At Devcom, the developer conference running alongside Gamescom, AI stole the show. Microsoft, Tencent, Google, Meta, and other giants brought more than twenty AI-focused sessions, with developers zeroing in on three themes: how AI raises game-art production efficiency, how AI tools integrate with traditional pipelines, and case studies in animation generation and scene construction.
Game-art fidelity demands have grown exponentially in recent years — and the workload has grown geometrically with them. Which is the backdrop for Tencent Games’ global launch of VISVISE, a full-chain AI solution for game creation: a suite of tools organized into four pipelines — animation production, model production, digital asset management, and smart NPCs — covering the entire game-art development flow, focused on the repetitive, mechanical, high-volume “manual labor” that consumes artists’ time.
The showcase numbers: in 3D animation, VISVISE’s MotionBlink takes a few user-supplied keyframes and auto-completes the in-between frames into a full sequence. Historically, hand-keying in-betweens consumed 60-70 percent of character-animation hours — a ten-second animation could take 3-7 person-days of hand-polish. Now, AI completes 200 animation frames in 4 seconds, with 1.5-second-interval optimization pushing some results toward optical-motion-capture quality. A German game-development student who tried it at the booth — an indie developer preparing his own startup — said motion capture was precisely his bottleneck, and MotionBlink significantly lowers the barrier for small teams and founders. And MotionBlink is just one tool: the Superman character-animation solution and GoSkinning auto-rigging have already shipped in real games, including Peace Elite and PUBG Mobile.
What Problem Is a Game AI Model Actually Solving?
In traditional game-art production, 50-60 percent of the work goes into asset creation, with 3D modeling and animation as the most tedious stages. Modeling turns concept art into static digital models; making characters move then requires skeleton binding → skinning → animation.
Skinning, as Tencent Games’ effectiveness-product lead Chen Dong explains, is like dressing a 3D character — not with cloth but with math, making skin and bones move in perfect coordination. When a bone moves, the system must compute how every mesh face follows it: which joints each face is associated with, and with what weights. Weights determine which joints influence a face and how strongly — “if a mesh face binds to only one joint, it only follows that joint’s motion.” Traditionally, artists hand-adjust thousands of weight points, sewing skin to bone with invisible threads of data; skin near joints needs distributed binding so elbow creases fold naturally when the arm rises. The problems: a character may have hundreds of bones, each demanding precise skin deformation control, and one slip causes clipping — an inner garment poking through an outer layer. Even senior riggers spend 60 percent of their time on these issues.
Animation is harder still. The two traditional routes: hand-keying, where frame-by-frame adjustment makes a ten-second animation take a week; and motion capture, which is faster but delivers unstable data quality requiring heavy correction — while stylized, non-generic motions like martial arts or exaggerated expressions often lack data entirely, forcing new capture sessions. AI wants to solve this, but game data’s complexity makes it brutal: skinning requires modeling complex vertex-bone relationships with highly heterogeneous 3D data that resists representation; animation detail is so rich it lacks unified effective representation (like the discretization that powered LLMs), and its spatial-temporal multi-dimensional modeling is an open problem. Above all, the industry’s quality bar is merciless — AI output must drop seamlessly into production pipelines and remain artist-adjustable. A company that does not understand game development, or lacks the data, probably cannot reach even the passing line. The real challenge: building AI that understands both the technology and the game.
Tencent’s Answer
For skinning, VISVISE ships GoSkinning, a two-stage AI: a general skinning foundation model (GCN plus Transformer architecture) predicts bone-chain and skeletal weights, then a local AI pass re-optimizes problem areas — skirt hems, wings, complex parts. Operation is simple: select the model and joints in the scene, click auto-full-skinning. For local fixes, when a character’s skirt shows sticking artifacts, select the wrong joints and vertices and GoSkinning repairs them one-click. Trained on high-quality data accumulated through real game development, the system processes a 20,000-vertex model in 30 seconds — an 8x efficiency gain.
Animation gets the same treatment: set the trajectory, add key poses like takeoff and landing, and MotionBlink’s MIB intelligent in-betweening fills the middle — an autoregressive diffusion approach whose underlying MotionGen model, trained on large-scale high-precision motion-capture data, supports climbing, dance, and other action styles, with some locomotion generations approaching professional mocap quality. The team’s smart-keyframe research was accepted to ACM SIGGRAPH 2025. Critically, the tools embed as plugins into existing software like Maya — developers adopt them without rebuilding their pipelines.
How VISVISE Was Built: Production First, Research Second
Unlike the “invent technology, then find applications” AI playbook, VISVISE was born from Tencent Games’ actual development needs. The timeline: 2016 — AI exploration begins, with deep learning applied to operations efficiency and process optimization. 2017 — reinforcement-learning game-playing research. 2018 — the technical focus extends to the art production (DCC) pipeline. 2022 — GoSkinning 1.0 ships: a data-driven machine-learning approach that learns from existing skinning data and predicts on similar garment datasets, deployed in Peace Elite and QQ Dance. Real production data showed GoSkinning 1.0 alone lifted Peace Elite’s animation-skinning efficiency by about 40 percent. The VISVISE team then kept a long-running collaboration with the Peace Elite project group, shipping custom features against concrete needs; GoSkinning has since iterated to version 4.2, adding skirt skinning, quadruped skinning, and facial binding, with animation-skinning efficiency up more than 60 percent. When generative AI exploded in 2023, the team recognized that large-scale data pretraining through neural networks yielded powerful generative capability — and doubled down on 3D model and animation generation, the most expensive stages of game development. In 2024, Tencent consolidated its scattered AI explorations into a systematic product matrix, and VISVISE formally launched: organized around the six core stages of game creation — model generation, texturing, rigging, animation, scene construction, rendering — with four AI production pipelines, each carrying purpose-built tools trained differentially across game genres.
The business results are already in Tencent’s filings: Q1 and Q2 2025 reports both emphasized AI’s contribution to R&D efficiency, with Q1 noting substantive contributions to evergreen titles and Q2 citing intensified AI adoption as Honor of Kings and Peace Elite evolve toward platform status. VISVISE now serves not only Tencent’s internal projects but numerous industry partners — GoSkinning alone runs in nearly a hundred games.
Games and AI: The Old Marriage
With ChatGPT-5 and Google Genie 3, a thesis is forming: multimodality is becoming AI’s key capability — and games, rich in 2D and 3D interactive scenes, are AIGC’s best application ground. As Chen Dong puts it: “Games and AI are a natural pair.” The history agrees. NVIDIA’s GPUs were built to render game graphics and became the compute foundation for training large models. DeepMind’s reinforcement-learning AI cut its teeth on StarCraft and DOTA 2. Even in hardware, the CUDA cores that run games and the Tensor cores that run AI sit physically adjacent — neighbors by design. The relationship is not coincidence but inevitability: a game’s loop of perceive-environment → decide → act → iterate is, structurally, an AI training ground. An AI that learns to “ambush from grass” in Honor of Kings is learning the same underlying skill as one that learns to “anticipate traffic” in the real world. Games are also one of the largest consumers of 3D assets — and their pursuit of ultimate experience keeps pushing AI past its limits. Developers wanted crying, laughing NPCs years ago; the technology only allowed wooden puppets. Now AI lets NPCs remember your name, mock your gameplay, even feign anger — and the VISVISE bet is that this is still not the ceiling.
