Video Generation

Today's Stories

1. 1

   Huawei builds 博観, local-history AI for China tourism

   Huawei (ファーウェイ) developed an AI model called 博観 that specializes in local history. It learns from museum collections and generates PR videos and mini-drama scripts for tourist spots. The model ties regional content to tourism demand, using Chinese motifs such as 昭陵六駿 — the beloved horses of the Tang dynasty emperor who unified China — to drive visitor traffic.

   Success hinges on whether generated content actually brings travelers to destinations. The article gives no launch date, pricing, or client list to watch yet.

2. 2

   Apple's CapQuiz grades video captions by quiz answers

   Apple researchers introduced CapQuiz, a benchmark that scores video captions by how well they answer human-verified multiple-choice questions from the video, covering 10 question types across 24 video domains, plus a composite metric, CapF1. Existing metrics match text against reference captions, which penalizes valid descriptions that differ in wording or visual focus. CapQuiz correlates significantly better with human judgments, and CapF1 combines factuality (CapP) and coverage (CapR) rather than collapsing them into one score.

   The test is whether reference-free, question-based scoring becomes a trusted way to compare video models, since it trades text matching for verifiable answers. Watch whether CapQuiz and CapF1 are adopted beyond this research setting.

3. 3

   TwelveLabs Marengo Embed 3.0 lands in Amazon Bedrock Knowledge Bases

   AWS announced the general availability of TwelveLabs Marengo Embed 3.0 as an embedding model in Amazon Bedrock Knowledge Bases, a fully managed RAG service that ingests video (MP4, MOV), images (JPEG, PNG), and audio. Previously, semantic search over video meant stitching together transcription, frame extraction, embedding models, vector databases, and synchronization logic; AWS says Managed MKB now handles segmentation, frame sampling, and transcription internally.

   Availability is limited to the US East (N. Virginia) and US West (N. California) Regions, so the test is whether AWS expands to more Regions. Embeddings generation with Marengo Embed 3.0 is charged at the standard Amazon Bedrock model invocation rate.

4. 4

   Amazon's Prime Video launches AI lip-sync dubbing on Maxton Hall

   Amazon's Prime Video is launching an AI-powered feature that aligns actors' lips with dubbed audio, starting with the English dub of the German series "Maxton Hall." This feature, currently available for seasons 1 and 2 of "Maxton Hall," will also be available for its third season on December 9th, expanding the reach of dubbed content.

   Prime Video plans to expand this lip-sync dubbing to additional titles in the future. The success of this feature may hinge on viewer acceptance of AI-altered visuals.

5. 5

   Adobe's Premiere Overhaul: Generate Media Directly on Timeline

   Adobe introduced Generative Media in Premiere, letting editors highlight an empty gap and generate context-aware, fully editable clips without leaving the timeline. Editors previously had to jump between browsers or external apps to find missing B-roll or audio; now multiple AI models like Adobe Firefly and Google Veo are accessible in the timeline.

   The tool's effectiveness hinges on how well it integrates with editors' workflows. New AI audio tools and an AI Assistant for After Effects are in beta.

6. 6

   World Labs unveils Atlas, a single AI model that generates, reconstructs, and simulates 3D worlds from just a few photos

   World Labs unveiled Atlas, an omni-model trained on text, images, video, and 3D data that anchors every input to a specific position in 3D space. It generates up to one minute of 1440p video with user-controlled camera paths, reconstructs real scenes from as few as two or three photos, and can handle over a hundred inputs. This addresses a problem Fei-Fei Li outlined in a November 2025 essay: current multimodal and video models break data into 1D or 2D sequences, making simple spatial tasks needlessly hard. Atlas's shared spatial understanding separates it from pure language or video models, and it outperforms specialized models in human evaluations, scoring 75% against MiniMax H3 and 94% against Seedance 2.5.

   Atlas's native 3D output supports point clouds and 3D Gaussian splats, matching the representation in Marble. It also serves as a real-to-sim tool for robotics, building on an August 2026 real-to-sim-to-real engine that ran for an hour each on five robot platforms without human intervention. Atlas is currently in an early-access program for select partners.