#AI Video

The Jenga Inverse Predictor (JIP-2) is a GPU-accelerated deep learning framework that reconstructs collapsed architectural structures using a physics engine and dual-stream ResNet-18 model. It predicts block removal probabilities and generates a 3D video of the reconstruction process, enhancing conservation efforts at sites like Uxmal, Yucatan.

The development of the Jenga Inverse Predictor (JIP-2) enables builders and project managers to assess and restore collapsed structures with greater accuracy and efficiency, potentially reducing costs and time in conservation projects. For investors, this technology represents a novel application of AI in heritage conservation, opening opportunities in both construction and preservation markets.

ComMem introduces a dual-memory system for test-time adaptation in vision-language models, outperforming existing methods on 15 benchmark datasets. By mimicking brain functions, it combines fast visual caching and slow textual refinement, achieving superior cross-modal consistency and adaptability under distribution shifts.

The development of ComMem, a dual-memory system for vision-language models, significantly enhances test-time adaptation capabilities, which is crucial for builders and PMs looking to create more robust AI applications. For investors, this advancement signals a potential leap in performance across various AI-driven products, increasing their market competitiveness and scalability.

The SoccerNet 2026 submission introduces a two-stage pipeline for player-centric ball action spotting, achieving a Macro-F1 score of 58.94, up from a baseline of 48.6. Key innovations include a Track-Aware Action Detector (TAAD) enhanced with a temporal transformer and a Denoising Sequence Transduction (DST) transformer employing a novel per-player attention mechanism. The ensemble approach effectively reduces false positives while maintaining recall.

The introduction of the Track-Aware Action Detector (TAAD) and Denoising Sequence Transduction (DST) transformer in SoccerNet 2026 significantly improves player-centric ball action spotting accuracy, as evidenced by a Macro-F1 score increase to 58.94. This advancement highlights the potential for enhanced analytics and real-time insights in sports tech, which can attract investment and drive product development in AI-driven sports applications.

COMPASS introduces a unified multimodal framework for composition-intent control, enhancing both perception and generation through a shared expert token. It significantly improves composition understanding and generation consistency, outperforming strong baselines on a newly constructed dataset, Comp-11, which features 11 classes and reasoning-augmented annotations.

The introduction of COMPASS, a unified multimodal framework for composition-intent control, represents a significant advancement in AI's ability to understand and generate content across different modalities. This development can enhance user experience in applications like content creation and interactive systems, making it a crucial consideration for builders and PMs looking to leverage multimodal AI capabilities.

Kuaishou is negotiating with General Atlantic for a $2 billion investment in its AI video generation unit, Kling AI, aiming for a post-money valuation of $18 billion. This move is part of Kuaishou's strategy to attract a prominent U.S. investor before its IPO.

Kuaishou's negotiation with General Atlantic for a $2 billion investment in its AI video generation unit, Kling AI, signals strong confidence in AI-driven content creation. For builders and PMs, this highlights the growing importance of AI in media, while investors should note the potential for high returns in AI startups as they prepare for IPOs.

DeLux introduces a cross-modal restoration method using neuromorphic event streams to effectively reduce lighting artifacts in RGB video, achieving an average MS-SSIM of over 0.99 and an 88% reduction in artifact severity in real-world footage. This approach outperforms existing RGB-only and event-guided HDR models, providing a significant advancement in video restoration techniques.

The introduction of DeLux's cross-modal restoration method using neuromorphic data significantly enhances video quality by reducing lighting artifacts, achieving an MS-SSIM of over 0.99. This advancement presents opportunities for builders and PMs in video technology and content creation, while investors may find potential in applications across various industries, including entertainment and surveillance.

Fine-tuning Gemini 2.5 Pro on 400 clinician-rated home videos improved ASD diagnosis accuracy by 53%, achieving 77% accuracy and an AUC of 86%. This approach enhances early diagnosis for 1 in 31 US children affected by autism.

The fine-tuning of Gemini 2.5 Pro for autism behavioral scoring demonstrates a significant 53% improvement in diagnosis accuracy, indicating the potential for AI to enhance early detection of autism in children. Builders and PMs should consider integrating such advanced models into healthcare applications, while investors may find opportunities in AI-driven solutions targeting mental health diagnostics.

CoIn introduces a multi-stage framework for 2D-3D inpainting, utilizing Gaussian Splatting for enhanced scene reconstruction. It achieves state-of-the-art performance in both object removal and insertion tasks, leveraging a diffusion model and adaptive feature attention for consistency across views.

The introduction of CoIn, a multi-stage framework for 2D-3D inpainting using Gaussian Splatting, significantly enhances scene reconstruction capabilities. This development is crucial for builders and PMs in industries like gaming and virtual reality, as it allows for more realistic object integration and manipulation, potentially driving innovation and investment opportunities in immersive technologies.

