Open-Source AI 3D, Video & Audio Models: Tencent, Vidu, Audio2Face and More

1. Tencent releases two open‑source 3D models: Hunyuan 3D‑Omni and Hunyuan 3D‑Part

Tencent recently open‑sourced Hunyuan 3D‑Omni and Hunyuan 3D‑Part, accelerating AI‑driven 3D modeling for games, 3D printing, AR/VR and related fields.

This is the first fully open‑source industrial‑grade 3D generation model that also supports PBR texture material generation.

Hunyuan 3D‑Omni breaks the limitation of image‑only input by integrating skeletons, point clouds and other conditions, allowing precise control of geometry and pose. Hunyuan 3D‑Part can automatically split and generate over 50 component types, making 3D model manipulation as flexible as building blocks.

Compared with the previous 2.0 version, 3D‑Omni 2.1 focuses on dual optimization of geometry and texture, achieving state‑of‑the‑art results on open‑source 3D models, especially in metallic texture fidelity.

The models provide full‑chain open‑source weights, training code and data pipelines, and can run on consumer‑grade GPUs with detailed deployment guides.

2. Shengshu Technology launches Vidu Q2 for AI‑driven facial video generation

Vidu Q2 is a next‑generation video generation model that captures subtle facial expressions with high precision, integrates push‑pull camera motion techniques, and improves generation speed and semantic understanding.

It supports video lengths from 2 to 8 seconds and multiple creative styles, enabling creators to produce expressive digital characters, complex acting scenes, and dynamic camera movements with fine‑grained control.

3. NVIDIA open‑sources Audio2Face for real‑time facial animation

Audio2Face analyzes audio input to drive realistic facial motions of virtual characters in real time, synchronizing lip movements and natural expressions.

The open‑source release includes core algorithms, SDK, plugins for major game engines, and a complete training framework, simplifying the workflow for developers and enhancing immersion in games and 3D applications.

4. Figma upgrades MCP server to generate code directly from design files

Figma’s new MCP server removes client‑side dependencies, allowing AI‑assisted code generation agents to read design data directly and convert Figma prototypes into front‑end code assets, greatly speeding up the design‑to‑product pipeline.

5. Google unveils Mixboard, an AI‑powered creative board tool

Mixboard helps users quickly build mood boards by generating visual concepts from text prompts or uploaded images, offering one‑click regeneration and image editing to accelerate creative design for interior décor, event planning and more.

6. Alibaba upgrades Qwen‑Image with multi‑image editing capabilities

Qwen‑Image‑Edit‑2509 adds support for editing multiple images simultaneously (e.g., person + person, person + product) and improves single‑image consistency for faces, products and text, while natively supporting ControlNet inputs such as depth, edge and keypoint maps.

7. Kuaishou Keling 2.5 Turbo model released for stronger, cheaper video generation

Keling 2.5 Turbo enhances text understanding, fine‑grained creative control, dynamic effects, and aesthetic style, reducing the cost of generating a 5‑second 1080p video by nearly 30% compared with the previous model.