Mohsen Ghafoorian

Denis Korzhenkov, Adil Karjauv, Animesh Karnewar et al.

Recently proposed pyramidal models decompose the conventional forward and backward diffusion processes into multiple stages operating at varying resolutions. These models handle inputs with higher noise levels at lower resolutions, while less noisy inputs are processed at higher resolutions. This hierarchical approach significantly reduces the computational cost of inference in multi-step denoising models. However, existing open-source pyramidal video models have been trained from scratch and tend to underperform compared to state-of-the-art systems in terms of visual plausibility. In this work, we present a pipeline that converts a pretrained diffusion model into a pyramidal one through low-cost finetuning, achieving this transformation without degradation in quality of output videos. Furthermore, we investigate and compare various strategies for step distillation within pyramidal models, aiming to further enhance the inference efficiency. Our results are available at https://qualcomm-ai-research.github.io/PyramidalWan.

Kai-En Lin, Alex Trevithick, Keli Cheng et al.

Portrait synthesis creates realistic digital avatars which enable users to interact with others in a compelling way. Recent advances in StyleGAN and its extensions have shown promising results in synthesizing photorealistic and accurate reconstruction of human faces. However, previous methods often focus on frontal face synthesis and most methods are not able to handle large head rotations due to the training data distribution of StyleGAN. In this work, our goal is to take as input a monocular video of a face, and create an editable dynamic portrait able to handle extreme head poses. The user can create novel viewpoints, edit the appearance, and animate the face. Our method utilizes pivotal tuning inversion (PTI) to learn a personalized video prior from a monocular video sequence. Then we can input pose and expression coefficients to MLPs and manipulate the latent vectors to synthesize different viewpoints and expressions of the subject. We also propose novel loss functions to further disentangle pose and expression in the latent space. Our algorithm shows much better performance over previous approaches on monocular video datasets, and it is also capable of running in real-time at 54 FPS on an RTX 3080.

Animesh Karnewar, Denis Korzhenkov, Ioannis Lelekas et al.

We introduce Neodragon, a text-to-video system capable of generating 2s (49 frames @24 fps) videos at the 640x1024 resolution directly on a Qualcomm Hexagon NPU in a record 6.7s (7 FPS). Differing from existing transformer-based offline text-to-video generation models, Neodragon is the first to have been specifically optimised for mobile hardware to achieve efficient and high-fidelity video synthesis. We achieve this through four key technical contributions: (1) Replacing the original large 4.762B T5xxl Text-Encoder with a much smaller 0.2B DT5 (DistilT5) with minimal quality loss, enabled through a novel Text-Encoder Distillation procedure. (2) Proposing an Asymmetric Decoder Distillation approach allowing us to replace the native codec-latent-VAE decoder with a more efficient one, without disturbing the generative latent-space of the generation pipeline. (3) Pruning of MMDiT blocks within the denoiser backbone based on their relative importance, with recovery of original performance through a two-stage distillation process. (4) Reducing the NFE (Neural Functional Evaluation) requirement of the denoiser by performing step distillation using DMD adapted for pyramidal flow-matching, thereby substantially accelerating video generation. When paired with an optimised SSD1B first-frame image generator and QuickSRNet for 2x super-resolution, our end-to-end Neodragon system becomes a highly parameter (4.945B full model), memory (3.5GB peak RAM usage), and runtime (6.7s E2E latency) efficient mobile-friendly model, while achieving a VBench total score of 81.61. By enabling low-cost, private, and on-device text-to-video synthesis, Neodragon democratizes AI-based video content creation, empowering creators to generate high-quality videos without reliance on cloud services. Code and model will be made publicly available at our website: https://qualcomm-ai-research.github.io/neodragon

Osman Ülger, Julian Wiederer, Mohsen Ghafoorian et al.

Graph neural networks have shown to learn effective node representations, enabling node-, link-, and graph-level inference. Conventional graph networks assume static relations between nodes, while relations between entities in a video often evolve over time, with nodes entering and exiting dynamically. In such temporally-dynamic graphs, a core problem is inferring the future state of spatio-temporal edges, which can constitute multiple types of relations. To address this problem, we propose MTD-GNN, a graph network for predicting temporally-dynamic edges for multiple types of relations. We propose a factorized spatio-temporal graph attention layer to learn dynamic node representations and present a multi-task edge prediction loss that models multiple relations simultaneously. The proposed architecture operates on top of scene graphs that we obtain from videos through object detection and spatio-temporal linking. Experimental evaluations on ActionGenome and CLEVRER show that modeling multiple relations in our temporally-dynamic graph network can be mutually beneficial, outperforming existing static and spatio-temporal graph neural networks, as well as state-of-the-art predicate classification methods.

