29 / Jul 14 – 20, 2025 · 11 RELEASES

AI News Week 29, 2025

Week 29 2025 AI releases.

Language Models (1)

Intern-S1
InternLM · Text Generation
Unable to analyze - page content unavailable. Based on URL path, appears to be a model release from InternLM organization on Hugging Face.

DINOv3
Meta AI Research · Visual Understanding
A self-supervised vision transformer model that learns dense visual representations without labels, enabling strong performance on downstream tasks like detection, segmentation, and retrieval through knowledge distillation from a teacher network.
AI Photo Editing in Google Photos
Google · Image Editing
Google Photos introduces AI-powered photo editing capabilities that automatically enhance images with features like magic eraser, face unblur, and real tone adjustments. The system uses machine learning models to intelligently remove unwanted objects and improve photo quality directly within the app.
Qwen-Image-Edit
Alibaba Qwen · Image Editing
An image editing model from the Qwen family that enables text-guided image manipulation and editing capabilities. Leverages Qwen's multimodal architecture to understand both images and natural language instructions for precise edits.

InfiniteTalk
Meituan / University of Chinese Academy of Sciences · Text-to-Video
Audio-driven video generation system for sparse-frame video dubbing that synthesizes full-body motion synchronized to audio while preserving identity and camera movements. Uses streaming architecture with context frames and adaptive reference frame conditioning to enable infinite-length video generation beyond traditional mouth-region editing.

TINKER
Zhejiang University · 3D Reconstruction
A generalizable 3D editing framework that achieves multi-view consistent edits from one or few input images without per-scene finetuning by leveraging pretrained diffusion models' latent 3D awareness. Includes a referring multi-view editor and any-view-to-video synthesizer for novel-view generation and scene completion.
MeshCoder
Shanghai AI Lab, Tsinghua University, University of Hong Kong, Harbin Institute of Technology, Beijing Institute of Technology, HKUST(GZ) · 3D Reconstruction
MeshCoder is an LLM-based framework that reconstructs 3D point clouds into editable Blender Python scripts using expressive APIs and a large-scale paired object-code dataset. The approach enables intuitive shape editing through code modification and enhances LLM reasoning for 3D shape understanding tasks.
GeoSAM2
VAST, University of Hong Kong, Tsinghua University, Sun Yat-sen University · 3D Reconstruction
GeoSAM2 extends SAM2 for 3D part segmentation by rendering multi-view normal and point maps, accepting 2D prompts (clicks/boxes) to guide interactive part selection, then back-projecting and aggregating masks across views with k-NN voting. Achieves state-of-the-art class-agnostic performance on PartObjaverse-Tiny and PartNetE without requiring text prompts or per-shape optimization.

Airi
moeru-ai · Agent Framework
Unable to analyze - page content unavailable. GitHub repository for Airi project by moeru-ai, likely an AI agent or framework based on naming conventions.

Lyzr Crawl
Lyzr · Web Scraping
Web crawling framework designed for AI applications, enabling efficient data extraction and processing from websites for LLM integration and agent automation.

RynnEC
Alibaba DAMO Academy · Research Paper
RynnEC appears to be a recurrent neural network-based framework from Alibaba's research division, likely focused on efficient computation or edge computing based on the repository structure and naming convention.

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