10 / Mar 3 – 9, 2025 · 9 RELEASES

AI News Week 10, 2025

Week 10 2025 AI releases.

Vision & Image (2)

Thera
ETH Zurich & University of Zurich · Image Restoration
Arbitrary-scale super-resolution method using neural heat fields with built-in physical observation model that guarantees alias-free output through principled anti-aliasing by design. Achieves state-of-the-art performance via pixel-wise local neural fields parameterized by a hypernetwork with frequency-dependent thermal activations.
Bokeh Diffusion
Armando Fortes, Tianyi Wei, Shangchen Zhou, Xingang Pan · Text-to-Image
A framework that enables precise defocus blur (bokeh) control in text-to-image diffusion models by conditioning on physical camera parameters, using a hybrid training pipeline with synthetic blur augmentations and a grounded self-attention mechanism to preserve scene content while adjusting blur strength.

Stable Virtual Camera
Stability AI · Text-to-Video
Multi-view video generation model with 3D camera control that enables consistent video synthesis from multiple viewpoints with precise camera trajectory manipulation.
Nvidia Cosmos
Nvidia · Text-to-Video
Nvidia Cosmos is a family of world foundation models designed to generate realistic and controllable videos and 3D content from text, images, or video prompts. The models enable simulation and content creation across various domains including robotics, autonomous vehicles, and digital twins.

SpatialLM
Manycore Tech Inc. & Hong Kong University of Science and Technology · Point Cloud
A 3D large language model that processes point clouds from diverse sources (monocular video, RGBD, LiDAR) to generate structured indoor scene understanding including walls, doors, windows, and semantic object bounding boxes. Outputs are compatible across multiple formats including 3D layouts, 2D floorplans, and industry-standard IFC.
LHM - Large Animatable Human Reconstruction Model
Alibaba Group (Tongyi Lab) · 3D Reconstruction
LHM reconstructs high-fidelity, animatable 3D human avatars from single images in seconds using 3D Gaussian splatting and a multimodal transformer architecture. The model preserves detailed clothing geometry, facial identity, and hand details without post-processing through a novel head feature pyramid encoding scheme.
StdGEN
Tencent AI Lab & Tsinghua University · 3D Generation
A 3D character generation pipeline that creates semantically decomposed characters (body, clothes, hair as separate components) from single images in three minutes using a Semantic-aware Large Reconstruction Model (S-LRM) with multi-view diffusion and hybrid implicit field reconstruction.

RD-Agent
Microsoft · Agent Framework
An agent framework for research and development tasks that enables autonomous planning and execution of complex workflows through tool integration and reasoning capabilities.
Computer Use Browser (CUA)
Browserbase and OpenAI · Web Agents
A web interface demonstrating OpenAI's computer-use model preview, enabling AI agents to autonomously browse the web, interact with web applications, and complete tasks like reviewing pull requests, playing games, and analyzing data through visual input and mouse/keyboard control.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.