This study introduces a framework that enhances motion generation by integrating large-scale synthetic human motion with a redesigned VQ-VAE tokenizer, significantly improving the diversity and compositionality of learned motion vocabularies. The approach demonstrates consistent performance gains in tasks like text-to-motion and motion continuation, indicating that expanding the motion representation space is crucial for better generalization in human motion synthesis.

The introduction of a redesigned VQ-VAE tokenizer for motion generation, combined with large-scale synthetic human motion, significantly enhances the diversity of motion vocabularies. This development is crucial for builders and PMs in the gaming and animation sectors, as it opens up new possibilities for more realistic and varied character animations, which can lead to improved user engagement and satisfaction.

MemoBench introduces a new benchmark for evaluating memory consistency in video generation models under dynamic conditions, focusing on the disappear-and-reappear paradigm. It includes 360 ground-truth clips and assesses eight state-of-the-art models, revealing critical insights into memory challenges in changing environments.

The introduction of MemoBench provides a standardized way to evaluate video generation models in dynamic environments, which is crucial for builders and PMs focused on developing applications in areas like AR/VR and autonomous systems. For investors, understanding the performance of these models can inform funding decisions in emerging AI technologies that require robust memory handling.

ReWorld introduces a novel representation learning framework for World Action Models (WAMs) in autonomous driving, enhancing video generation performance by 23.9% in FVD and improving closed-loop PDMS from 89.1 to 90.4 without post-training methods. The framework optimizes intermediate representations directly, significantly accelerating convergence by approximately 2x on benchmarks like nuScenes and NAVSIM.

The introduction of ReWorld's representation learning framework for World Action Models (WAMs) significantly enhances video generation performance and accelerates convergence in autonomous driving applications. This development is crucial for builders and PMs as it improves the efficiency and effectiveness of training models, while investors should note its potential to advance autonomous vehicle technologies and reduce development time.

The Frame Forgetting Network (FFN) introduces a novel approach to Test Time Training (TTT) for long videos, optimizing computational efficiency by processing only three frames at a time. This method reduces unnecessary computations and adapts to new information effectively, demonstrating significant performance improvements on dense-segmentation and video classification tasks using a new dataset of up to 3-hour long videos.

The introduction of the Frame Forgetting Network (FFN) for Test Time Training (TTT) optimizes video processing by focusing on three frames at a time, which enhances computational efficiency and adaptability. This development is crucial for builders and PMs in video analytics and AI applications, as it enables more effective handling of long video content with reduced resource consumption.

The proposed self-supervised framework learns implicit 3D physics from video signals using a Volumetric Latent Space, achieving high structural stability and physical plausibility on benchmarks like CLEVERER and PhysInOne, without relying on traditional physics engines.

The development of Neural Voxel Dynamics introduces a self-supervised framework that learns 3D physics from video signals, which could significantly reduce reliance on traditional physics engines in game development and simulations. This innovation offers builders and PMs a more efficient way to create realistic environments, while investors may see potential for cost savings and enhanced product offerings in the gaming and simulation markets.

The Physics Question Scene Graph (PQSG) introduces a novel evaluation method for text-to-video generation, assessing physical plausibility through a hierarchical question framework. Validated with the FinePhyEval dataset, PQSG shows higher correlation with human judgments than previous methods and ranks closed-source models higher in physical realism than Wan 2.1.

The introduction of the Physics Question Scene Graph (PQSG) for evaluating text-to-video generation marks a significant advancement in assessing physical plausibility, which is crucial for developers aiming to create realistic AI-generated content. This method's higher correlation with human judgments suggests that builders and PMs can leverage it to enhance user engagement and realism in their products, while investors may see potential in supporting projects that utilize this robust evaluation framework.

MJEPA introduces a unified joint-embedding predictive architecture for audio-visual learning, outperforming prior models by over 6.8 mAP on AudioSet-20K. Utilizing a single predictive objective across modalities, it enhances representation synergy while using 10x less video data, demonstrating significant efficiency and effectiveness.

The introduction of MJEPA, a joint-embedding predictive architecture for audio-visual learning, significantly improves model performance while reducing data requirements by 90%. This efficiency allows builders and PMs to develop more effective multimedia applications with lower data costs, making it an attractive proposition for investors looking for scalable AI solutions.

The Chorus II framework introduces cross-request sparsity reuse for image-to-video generation, achieving a 2.16× speedup by leveraging shared sparse masks from historical requests, minimizing online mask prediction overhead. This method enhances efficiency while maintaining generation quality, addressing the computational challenges of diffusion models in large-scale deployments.

The introduction of the Chorus II framework for image-to-video generation, which achieves a 2.16× speedup through cross-request sparsity reuse, is significant for builders and PMs as it reduces computational costs and enhances scalability for large-scale deployments. For investors, this development signals a potential increase in efficiency and profitability in AI-driven content creation technologies.

