Quanta AI Labs

06 / Feb 3 – 9, 2025 · 8 RELEASES

AI News Week 6, 2025

Week 6 2025 AI releases.

Video & Animation (2)

  • RIFLEx

    Tsinghua University, ShengShu, UT-Austin, RUC, Pazhou Lab · Video Editing

    RIFLEx enables training-free 2× temporal length extrapolation in video diffusion transformers by reducing intrinsic frequency components in positional encodings to prevent temporal repetition and slow motion artifacts. The method requires only a single line of code modification and achieves high-quality video extension from 129 to 261 frames without additional training.

  • Hunyuan Video I2V

    Tencent · Text-to-Video

    Image-to-video generation model that synthesizes dynamic video content from static images. Enables temporal coherence and motion generation for converting photographs into short video clips.

Audio & Speech (2)

  • Sesame Voice - Uncanny Valley Analysis

    Sesame · Text-to-Speech

    Research on crossing the uncanny valley in voice synthesis, likely demonstrating improvements in naturalness and human-likeness of synthesized speech through acoustic modeling or neural vocoding techniques.

  • Spark-TTS

    Hong Kong University of Science and Technology, SparkAudio, Shanghai Mobvoi, Shanghai Jiao Tong University, Northwestern Polytechnical University, NetEase · Text-to-Speech

    Spark-TTS is an LLM-based text-to-speech model using BiCodec, a single-stream speech codec that decouples speech into semantic tokens and speaker-specific global tokens, enabling zero-shot voice cloning and fine-grained control over linguistic content and speaker attributes via chain-of-thought generation with Qwen2.5.

Agents & Automation (1)

  • TheoremExplainAgent

    University of Waterloo / Tiger AI Lab · Agent Framework

    An agentic system that generates long-form theorem explanation videos (5+ minutes) using Manim animations through a two-agent architecture (planner + coding agent). Introduces TheoremExplainBench with 240 theorems and 5 evaluation metrics; o3-mini achieves 93.8% success rate, revealing that multimodal explanations expose reasoning flaws invisible to text-based evaluation.

Developer Tools (1)

  • ComfyUI-HunyuanVideoWrapper

    kijai · AI Infrastructure

    A ComfyUI wrapper/integration for Hunyuan Video model, enabling video generation within the ComfyUI node-based interface. Provides streamlined access to Hunyuan's video synthesis capabilities through ComfyUI's modular workflow system.

Research & Safety (2)

  • SynCD: Multi-Image Synthetic Data for Text-to-Image Customization

    CMU & Meta · Research Paper

    SynCD is a data generation pipeline that creates multi-image synthetic datasets of objects in different contexts to train encoder-based models for customized text-to-image generation. The approach uses 3D Objaverse assets, masked shared attention mechanisms, and enforces consistent object identity across generated images, enabling high-quality customization from single or multiple reference images.

  • DICEPTION

    Zhejiang University, Shanghai AI Laboratory · Research Paper

    A generalist diffusion model that solves multiple vision perception tasks (depth estimation, segmentation, pose estimation, normal estimation) with a single model, achieving competitive performance with SAM using only 0.06% of the training data and enabling adaptation with minimal parameters.

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