Florian Langer, Jihong Ju, Georgi Dikov et al.

Digitising the 3D world into a clean, CAD model-based representation has important applications for augmented reality and robotics. Current state-of-the-art methods are computationally intensive as they individually encode each detected object and optimise CAD alignments in a second stage. In this work, we propose FastCAD, a real-time method that simultaneously retrieves and aligns CAD models for all objects in a given scene. In contrast to previous works, we directly predict alignment parameters and shape embeddings. We achieve high-quality shape retrievals by learning CAD embeddings in a contrastive learning framework and distilling those into FastCAD. Our single-stage method accelerates the inference time by a factor of 50 compared to other methods operating on RGB-D scans while outperforming them on the challenging Scan2CAD alignment benchmark. Further, our approach collaborates seamlessly with online 3D reconstruction techniques. This enables the real-time generation of precise CAD model-based reconstructions from videos at 10 FPS. Doing so, we significantly improve the Scan2CAD alignment accuracy in the video setting from 43.0% to 48.2% and the reconstruction accuracy from 22.9% to 29.6%.

Mohsen Ghafoorian, Denis Korzhenkov, Amirhossein Habibian

Transformer-based video diffusion models (VDMs) deliver state-of-the-art video generation quality but are constrained by the quadratic cost of self-attention, making long sequences and high resolutions computationally expensive. While linear attention offers sub-quadratic complexity, previous approaches have failed to match the expressiveness of softmax attention unless retrained at significant computational cost. We introduce Attention Surgery, an efficient framework that enables linear or hybrid attention in pretrained VDMs, eliminating the need for training from scratch. Inspired by recent advances in language models, our method combines a novel hybrid attention mechanism-mixing softmax and linear tokens-with a lightweight distillation and fine-tuning pipeline requiring only a few GPU-days. Additionally, we incorporate a cost-aware block-rate strategy to balance expressiveness and efficiency across layers. Applied to Wan2.1 1.3B, a state-of-the-art efficient transformer VDM and evaluated on VBench, VBench2.0 and a human preference study, Attention Surgery achieves competitive results. Furthermore, measurements of on-mobile latency, memory usage, and FLOPs demonstrate notable improvements in scaling behavior for longer videos. Project page is available at: https://qualcomm-ai-research.github.io/attention-surgery.

Aritra Bhowmik, Denis Korzhenkov, Cees G. M. Snoek et al.

Text-to-video diffusion models have enabled high-quality video synthesis, yet often fail to generate temporally coherent and physically plausible motion. A key reason is the models' insufficient understanding of complex motions that natural videos often entail. Recent works tackle this problem by aligning diffusion model features with those from pretrained video encoders. However, these encoders mix video appearance and dynamics into entangled features, limiting the benefit of such alignment. In this paper, we propose a motion-centric alignment framework that learns a disentangled motion subspace from a pretrained video encoder. This subspace is optimized to predict ground-truth optical flow, ensuring it captures true motion dynamics. We then align the latent features of a text-to-video diffusion model to this new subspace, enabling the generative model to internalize motion knowledge and generate more plausible videos. Our method improves the physical commonsense in a state-of-the-art video diffusion model, while preserving adherence to textual prompts, as evidenced by empirical evaluations on VideoPhy, VideoPhy2, VBench, and VBench-2.0, along with a user study.

Babak Ehteshami Bejnordi, Gaurav Kumar, Amelie Royer et al.

Jointly learning multiple tasks with a unified model can improve accuracy and data efficiency, but it faces the challenge of task interference, where optimizing one task objective may inadvertently compromise the performance of another. A solution to mitigate this issue is to allocate task-specific parameters, free from interference, on top of shared features. However, manually designing such architectures is cumbersome, as practitioners need to balance between the overall performance across all tasks and the higher computational cost induced by the newly added parameters. In this work, we propose \textit{InterroGate}, a novel multi-task learning (MTL) architecture designed to mitigate task interference while optimizing inference computational efficiency. We employ a learnable gating mechanism to automatically balance the shared and task-specific representations while preserving the performance of all tasks. Crucially, the patterns of parameter sharing and specialization dynamically learned during training, become fixed at inference, resulting in a static, optimized MTL architecture. Through extensive empirical evaluations, we demonstrate SoTA results on three MTL benchmarks using convolutional as well as transformer-based backbones on CelebA, NYUD-v2, and PASCAL-Context.

