Realtime image generator, new GPT, fast AI videos, new vision models, robot MMA, AI music on phone

Key Concepts

• 3D Model Generation: Creating 3D objects from 2D reference images.
• Vision Language Model (VLM): AI that understands both images and text.
• Text-to-Audio Generation: Creating audio from text prompts.
• Audio-to-Audio Style Transfer: Applying the style of a reference audio clip to a new audio generation.
• Video Generation: Creating videos from text prompts or images.
• Realtime Image Generation: Generating images almost instantaneously.
• Light Manipulation: Adjusting lighting in images, including brightness, color, and position.
• AI Agents: Autonomous systems that can perform tasks independently.
• Deep Research Agents: AI systems that can conduct research, analyze data, and generate reports.
• GPT (Generative Pre-trained Transformer): A type of large language model.
• VRAM (Video RAM): Memory used by the graphics processing unit (GPU).
• Quantization: Reducing the precision of numerical data to decrease model size and VRAM requirements.
• Stems: Individual audio tracks that make up a complete song.
• DAW (Digital Audio Workstation): Software used for recording, editing, and producing audio.

3D Model Generation with Step 1X 3D

• Functionality: Step 1X 3D generates 3D models and textures from reference images.
• Accuracy: It is considered one of the best 3D model generators in terms of accurately following reference images, comparable to Hunyen 3D, Trellis, and Trippo.
• Features:
  • Symmetry Adjustment: A slider to control the symmetry of the generated object.
  • Sharpness/Smoothness Control: A slider to adjust the sharpness or smoothness of the object's geometry.
• Hugging Face Demo: A free online demo is available on Hugging Face.
  • Guidance Scale: Determines how literally the AI follows the image.
  • Inference Steps: Number of iterations the AI goes through (sweet spot around 50).
  • Max Face Number: Maximum number of faces for the object (default 400,000).
• Availability: Models are released on Hugging Face, and a GitHub repo provides instructions for local installation.

ByteDance's Seed 1.5 VL: A Powerful Vision Language Model

• Functionality: Seed 1.5 VL is an open-source vision language model that understands both images and text and performs complex visual reasoning.
• Capabilities:
  • Image Identification: Correctly identifies locations in images (e.g., Lombard Street).
  • Data Extraction: Extracts data from images like receipts into markdown tables.
  • Object Detection and Counting: Accurately counts and identifies objects in images (e.g., cats, strawberries, birds).
  • Visual Puzzle Solving: Solves visual puzzles by understanding image sequences.
  • AI Agent for Computer Navigation: Navigates computer windows and performs actions based on prompts.
• Performance: Achieves state-of-the-art performance on various visual reasoning tasks, even surpassing Google's Gemini 2.5 Pro and OpenAI's 01 on some benchmarks.
• Model Size: Relatively small, using a 532 million parameter vision encoder and a 20 billion parameter mixture of experts model.
• Availability: GitHub repo with instructions for download and usage under the Apache 2 license. A free Hugging Face space is also available for online use.

Stability AI's Stable Audio Open Small: Efficient Audio Generation

• Functionality: Stable Audio Open Small is a super-efficient audio generator that creates music and sound effects from text prompts.
• Efficiency: Generates 12 seconds of audio in 7 seconds on a mobile phone.
• Stereo Sound: Generates audio with stereo width and 3D effect.
• BPM Accuracy: Accurately follows specified BPM in prompts.
• Style Transfer: Transfers the style of a reference audio clip to a new generation.
• Stem Generation: Generates individual stems that complement existing tracks, allowing for layered compositions.
• Optimization: Optimized to run on ARM CPUs, making it suitable for smartphones.
• Technique: Uses adversarial relativistic contrastive postraining for faster and better audio generation.
• Availability: Models are released on Hugging Face, and a GitHub repo provides instructions for local installation.

Delete Me: Protecting Personal Data Online

• Problem: Data brokers sell personal information scraped from social media and other records, leading to risks like stalking and scams.
• Solution: Delete Me scans hundreds of data broker sites, finds personal information, and removes it.
• Functionality: Provides a dashboard showing the number of listings reviewed, data brokers with information, removals in progress, and listings cleaned.
• Sponsor: Delete Me is the sponsor of the video.
• Offer: 20% off with promo code AISEARCH.

