Superpower Daily
Posts
NVIDIA Reveals Mind-Blowing Real-Time AI: NPCs Respond to Human Speech in Video Games!

NVIDIA Reveals Mind-Blowing Real-Time AI: NPCs Respond to Human Speech in Video Games!

Extracting Multiple Concepts from a Single Image. Introducing Break-A-Scene

Saeed Ezzati
May 29, 2023

In today’s email:

🕹️ NVIDIA demos NPCs Respond to Human Speech in Video Games!.
🗂️ Organize your files using natural language and GPT-4.
🎇 Extracting Multiple Concepts from a Single Image
🛠 Various AI-related tools and platforms, including Decktopus, Quivr, roop, Plask, Swell, Autocode, Muzify, and more.

Highlights💡

NVIDIA CEO Unveils Gen AI Platforms for Every Industry [Link]

NVIDIA founder and CEO Jensen Huang delivered his first live keynote since the pandemic this Sunday at the COMPUTEX conference in Taipei. Watch the full video here.

He announced platforms for leveraging generative AI that is transforming industries such as advertising, manufacturing, and telecom.
The announcements included the DGX GH200, a large-memory AI supercomputer combining the Grace CPU and H100 Tensor Core GPU.
The NVIDIA MGX modular reference architecture for accelerated servers was introduced.
Collaborations with SoftBank and Microsoft were announced to build 5G and generative AI data centers and drive innovation for Windows PCs in the generative AI era, respectively.
Partnerships with WPP and electronics manufacturers were established for generative AI-enabled content creation and digital smart factories.
Huang highlighted the widespread adoption of NVIDIA technologies, with thousands of companies and startups using them and millions of downloads of CUDA software.

Break-A-Scene: Extracting Multiple Concepts from a Single Image [Project], [Paper], [Video]

The paper introduces "Break-A-Scene," a method for textual scene decomposition.
The method aims to extract distinct text tokens for each concept in a single image containing multiple concepts.
It enables fine-grained control over generating scenes by synthesizing individual concepts or combinations of them in different contexts.
The customization process involves a two-phase approach of optimizing textual embeddings (handles) and model weights.
A masked diffusion loss is used to ensure each handle can generate its designated concept without penalizing entanglement.
A loss on cross-attention maps is employed to prevent handle entanglement by ensuring each handle attends only to areas related to its target concept.
Union-sampling, a training strategy, is used to improve the generation of concept combinations.
The method is evaluated using automatic metrics and a user study, showcasing various applications.

GPTFile - Proof-of-concept illustration to show how LLM's could manage and organize files. [GitHub]

here’s gptfile, a way to organize files with natural language using gpt-4.
new operating system paradigms are on the horizon
repo below
— Aaron Ng (@localghost)
8:02 PM • May 29, 2023

Tools & Links 🛠️

roop - one-click deepfake (face swap) [GitHub], [More Demos]
Take a video and replace the face in it with a face of your choice. You only need one image of the desired face. No dataset, no training.

.local domains - Develop your apps locally with .local domains [Link]

Let's create some surreal influencer selfies using midjourney [Link]

Decktopus creates presentations that are both visually stunning and incredibly professional. [Link]
Quivr - Dump all your files and thoughts into your GenerativeAI Second Brain and chat with it [GitHub], [Demo], [App]

Swell AI - Build chat bots for your podcast episodes [Link]
Plask AI - You frame it, AI creates the rest.[Link]

Descript - A new way to make video and podcasts. A good way! [Link]

AutoCode - Your Ideas, Transformed into Code. Starting with “Text to Chrome Extension” [Link]
Turn yourself into an animated Pixar character with voice in 3 easy steps [Link]
Podcast Embeddings - 1000s hours of audio transcripts across the best podcasts turned into over 1M embeddings. And now you can chat with it [Link]
Getting started with LLMs? These 2 cheat sheets are a great starting point. [Link]
Chainlit lets you create ChatGPT-like UIs on top of any Python code in minutes [GitHub], [Docs]

Using the new Generative Fill feature from Photoshop to expand famous album covers [Link]

Hacking around ChatGPT’s Character Limits with the Code Interpreter [Link]
Muzify - Turn your book into an AI generated music playlist. [Link]
Automatic1111 - Stable Diffusion web UI [GitHub], [Tutorial]
Amazon Is Being Flooded with Books Entirely Written by Ai [Link]
Profile: Before His Killing, Tech Executive Bob Lee Led an Underground Life of Sex and Drugs [Link]