ARC Lab’s Blueprint: Turning Multimodal AI Research into Real-World Impact

ARC Lab (PCG Application Research Center) was founded in Tencent’s 930 project, rooted in content and social applications, growing steadily while exploring rational development space for corporate research.

Background

Corporate Research ("司研") refers to internal research organizations established by companies such as GE, Bell Labs, DuPont, Kodak, and IBM in the early 20th century. After a golden era post‑World War II, corporate labs declined in the late 1970s due to antitrust actions and reduced valuation in M&A. By the end of 2018, corporate research faced a bleak outlook, even as deep learning began to emerge. In this context, ARC Lab was created with the mission to conduct “top‑level” research.

Mechanism

Small Team

ARC seeks the best talent not only by qualifications but also by fit, passion for frontier exploration, and a strong desire to push technology to its limits. By 2019 the team had only three members; today it maintains fewer than 20 full‑time staff and about 30 interns, with 85% of full‑time members holding PhDs and 70% of interns being PhD candidates from top universities.

Technology‑Impact Funnel

Traditional corporate labs are backed by the company but can be distant from business. ARC leverages PCG’s content and social ecosystem, focusing on projects that sit at the intersection of business demand and frontier interest. The funnel consists of four layers: academic impact, industry impact, business impact, and societal impact. Each year ARC publishes 30‑40 top‑conference/journal papers (academic layer). Roughly one‑quarter of these projects gain community attention (industry layer), another quarter lead to product deployment (business layer), and a few achieve broader societal recognition.

Performance System

Since 2019 ARC has used a quantitative performance system aligned with the impact funnel. “North‑star” projects are those accepted by top conferences/journals and widely deployed in products. Scores are calculated from project impact depth and individual time contribution, fostering a shared consensus on technology impact.

Investment Direction

ARC targets high‑uncertainty problems, specifically multimodal understanding and generation—integrating text, images, video, and 3D. This aligns with both cutting‑edge research and PCG’s content/social business needs.

Academic Impact: The H‑Index Story

The H‑Index measures both quantity and quality of publications. ARC’s 2025 H‑Index is 66, meaning 66 papers have each been cited at least 66 times. The lab’s most‑cited work (Real‑esrgan, 2021) has 1,769 citations, contributing one point to the H‑Index.

Key high‑citation papers span 2D/3D/4D generation, multimodal unified models, and multimodal parsing.

Multimodal Unified Large Model

ARC’s work on unified multimodal models (e.g., the SEED series) aims to combine understanding and generation across modalities. Early models achieved conceptual-level fusion, while later models like MindOmni enable fine‑grained editing and global transformation of images.

Industry Impact: Open‑Source Contributions

ARC’s open‑source projects have garnered significant community attention, with 26 projects receiving over 500 stars and more than 140,000 total stars.

Video Re‑generation Technologies

DepthCrafter estimates per‑pixel depth in video, enabling high‑quality video editing. It was accepted as a Highlight paper at CVPR 2025, downloaded over 1.5 million times on HuggingFace, and earned 1.3 K stars. Variants such as ViewCrafter (single‑image camera trajectory) and TrajectoryCrafter (video camera trajectory) have also achieved strong community adoption.

Business Impact: Product Integration

ARC’s research has been integrated into products such as PhotoMaker (avatar generation) and the Hunyuan‑ARC‑7B multimodal model, which provides video‑level understanding and generation capabilities. The model improves search accuracy from 74 % to 90 % and boosts CTR by 5.88 % in QB search, and powers video summarization features in WeChat Yuanbao.

Social Impact: Technology with Warmth

ARC’s societal contributions include virtual heritage projects (e.g., collaborations with SSV and the Sanxingdui Museum) that received widespread media coverage.

Talent Impact

The technology‑impact funnel also enhances talent attraction, retention, and conversion. Academic impact draws top PhD talent, while open‑source and industry impact provide practical experience, creating a virtuous cycle that strengthens both research and product development.

Afterword

ARC Lab leverages technology impact to generate talent impact, attracting and retaining top AI talent, which in turn fuels further technological breakthroughs, forming a positive feedback loop in the rapidly evolving AI era.