Mohsen Ghafoorian, Amirhossein Habibian

Recent advances in video diffusion models have shifted towards transformer-based architectures, achieving state-of-the-art video generation but at the cost of quadratic attention complexity, which severely limits scalability for longer sequences. We introduce ReHyAt, a Recurrent Hybrid Attention mechanism that combines the fidelity of softmax attention with the efficiency of linear attention, enabling chunk-wise recurrent reformulation and constant memory usage. Unlike the concurrent linear-only SANA Video, ReHyAt's hybrid design allows efficient distillation from existing softmax-based models, reducing the training cost by two orders of magnitude to ~160 GPU hours, while being competitive in the quality. Our light-weight distillation and finetuning pipeline provides a recipe that can be applied to future state-of-the-art bidirectional softmax-based models. Experiments on VBench and VBench-2.0, as well as a human preference study, demonstrate that ReHyAt achieves state-of-the-art video quality while reducing attention cost from quadratic to linear, unlocking practical scalability for long-duration and on-device video generation. Project page is available at https://qualcomm-ai-research.github.io/rehyat.

Erik Stammes, Tom F. H. Runia, Michael Hofmann et al.

Semantic segmentation is a task that traditionally requires a large dataset of pixel-level ground truth labels, which is time-consuming and expensive to obtain. Recent advancements in the weakly-supervised setting show that reasonable performance can be obtained by using only image-level labels. Classification is often used as a proxy task to train a deep neural network from which attention maps are extracted. However, the classification task needs only the minimum evidence to make predictions, hence it focuses on the most discriminative object regions. To overcome this problem, we propose a novel formulation of adversarial erasing of the attention maps. In contrast to previous adversarial erasing methods, we optimize two networks with opposing loss functions, which eliminates the requirement of certain suboptimal strategies; for instance, having multiple training steps that complicate the training process or a weight sharing policy between networks operating on different distributions that might be suboptimal for performance. The proposed solution does not require saliency masks, instead it uses a regularization loss to prevent the attention maps from spreading to less discriminative object regions. Our experiments on the Pascal VOC dataset demonstrate that our adversarial approach increases segmentation performance by 2.1 mIoU compared to our baseline and by 1.0 mIoU compared to previous adversarial erasing approaches.

Hessam Sokooti, Bob de Vos, Floris Berendsen et al.

We propose a supervised nonrigid image registration method, trained using artificial displacement vector fields (DVF), for which we propose and compare three network architectures. The artificial DVFs allow training in a fully supervised and voxel-wise dense manner, but without the cost usually associated with the creation of densely labeled data. We propose a scheme to artificially generate DVFs, and for chest CT registration augment these with simulated respiratory motion. The proposed architectures are embedded in a multi-stage approach, to increase the capture range of the proposed networks in order to more accurately predict larger displacements. The proposed method, RegNet, is evaluated on multiple databases of chest CT scans and achieved a target registration error of 2.32 $\pm$ 5.33 mm and 1.86 $\pm$ 2.12 mm on SPREAD and DIR-Lab-4DCT studies, respectively. The average inference time of RegNet with two stages is about 2.2 s.

Laurens Samson, Nanne van Noord, Olaf Booij et al.

Adversarial training has been recently employed for realizing structured semantic segmentation, in which the aim is to preserve higher-level scene structural consistencies in dense predictions. However, as we show, value-based discrimination between the predictions from the segmentation network and ground-truth annotations can hinder the training process from learning to improve structural qualities as well as disabling the network from properly expressing uncertainties. In this paper, we rethink adversarial training for semantic segmentation and propose to formulate the fake/real discrimination framework with a correct/incorrect training objective. More specifically, we replace the discriminator with a "gambler" network that learns to spot and distribute its budget in areas where the predictions are clearly wrong, while the segmenter network tries to leave no clear clues for the gambler where to bet. Empirical evaluation on two road-scene semantic segmentation tasks shows that not only does the proposed method re-enable expressing uncertainties, it also improves pixel-wise and structure-based metrics.

Hugo J. Kuijf, J. Matthijs Biesbroek, Jeroen de Bresser et al.

