New DeepSeek & Gemini, Wan 2.5, Alibaba dominates AI, 3D scenes, new deepfake lipsync: AI NEWS

Key Concepts

• Video generation from 3D models
• Reference to video generation (face cloning)
• 3D model segmentation and reconstruction
• 3D scene generation from images/videos (4D)
• Language model upgrades (DeepSeek Terminus, Gemini 2.5 Flash)
• Object insertion into videos
• Humanoid robot balance and adaptation
• Multimodal AI (text, image, audio, video)
• Vision language models
• Image editing
• Code generation
• AI agents
• Personalized AI updates
• AI music generation
• Realistic lip-sync animation

Video Generation from 3D

• Video from 3D: A new method for creating videos by applying the texture and style of a reference image onto a moving 3D model.
  • Allows precise camera control and object movement.
  • Can add effects like changing the season or adding fire through prompts.
  • More consistent than other similar tools, with fewer hallucinations.
  • Process:
    1. Takes a reference image and the edge/canny map of the video.
    2. Uses a SAG (Sparse Anchor View Generation) module to generate a few keyframes (anchor views).
    3. Uses a GGI (Geometry Guided Generative Inbetweening) module to interpolate the frames between the anchor views, creating a smooth video.
  • GitHub repository available for local installation.

Reference to Video Generation (Links by ByteDance)

• Links: An open-source video generator that clones a reference character's face onto a video, allowing the character to perform actions specified by a prompt.
  • Higher quality than previous competitors (Phantom, Standin) in terms of face accuracy, video quality, and prompt following.
  • Can control subtle expressions and complex actions.
  • Released under the Apache 2 license, allowing commercial use.
  • GitHub repository available for local installation.

3D Model Segmentation and Reconstruction (Hunen 3D Part by Tencent Hunen)

• Hunen 3D Part: Consists of two tools:
  • P3 SAM: Segments a 3D model into meaningful parts by converting it into a point cloud and using AI.
  • XART: Generates individual shapes for the segmented parts, completing incomplete parts and allowing for further editing.
• Online demo available on Hugging Face.
• GitHub repositories available for local installation of both P3 SAM and XART.

3D Scene Generation from Images/Videos (LRA by Nvidia)

• LRA: Generates an entire 3D scene from an input image or video.
  • Can handle different characters and scenes.
  • Can generate 3D videos (4D, with the added dimension of time).
  • Potentially useful for training autonomous driving systems.
  • Requires significant VRAM (80GB recommended, minimum 43GB with memory offloading).
  • GitHub repository available for local installation.

Language Model Upgrades

• DeepSeek V3.1 Terminus: An upgraded language model with improved language consistency and agentic performance.
  • Shows improvements in benchmarks like GPQA Diamond, Humanity's Last Exam, and software engineering tasks (SU Verified, Bench Multilingual, Terminal Bench).
  • Tied with GBT OSS as a leading open-source model according to the Artificial Analysis leaderboard.
  • Available on Hugging Face.
• Google Gemini 2.5 Flash and Flash Light: Updated versions with better instruction following, stronger multimodal and translation capabilities, and improved efficiency.
  • Faster response times and higher intelligence index compared to previous versions.
  • Available on Google's AI Studio platform for free testing.

Object Insertion into Videos (Omniinsert by ByteDance)

• Omniinsert: Inserts characters or objects into existing videos using an image and/or a prompt.
  • More accurate and consistent than previous tools like PA.
  • GitHub repository available, with plans to release inference codes.

Humanoid Robot Balance and Adaptation

• Unitree G1 Robot: Demonstrated impressive balance and recovery capabilities.
• Skilled AI: A generalist AI system that can be inserted into any robot body and adapt to numerous tasks without previous training.
  • Trained on a thousand years' worth of simulated walking across 100,000 different robotic bodies.
  • Can adapt to changes like leg removal, broken legs, or stilts.

Alibaba's AI Releases

• One 2.5: A video generator with natively built-in audio.
  • Supports text-to-video and image-to-video.
  • Can generate videos up to 1080p resolution and 10 seconds in length.
  • Available for free trial on Alibaba's One platform.
• Quenfree Max: A language model that competes with Claude Opus 4 and DeepSeek.
  • Thinking model achieves 100% scores in mathematical reasoning benchmarks.
  • Available for free use in Quen Chat and via API.
• Quen 3 Omni: An end-to-end multimodal model that can handle text, image, audio, and video input and output.
  • Supports text interaction in 119 languages, speech understanding in 19 languages, and speech generation in 10 languages.
  • Latency as low as a few hundred milliseconds.
  • Can analyze audio up to 30 minutes long.
  • Outperforms some proprietary models like GPT and Gemini in benchmarks.
  • Available for free use in Quen Chat.
  • 30B parameter models released on Hugging Face, with a GitHub repository for local installation.
• Quen 3VL: A vision language model with visual reasoning capabilities.
  • Can analyze images, identify brands, and search the web for information.
  • Can generate code from sketches and plot data from images.
  • Outperforms Gemini 2.5 Pro thinking, GPT5, and Claude Opus 4.1 in benchmarks.
  • Models released on Hugging Face, with a GitHub repository for local installation.
• Quen ImageEdit: An image editor that competes with Nano Banana.
  • Free and open source.
  • Can generate photos of people in specific poses using pose skeletons.
• Quen 3 Coder: A code generation model.
  • Significantly better than the previous version.
  • Comparable to Claude for Sonnet in Sweetbench verified performance.
  • Available for free use in Quen Chat.
  • GitHub repository available for local installation.

AI Agents

• Kimmy's "Okay, Computer": An agent mode for the Kimmy AI model.
  • Can autonomously execute tasks using multiple steps.
  • Can create websites, slideshows, and interactive dashboards.
  • Available on kimmy.com with three free sessions.

Personalized AI Updates

• OpenAI's Chat GPT Pulse: A new feature that provides personalized updates based on user data.
  • Can synthesize information and provide daily updates on topics and goals.
  • Can connect to Gmail and Google Calendar for additional context.
  • Currently only available to pro users on mobile (200 USD/month).

AI Music Generation

• Sunno V5: The latest version of the Sunno AI music generator.
  • More structurally coherent and creative.
  • Voices and instruments remain consistent across the entire track.
  • Faster generation times.
  • Available to pro and premier users.

Realistic Lip-Sync Animation

• Omnihuman 1.5: A realistic deep fake animator that animates and lip-syncs images with audio.
  • Understands the context of the audio and animates the person accordingly.
  • Available on foul.ai for a fee (16 cents per second).

Conclusion

The AI landscape is rapidly evolving, with significant advancements in video generation, 3D modeling, language models, multimodal AI, and robotics. Alibaba has emerged as a major player, releasing a suite of powerful and open-source AI tools. The trend towards more capable and accessible AI models continues, empowering users to create and innovate in various domains.