FreeStory introduces a training-free framework for visual storytelling that enhances character consistency without structured prompts. By utilizing entity-grounded feature reuse, it outperforms existing methods on structured benchmarks and maintains stronger consistency in free-form prompts. The new benchmark, FreeStoryBench, supports both single and multi-character narratives.

The introduction of FreeStory, a training-free framework for visual storytelling, allows builders and PMs to create more consistent character narratives without the need for structured prompts, potentially reducing development time and costs. For investors, this innovation signals a shift towards more accessible AI tools in creative industries, enhancing the market potential for storytelling applications.

CoGeoAD introduces a unified CLIP-based framework for zero-shot 3D anomaly detection, effectively fusing 2D color images and 3D geometric structures. Its innovative Data-Driven Multi-View Attention mechanism and Multi-Stage Color-Geometric Fusion module achieve state-of-the-art performance on MVTec3D-AD and Eyecandies benchmarks, addressing critical industrial quality inspection challenges.

The development of CoGeoAD, a unified CLIP-based framework for zero-shot 3D anomaly detection, is significant for builders and PMs as it enhances industrial quality inspection processes by integrating 2D and 3D data. For investors, this technology presents opportunities in automation and AI-driven quality control, potentially reducing costs and improving product reliability in manufacturing.

The KidRisk dataset, comprising 2,500 videos and 10,000 images, enables improved recognition of children's dangerous actions, achieving 83.53% accuracy in action classification and 96.14% in danger recognition using vision-language models, outperforming traditional deep learning methods.

The release of the KidRisk dataset, which achieves high accuracy in recognizing dangerous actions among children, is significant for builders and PMs focused on safety applications and child monitoring technologies. Investors should note its potential to enhance AI-driven safety solutions, paving the way for innovative products in child protection and surveillance.

NaviGen enhances personalized multimodal content generation by transforming user interaction history into executable instructions, addressing the challenges of behavior encoding and instruction writing. The model improves image and video generation across various domains, yielding more relevant and visually generatable outputs.

NaviGen's ability to transform user interaction history into executable instructions for personalized multimodal content generation enhances the relevance and quality of AI-generated images and videos. This development signals a significant advancement for builders and PMs in creating more engaging user experiences, while investors should note its potential to capture market interest in personalized content solutions.

The Sol Video Inference Engine is a training-free acceleration framework for video diffusion models, achieving over 2x end-to-end acceleration with near-lossless VBench quality across models like Cosmos3-Super and LTX-2.3. By utilizing techniques such as cache, sparse attention, and token pruning, it optimizes performance with minimal human intervention.

The launch of the Sol Video Inference Engine, which offers over 2x acceleration for video diffusion models, is significant for builders and PMs as it reduces development time and costs while maintaining high quality. Investors should note this advancement as it enhances the competitive edge of products relying on video generation technologies, potentially leading to increased market demand.

This study introduces ingredient-level semantic segmentation using SegFormer-B0 and B1 on the FoodSeg103 dataset, achieving pixel accuracies of 0.7709 and 0.7929, respectively. The B1 model outperformed B0 with a mean IoU improvement of 0.0683, providing a visual summary of ingredient areas for enhanced nutrition awareness.

The introduction of ingredient-level semantic segmentation using SegFormer-B1, achieving a mean IoU improvement of 0.0683, presents a significant advancement for applications in nutrition tracking and food analysis. Builders and PMs can leverage this technology to create more accurate food recognition tools, while investors may find opportunities in health and wellness startups focused on personalized nutrition solutions.

HANCLIP introduces a new family of Vision-Language Models that enhance negation sensitivity by restructuring the embedding space. Trained on 20,000 image-text quadruplets, it shows improved performance on the NegBench benchmark while maintaining competitive results on standard tasks. This model-agnostic framework can be integrated into existing models like CLIP without extensive retraining.

The introduction of HANCLIP, a family of Vision-Language Models that enhances negation sensitivity, is significant for builders and PMs as it offers a model-agnostic framework that can be easily integrated into existing systems like CLIP, improving performance on specific tasks without extensive retraining. For investors, this development signals a potential for more robust AI applications in understanding complex language nuances in visual contexts.

TheProfessor introduces a multi-teacher approach for prompt distillation in vision-language models, enhancing performance on datasets like EuroSAT, achieving a +5.78 HM improvement. By leveraging a two-teacher ensemble, it outperforms single-teacher methods, demonstrating significant gains in domain-shifted scenarios.

The introduction of The Professor's multi-teacher approach for prompt distillation in vision-language models significantly enhances performance, as evidenced by a +5.78 HM improvement on datasets like EuroSAT. This development indicates that leveraging ensemble methods can lead to better model robustness and adaptability in real-world applications, which is crucial for builders and PMs looking to enhance AI product capabilities.