Quantification of cerebral white matter hyperintensities (WMH) of presumed vascular origin is of key importance in many neurological research studies. Currently, measurements are often still obtained from manual segmentations on brain MR images, which is a laborious procedure. Automatic WMH segmentation methods exist, but a standardized comparison of the performance of such methods is lacking. We organized a scientific challenge, in which developers could evaluate their method on a standardized multi-center/-scanner image dataset, giving an objective comparison: the WMH Segmentation Challenge (https://wmh.isi.uu.nl/). Sixty T1+FLAIR images from three MR scanners were released with manual WMH segmentations for training. A test set of 110 images from five MR scanners was used for evaluation. Segmentation methods had to be containerized and submitted to the challenge organizers. Five evaluation metrics were used to rank the methods: (1) Dice similarity coefficient, (2) modified Hausdorff distance (95th percentile), (3) absolute log-transformed volume difference, (4) sensitivity for detecting individual lesions, and (5) F1-score for individual lesions. Additionally, methods were ranked on their inter-scanner robustness. Twenty participants submitted their method for evaluation. This paper provides a detailed analysis of the results. In brief, there is a cluster of four methods that rank significantly better than the other methods, with one clear winner. The inter-scanner robustness ranking shows that not all methods generalize to unseen scanners. The challenge remains open for future submissions and provides a public platform for method evaluation.

Hans Meine, Grzegorz Chlebus, Mohsen Ghafoorian et al.

Various approaches for liver segmentation in CT have been proposed: Besides statistical shape models, which played a major role in this research area, novel approaches on the basis of convolutional neural networks have been introduced recently. Using a set of 219 liver CT datasets with reference segmentations from liver surgery planning, we evaluate the performance of several neural network classifiers based on 2D and 3D U-net architectures. An interesting observation is that slice-wise approaches perform surprisingly well, with mean and median Dice coefficients above 0.97, and may be preferable over 3D approaches given current hardware and software limitations.

Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka et al.

Convolutional neural networks have been successfully applied to semantic segmentation problems. However, there are many problems that are inherently not pixel-wise classification problems but are nevertheless frequently formulated as semantic segmentation. This ill-posed formulation consequently necessitates hand-crafted scenario-specific and computationally expensive post-processing methods to convert the per pixel probability maps to final desired outputs. Generative adversarial networks (GANs) can be used to make the semantic segmentation network output to be more realistic or better structure-preserving, decreasing the dependency on potentially complex post-processing. In this work, we propose EL-GAN: a GAN framework to mitigate the discussed problem using an embedding loss. With EL-GAN, we discriminate based on learned embeddings of both the labels and the prediction at the same time. This results in more stable training due to having better discriminative information, benefiting from seeing both `fake' and `real' predictions at the same time. This substantially stabilizes the adversarial training process. We use the TuSimple lane marking challenge to demonstrate that with our proposed framework it is viable to overcome the inherent anomalies of posing it as a semantic segmentation problem. Not only is the output considerably more similar to the labels when compared to conventional methods, the subsequent post-processing is also simpler and crosses the competitive 96% accuracy threshold.

Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi et al.

Deep learning algorithms, in particular convolutional networks, have rapidly become a methodology of choice for analyzing medical images. This paper reviews the major deep learning concepts pertinent to medical image analysis and summarizes over 300 contributions to the field, most of which appeared in the last year. We survey the use of deep learning for image classification, object detection, segmentation, registration, and other tasks and provide concise overviews of studies per application area. Open challenges and directions for future research are discussed.

Mohsen Ghafoorian, Nico Karssemeijer, Tom Heskes et al.

The anatomical location of imaging features is of crucial importance for accurate diagnosis in many medical tasks. Convolutional neural networks (CNN) have had huge successes in computer vision, but they lack the natural ability to incorporate the anatomical location in their decision making process, hindering success in some medical image analysis tasks. In this paper, to integrate the anatomical location information into the network, we propose several deep CNN architectures that consider multi-scale patches or take explicit location features while training. We apply and compare the proposed architectures for segmentation of white matter hyperintensities in brain MR images on a large dataset. As a result, we observe that the CNNs that incorporate location information substantially outperform a conventional segmentation method with hand-crafted features as well as CNNs that do not integrate location information. On a test set of 46 scans, the best configuration of our networks obtained a Dice score of 0.791, compared to 0.797 for an independent human observer. Performance levels of the machine and the independent human observer were not statistically significantly different (p-value=0.17).