# NVIDIA’s New AI Just Leveled Up Video Editing

Source: https://www.youtube.com/watch?v=RaNay3x0Fmk
Recap page: https://rapidrecap.app/video/RaNay3x0Fmk
Generated: 2026-02-06T14:36:52.548+00:00

---
## Quick Overview

The video showcases the capabilities of Lambda Labs' new AI services, particularly highlighting the performance of their GPU instances (like the 671B model) for running large language models quickly and reliably, contrasting it with older, slower, or less stable methods like previous papers on video inpainting and demonstrating superior metrics across various benchmarks.

**Key Points:**
- The presentation features demonstrations of object removal from video using new AI techniques, showing superior results compared to older methods like DiffuEraser (00:00-00:10, 00:24-00:40).
- The novel technique, OmniMatteZero, excels at removing foreground objects (like a dog or a person) and maintaining consistent shadows across frames (00:05-00:09, 02:08-02:27).
- OmniMatteZero achieves state-of-the-art performance, sweeping previous techniques across Movie, Kubric, and Average metrics according to the benchmark table (21:20-21:30, 23:24-23:45).
- The presentation transitions to introducing Lambda's 'Superintelligence Cloud' offering powerful NVIDIA GPU instances for training and inference, emphasizing pay-by-the-minute pricing (08:43-09:07).
- The speed of running a large model (Deepseek-R1:671B) on their infrastructure is extremely fast, processing 671 billion parameters per second, making LLM experimentation highly efficient (08:43-08:54).
- The video concludes by referencing the availability of the source code and encouraging viewers to check out lambda.ai/papers for more information (08:58-09:13).

**Context:** This video serves as a promotional and technical overview for new AI video editing and foundation model infrastructure advancements, primarily from Lambda Labs. It compares a new, highly effective video object removal technique called OmniMatteZero against older state-of-the-art methods, demonstrating its ability to handle complex occlusions and shadows consistently across video frames. The latter half of the video pivots to promoting Lambda's cloud computing services, emphasizing the raw speed and reliability of their GPU offerings for running massive models like the 671B parameter LLM.

## Detailed Analysis

The video begins by demonstrating the superior video object removal capabilities of a new technique called OmniMatteZero, comparing it side-by-side with previous methods like DiffuEraser (00:00-00:40). OmniMatteZero successfully removes foreground elements, such as a dog on a beach (00:00-00:07), people in a trampoline park (00:03-00:05), and a figure walking outdoors (02:17-02:22), while also correctly handling complex secondary effects like shadows, which older methods failed to remove or rendered poorly (00:38-00:40, 02:08-02:17). The technique is shown to maintain temporal coherence, ensuring that removed elements do not flicker back in subsequent frames (04:58-05:12). The presentation then shifts to technical validation, displaying a comparison table (21:20) where OmniMatteZero (LTXVideo and Wan2.1 versions) consistently outperforms all listed competitors across PSNR, LPIPS, and SSIM metrics on Movie and Kubric datasets (23:24-23:45). Following this, the video transitions into promoting Lambda Labs' infrastructure, featuring 'The Superintelligence Cloud' (08:58-09:04) designed for training and inference using 1 to 8 NVIDIA GPU instances. A key performance metric highlighted is the actual speed of running a 671B parameter model, achieving 671 billion operations per second, which is described as super fast and reliable (08:43-08:54). The presenter expresses excitement over this achievement, contrasting it with previous research papers that required complex workarounds or suffered from instability (08:56-09:05). The video concludes by directing viewers to lambda.ai/papers for the source code and further details (09:09-09:13).

### Video Object Removal Comparison

- OmniMatteZero removes foreground subjects (dog, person, swan) perfectly, including shadows, unlike previous techniques like DiffuEraser which left artifacts (00:00-00:40, 02:08-02:33, 04:57-05:03).

### OmniMatteZero Technical Superiority

- Benchmark table shows OmniMatteZero achieving the highest PSNR, LPIPS, and SSIM scores across Movie and Kubric datasets compared to competitors (21:20-21:30).

### Infrastructure Promotion

- Lambda Labs promotes 'The Superintelligence Cloud' offering on-demand NVIDIA GPU instances for training and serving models, emphasizing pay-by-the-minute pricing (09:02-09:07).

### LLM Performance Benchmark

- Running the Deepseek-R1:671B model achieves an actual speed of 671 billion parameters per second, demonstrating high-speed, reliable execution (08:43-08:54).

### Transformer Explanation (Emoji Edition)

- A brief section uses emojis to explain the transformer mechanism: Input (text to tokens, positional encoding) -> Self-Attention (words focus on each other) -> Layers Stacked (deep processing, no sequence dependency) (08:50-08:57).

### Call to Action

- Viewers are directed to lambda.ai/papers, with the source code promised to be available to reward those who follow the work (08:58-09:13).

![Screenshot at 00:01: Side-by-side comparison showing OmniMatteZero removing a blue, semi-transparent dog from a beach scene, with the 'After' panel showing a completely clean background.](https://ss.rapidrecap.app/screens/RaNay3x0Fmk/00-00-01.jpg)
![Screenshot at 00:34: Comparison showing OmniMatteZero successfully removing multiple blue-masked puppies from a grassy field, leaving behind natural grass and correctly filling in the shadows cast by the removed subjects.](https://ss.rapidrecap.app/screens/RaNay3x0Fmk/00-00-34.jpg)
![Screenshot at 01:25: Side-by-side comparison demonstrating the removal of the blinking colon from a digital clock display, highlighting the method's ability to handle small, high-frequency details.](https://ss.rapidrecap.app/screens/RaNay3x0Fmk/00-01-25.jpg)
![Screenshot at 01:47: Slide introducing OmniMatteZero, highlighting that it is 'Training Free' and showing its ability to cleanly separate foreground \(swan\) from background in a low-light water scene.](https://ss.rapidrecap.app/screens/RaNay3x0Fmk/00-01-47.jpg)