--- title: "Ninety projects share one animation pipeline, and the boring half of the job is disappearing" date: 2026-11-03 category: Generative Media site: NeuroAI canonical: https://neuroai.site/a/na-aigc-motus-animation-pipeline language: en --- # Ninety projects share one animation pipeline, and the boring half of the job is disappearing > Tencent's Motus AI animation pipeline is wired into more than 90 internal and external projects, taking complex costume rigging from days down to a few hours. An animator in Chengdu opens a file, asks for a boss's three-hit attack sequence, and gets back a motion set with the skeleton already bound. Twenty minutes later the same character has hair physics that swing correctly when she turns. The part being automated was never the fun part. It was the binding, the skinning weights, the first pass of the motion. ## Key takeaways - Tencent's Motus AI pipeline has been connected to more than 90 internal and external projects, per the team's own Gamescom Dev presentation, with work reported in PUBG MOBILE, Wuthering Waves (鸣潮) and QQ Dancer (QQ炫舞). - Tencent states automation runs at roughly 50–70% for skeleton generation and 70–85% for general skinning, rising above 85% once post-processing is added. - One-click skinning completion passes 80% on the team's test cases, with multi-layer costumes handled by a dedicated proxy-mesh model. - Tencent says complex costume rigging, including the artist's clean-up, drops from as much as four days to one to four person-hours; a simple costume took about three person-days. - The Gamescom Dev 2026 survey of 100 speakers found 47% expecting machine-assisted development to be the biggest industry trend of the next three years, while 30% want as little machine involvement as possible. ## The unglamorous end of the pipeline Animation is not one task. It is a chain with a glamorous half and a dull half. The glamorous half is performance: the timing of a hit, the arc of a jump, a walk that reads as confident versus one that reads as a commute. Nobody wants that automated, because that is where the craft sits. The dull half is the setup that makes performance possible. Before a character can be animated at all, a technical artist has to build a skeleton inside the mesh, compute skinning weights so the surface deforms with the bones instead of collapsing, and generate a first pass of motion a human will then argue with. Motus AI Team lead Zeng Zijiao put the old numbers plainly at Gamescom Dev: a simple costume took about three person-days to rig and skin, a complex one up to four days. Motus is aimed precisely at that half, structured as three layers. ## The three layers The first layer generates a skeleton from an arbitrary character mesh — from scratch, from a reference rig re-fitted to a new character, or by adding physics-assist bones for hair and skirts. It uses an autoregressive architecture of more than one billion parameters, and Tencent says it removes roughly 30% of the elapsed time there. The second layer is general skinning: a coarse-to-fine framework combining a graph convolution network with a transformer to work out a weight for every vertex-bone pair. On a typical character with a layered skirt and head accessories — around 25,000 vertices and 180 bones — the team says the process takes about ten seconds to two minutes, with one-click completion passing 80%. The model is over 1.3 billion parameters, trained on 80 GPUs for 90 days. The third layer exists because the second one is not enough. Multi-layer costumes make layers interpenetrate and deform unevenly, and Tencent says more than half the characters in its projects wear that kind of clothing. So the company built a separate skirt-skinning model that generates a cleaner proxy mesh, skins the proxy, migrates the weights back and optimises the deformation. The team's own presentation concedes that automatic rigging still struggles with characters carrying many corrective bones. ## What 90 projects actually means A demo of a model generating a walk cycle is a demo. A pipeline connected to more than 90 projects is a different claim: the output survived contact with real production — different art styles, engines, hardware targets, and animators who had to accept the result or fix it. Tencent describes Motus as sitting inside its shared public technology line (腾讯游戏公共技术线) — infrastructure studios consume rather than rebuild. It is reachable through the web and through plug-ins for 3ds Max, Blender, Maya and Unreal Engine, which is how it arrives inside an existing workflow, not beside it. The honest reading is that the pipeline changes who does the first hour of setup work, not that the hour vanishes. Tencent says the same from the other side: efficiency gains did not remove artists, because editability and artist control are requirements for a production-grade system. ## The industry mood, measured Motion generation runs on a unified diffusion transformer of about 1.5 billion parameters accepting text, video, keyframes or audio. Tencent reports the shift to that backbone cut pose error by 33.8% and foot-skating error by 60.2% against its earlier work. Gamescom is where this sector argues about itself, and the 2026 numbers show the argument has moved past replacement. The Gamescom Dev survey of 100 speakers found 47% expecting machine-assisted development to be the trend with the biggest impact over the next three years — ahead of hardware and memory pricing at 28%. But 30% would prefer as little machine involvement as possible, and only 34% saw the greatest value in code and production. On jobs, 36% expected roles to change rather than team sizes to fall, and 17% expected no cuts at all. The split matters. The people who attended have to live with the tooling, and they describe a division of labour: let the machine take the setup, keep the performance. According to Li An, chief scientist at BrainNet (脑机网), the survey result is a labour-market signal rather than a quality verdict: what separates studios is the shape of their own pipeline, not which model wins benchmarks. ## Honest limitations The 90-plus project figure, the automation rates and the day-to-hour comparisons are Tencent's own disclosures, presented by the team at its own conference session; they are not independently audited, and "automation rate" is not defined in the public material in a way a third party could reproduce. The same applies to the model sizes, the training budget of 80 GPUs for 90 days, and the error-reduction percentages, which are measured against the team's earlier internal versions rather than against outside systems. The Gamescom Dev survey is a small sample of 100 self-selected speakers, and it measures expectations rather than outcomes. Note also that Tencent was the only Chinese company presenting in the AI track at Gamescom Dev in 2026, per China Daily and China Securities Journal — a fact about the conference programme, not a ranking of capability. Finally, the deployment figure counts connected projects, not the share of any title's animation work it handles.## What readers can do now - If you run animation or technical art, time one character end to end — bind, skin, first motion pass — and find the hours that are not performance decisions. That is the part these tools remove. - If you evaluate a vendor's automation claims, ask for the post-processing rate and the clean-up rate separately, and ask what the artist still touches. A pipeline automating 85% of steps that then needs a day of manual correction is not an 85% saving. - If you are hiring into this area, weight the portfolio over the prompt: the durable skill is knowing which fixes are cheap to automate and which are the performance. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-motus-animation-pipeline Free to quote with attribution and a link to the original.