Naver’s HyperCLOVA X Imaginative and prescient proves {that a} image is price a thousand phrases
An image is price a thousand phrases, because the saying goes, emphasizing the facility of imaginative and prescient over textual content.
Individuals additionally say that eyes are home windows to the soul, emphasizing the significance of people’ skill to obtain refined visible info.
Naver Corp., a number one South Korean expertise large, stated Thursday that it has educated the mind of its newest synthetic intelligence platform, HyperCLOVA X, to grasp pictures in addition to textual content.
On August 27, Naver plans to unveil HyperCLOVA X Imaginative and prescient (HCX Imaginative and prescient), one other upgraded model of HyperCLOVA X, after being educated with giant quantities of textual content and picture information to course of visible info, together with paperwork.
“We’re including picture capabilities to HyperCLOVA X with out compromising its textual content capabilities,” the corporate stated in a press release.
HyperCLOVA X Naver Imaginative and prescient
Naver stated that HCX Imaginative and prescient has moved from a big language mannequin (LLM) to a big imaginative and prescient language mannequin (LVLM).
Skilled on visible information and broad language, HCX Imaginative and prescient helps each textual content and picture modes and performs duties in varied eventualities, comparable to doc recognition and understanding textual content inside pictures, he stated.
SCORE HIGHER THAN GPT-4o
Naver stated it makes use of over 30 benchmarks to trace the efficiency of HCX Imaginative and prescient in comparison with industrial AI fashions Open AI GPT-4v and GPT-4o.
One benchmark that Naver used to measure and showcase its mannequin’s Korean capabilities was the Korean Common Instructional Growth (Okay-GED) checks, that are major and secondary training equivalency diplomas.
HyperCLOVA X Naver Imaginative and prescient
The benchmark consisted of 1,480 four-option multiple-choice questions. When testing with picture inputs, HCX Imaginative and prescient answered 83.8% of questions appropriately, surpassing the Okay-GED check go threshold of 60% and the 77.8% scored by GPT-4o, in keeping with Naver.
Beneath the picture caption class, he stated HCX Imaginative and prescient can precisely establish and describe small particulars in a picture with out utilizing a separate object detection mannequin.
HCX Imaginative and prescient can title historic figures, landmarks, merchandise and meals with simply picture inputs. It might probably additionally purpose and predict attainable subsequent steps primarily based on pictures.
HyperCLOVA X Naver Imaginative and prescient
UNDERSTAND CHARTS, TABLES AND GRAPHS
Naver stated the AI mannequin additionally understands charts, tables and information in an Excel file.
“If the information is a screenshot of a picture, getting responses to your prompts is extra difficult as a result of the mannequin should first acknowledge textual content and perceive how the numbers are associated,” he stated.
HCX Imaginative and prescient helps paperwork in Korean, English, Japanese and Chinese language, he stated.
Naver stated that HCX Imaginative and prescient has been educated on numerous picture and textual content pairs and may even perceive humor and memes.
HyperCLOVA X Naver Imaginative and prescient
Different skills embody understanding equations; generate code utilizing shapes, charts or graphs; remedy maths issues involving shapes; and artistic writing comparable to poems.
“At the moment, HyperCLOVA X Imaginative and prescient can perceive one picture at a time. However quickly, with assist for context lengths within the hundreds of thousands, we anticipate HCX Imaginative and prescient to grasp hour-long films and video streams,” Naver stated.
X’s speech
On Thursday, Naver additionally unveiled Speech X, a voice synthesis expertise primarily based on its HyperCLOVA X.
HyperCLOVA X Naver Imaginative and prescient
Naver stated that Speech X is a extra superior mannequin than present voice synthesis and recognition expertise, with higher language construction and pronunciation accuracy. It might probably additionally categorical feelings like a human being, Naver stated.
The corporate has already confirmed its technological competitiveness with varied AI voice companies comparable to AI CLOVA Notice voice recording, CLOVA Care Name AI phone service and CLOVA Dubio AI voice synthesis.
“HCX, which began as a large-scale language mannequin, is evolving into an enormous visible language mannequin with further picture understanding capabilities, and additional right into a multimodal voice language mannequin,” stated Sung Nako, head of Hyperscale AI Expertise at Naver Cloud Corp., Affiliate AI Naver Corp.
“We’ll broaden our HCX ecosystem by making use of the superior capabilities of HCX to numerous Naver companies, together with CLOVA X.”
Write to Seung-Woo Lee at leeswoo@hankyung.com
In-Soo Nam edited this text.