LTX Video: Fast and Efficient Video Generation

• Functionality: LTX Video is an open-source video generator that balances quality and speed.
• Update: A 13B distilled model has been released with improved generation speed and low VRAM requirements (as low as 12 GB, or even lower with distilled FP8 or GGUF versions).
• Speed: Videos can be generated in 4-8 steps.
• Hugging Face Demo: A free Hugging Face space is available for online use.
• Installation: A previous video provides step-by-step instructions for local installation.

Tencent's Hunyan Image 2.0: Realtime Image Generation

• Functionality: Hunyan Image 2.0 is an image generator that creates high-resolution images in milliseconds.
• Realtime Canvas: Features a realtime canvas where users can sketch and see generated results immediately.
• Availability: Requires signing up with an email on the official website and joining a waitlist.

Google AI Updates: Alpha Evolve and Light Lab

• Alpha Evolve: An autonomous system that makes scientific breakthroughs. (Refer to a separate video for details).
• Light Lab: A tool that accurately changes the lighting in a picture, adjusting brightness, color, and adding new lights.
  • Features: Detects lights in the photo, adjusts brightness using sliders, creates ambient light, and adjusts the color of light.
  • Technology: Uses a segmentation map to detect lights, a depth map to estimate image depth, and a light-controlled diffusion model to generate the final image.
  • Availability: Currently only a technical paper is available.

Robot Fighting Tournament in China

• Event: A robot fighting tournament is taking place in Hongjo, China.
• Format: Four teams will remotely control Uni Tree robots in real time.

Vase by Alibaba: Reference-to-Video Generation

• Functionality: Vase is a free and open-source reference-to-video generator.
• Capabilities: Replaces characters in videos, adds multiple reference characters, and transfers motion from one video to another.
• Update: Official non-preview versions have been released, including a 14 billion parameter model (1280x720 resolution).
• Licensing: Models are under the Apache 2 license.
• VRAM Requirements: The official version requires 80 GB of VRAM, but a quantized version can run on as low as 8 GB.

Blip 30 by Salesforce: Open-Source Image Generator

• Functionality: Blip 30 is a family of multimodal models for image understanding and generation.
• Models: Two models are available: 4 billion and 8 billion parameters.
• Technology: Uses a combination of auto-regressive and diffusion models.
• Capabilities: Chatbot interaction, image analysis, image generation, and image editing.
• Availability: Models are released, and a GitHub repo provides instructions. A free online demo is also available.
• Limitations: Lacks quality and understanding of human anatomy compared to other image generators.

OpenAI Updates: GPT 4.1 and Codeex

• GPT 4.1: A specialized model in ChatGPT that excels at coding tasks and instruction following.
  • Availability: Available to Plus, Pro, and Team users, with Enterprise and Education users gaining access later.
  • Performance: Decent performance, better than GPT-4.0 but not as good as other leading models.
  • GPT 4.1 Mini: Replaces GPT-4.0 Mini in ChatGPT for all users.
• Codeex: A coding agent that autonomously performs coding tasks like writing code, fixing bugs, and running tests.
  • Powered by: Codeex 1, a specialized version of the 03 model.
  • Functionality: Identifies issues in codebases and suggests fixes.
  • Codeex CLI: An open-source command-line version that requires connection to OpenAI's API.

Deer Flow by ByteDance: Deep Research Agent

• Functionality: Deer Flow is a free and open-source system of AI agents that helps with research and other tasks autonomously.
• Capabilities: Searches for information, reads articles, analyzes data, writes reports, and generates slides or podcasts.
• Functionality: Creates research plans, gathers information, and generates reports.
• Availability: GitHub repo with instructions for local installation.
• Integrations: Supports various search engines (DuckDuckGo, Arcive) and AI models (Quinn, DeepSeek, Llama, OpenAI, Anthropic).

Conclusion

This week in AI saw significant advancements across various domains, including 3D modeling, vision language models, audio generation, video generation, and AI agents. Notable releases include Step 1X 3D for accurate 3D model generation, ByteDance's Seed 1.5 VL for powerful visual reasoning, Stability AI's Stable Audio Open Small for efficient audio generation, and Deer Flow for autonomous research. Additionally, updates from OpenAI with GPT 4.1 and Codeex, along with Tencent's Hunyan Image 2.0 for realtime image generation, highlight the rapid pace of innovation in the field. The increasing availability of free and open-source tools empowers developers and researchers to explore and contribute to the ongoing evolution of AI technology.