrated into existing diffusion pipelines, accelerating adoption and reducing computational costs.

The emphasis on AI safety and interpretability through concept erasure and rigorous certification (Certifying Concept Unlearning in Text-to-Image Diffusion Models from Imperial College London) is crucial for building trustworthy AI. Furthermore, bridging the gap between computational models and human cognition, as seen in Georgia Institute of Technology’s analysis of Diffusion Models and Concept Formation, promises deeper understanding and more human-aligned AI.

Looking ahead, we can anticipate further integration of physics-aware models (Physics-Aware Video Generation via Agentic Planning and Graph-Guided Optimization from University of Science, Ho Chi Minh City), more sophisticated multimodal understanding (Zero-shot video highlight detection based on text descriptions and synthetic images from Samsung AI Center), and adaptive learning for resource-constrained environments (AdaVLA: Adaptive Step Flow Matching for Training-free Acceleration of Vision-Language-Action Models from Sogang University). The journey towards universally capable and ethically sound generative AI continues at a breathtaking pace, driven by the ingenuity showcased in these recent diffusion model breakthroughs.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.