PyramidalWan: On Making Pretrained Video Model Pyramidal for Efficient Inference

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.

ReHyAt: Recurrent Hybrid Attention for Video Diffusion Transformers

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.

Neodragon: Mobile Video Generation using Diffusion Transformer

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

MoAlign: Motion-Centric Representation Alignment for Video Diffusion Models

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.

FastCAD: Real-Time CAD Retrieval and Alignment from Scans and Videos

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%.

InterroGate: Learning to Share, Specialize, and Prune Representations for Multi-task Learning

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.

PVP: Personalized Video Prior for Editable Dynamic Portraits using StyleGAN

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.

Multi-Task Edge Prediction in Temporally-Dynamic Video Graphs

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.

Find it if You Can: End-to-End Adversarial Erasing for Weakly-Supervised Semantic Segmentation

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.

3D Convolutional Neural Networks Image Registration Based on Efficient Supervised Learning from Artificial Deformations

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.