Turing Develops Visual Data Compression Technology for Large-Scale AIAchieving Significant Capacity Reduction with High-Quality Reconstruction
Turing Inc. (Headquartered in Shinagawa, Tokyo; CEO: Issei Yamamoto, hereinafter “Turing”) has developed a new technology (patent pending) that efficiently compresses large-scale video and image data while retaining high accuracy in a format optimized for AI. By combining a learning mechanism that aggregates key information locally with a data allocation approach based on importance, this technology enables high-speed and high-precision data utilization for autonomous driving AI, multimodal AI, and more.
Release Background
In recent years, multimodal large language models (MLLMs), which handle multiple data types such as images and text simultaneously, have drawn increasing attention. The demand for advanced development capable of processing vast amounts of data continues to rise. However, conventional image embedding techniques face challenges in efficiently delivering information in a format truly optimized for AI.
Technology Overview
Turing’s newly developed technology provides a mechanism to compress massive amounts of data while preserving essential information with high accuracy. By converting various data—such as text and images—into a sequence of tokens (the smallest units processed by AI) and introducing a variable-length compression approach, users can add or remove tokens as needed. This makes it possible to significantly reduce data size without compromising required image quality or analytical precision.
A technique called Tail Token Drop is employed during training. It randomly deletes tokens from the end of the sequence and compares the differences to optimize the model. This process effectively concentrates critical information toward the beginning of the token sequence, minimizing the loss of crucial data even under higher compression ratios.
Furthermore, images can be reconstructed from the token sequences. Compared to traditional image formats such as JPEG or WebP, this method can recreate visually natural images using fewer bytes, opening up new possibilities for real-time autonomous driving systems and cloud-based platforms where communication costs and responsiveness are vital.
Furthermore, details on this technology have been published in the paper One-D-Piece: Image Tokenizer Meets Quality-Controllable Compression and are explained in depth on our tech blog. The model files and source code are available for commercial use under the Apache License 2.0.
| Project Page | https://turingmotors.github.io/one-d-piece-tokenizer/ |
|---|---|
| Tech Blog | https://zenn.dev/turing_motors/articles/6d77c5a3b3712e |
Potential Applications
1. Autonomous Driving
By compressing onboard camera footage into fewer tokens, this technology can deliver visual data to autonomous driving base models more efficiently. Preventing excessive computational load when handling large volumes of data enables large-scale AI models to rapidly recognize and evaluate the surrounding environment in real time.
2. Multimodal Models and World Models
Images and videos tokenized using this technology are expected to be directly used as inputs and outputs for MLLMs or world models, similar to language tokens. By adjusting the number of tokens according to context and data volume, one can ease the computational burden of training and inference in these models while maintaining overall accuracy.
Company Overview
Company Name: Turing Inc.
Location: 4th Floor, East Tower, Gate City Osaki, 1-11-2 Osaki, Shinagawa-ku, Tokyo
Representative: CEO Issei Yamamoto
Founded: August 2021
Business: Development of fully autonomous driving technology
URL: https://tur.ing/
Career Opportunities
Turing is seeking individuals who are eager to change the world by making fully autonomous driving a reality. We frequently host company introduction events and autonomous driving experience sessions, so feel free to reach out to us.
Careers Page: https://tur.ing/jobs
Event information: Connpass
Media Inquiries
PR Contact (Hiraku Abe): pr@turing-motors.com