34 / Aug 18 – 24, 2025 · 10 RELEASES

AI News Week 34, 2025

Week 34 2025 AI releases.

Language Models (1)

DeepSeek-R1-0528
DeepSeek · Reasoning
Updated reasoning model with improved benchmark performance, reduced hallucinations, and support for JSON output function calling. Maintains backward compatibility with existing API integrations.

Chain-of-Zoom (CoZ)
KAIST AI / Bryan Sangwoo Kim et al. · Image Editing
CoZ is a model-agnostic framework that enables extreme super-resolution (16x-256x magnification) by decomposing upscaling into an autoregressive chain of intermediate scales with VLM-generated multi-scale-aware prompts, fine-tuned using GRPO for human preference alignment. A standard 4x diffusion SR model wrapped in CoZ achieves beyond 256x enlargement without additional training.

Chatterbox
Resemble AI · Voice Agents
Open-source conversational AI framework leveraging text-to-speech and voice synthesis capabilities for building interactive voice-based applications.

Direct3D-S2
Nanjing University, DreamTech, Fudan University, University of Oxford · 3D Generation
A scalable 3D generation framework that generates high-resolution 3D shapes from images using sparse volumetric representations and a novel Spatial Sparse Attention mechanism, achieving 3.9x-9.6x speedups and enabling 1024³ resolution training on 8 GPUs.
EVA: Expressive Virtual Avatars
Max Planck Institute for Informatics · 3D Generation
EVA generates high-fidelity, real-time photorealistic human avatars from multi-view videos using a two-layer architecture combining expressive template geometry with a decoupled 3D Gaussian appearance model. The system enables independent control of facial expressions, body movements, and hand gestures through disentangled skeletal and expression parameters.

Paper2Poster
University of Waterloo, University of Oxford, Vector Institute · Agent Framework
A multi-agent AI system that automatically generates academic posters from scientific papers using a Parser-Planner-Painter pipeline, paired with PaperQuiz—a VLM-based evaluation metric measuring poster effectiveness at conveying paper content. Outperforms GPT-4o while using 87% fewer tokens.
Deer-Flow
ByteDance · Agent Framework
A workflow orchestration framework for multi-agent systems that enables complex task decomposition and parallel execution across distributed agents.
Firecrawl Deep Job Researcher
Mendable AI · Web Agents
An AI agent application that uses web scraping and LLM capabilities to research job opportunities by crawling and analyzing job listings across the web. Built on Firecrawl's web crawling infrastructure for automated job market research.

Paper2Poster
GitHub Community · AI Infrastructure
Tool that automatically converts academic papers into visual poster presentations, likely using layout algorithms and text extraction to generate presentation-ready graphics from PDF documents.

OmniConsistency
ShowLab · Research Paper
A framework for generating consistent multi-modal content across images, videos, and 3D scenes while maintaining temporal and spatial coherence. Enables unified consistency constraints across different content modalities and generation tasks.

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