# Runway Characters 入选 SIGGRAPH 2026 Real-Time Live！ 现场演示 - 来源：Runway：News（网页） - 作者：Leah Retta - 发布时间：2026-08-01 04:34 - AIHOT 分数：49 - AIHOT 链接：https://aihot.virxact.com/items/cms9eics20crbro9kgoo68l7v - 原文链接：https://runwayml.com/news/company-news/runway-characters-siggraph-2026 ## AI 摘要 Runway 的实时交互数字人系统 Characters 入选 SIGGRAPH 2026 的 Real-Time Live! 环节，团队在现场用一张照片数秒内生成可对话的角色。该系统从单张图片出发，无需微调即可适配任意风格，逐帧生成画面以支持长达 30 分钟以上的连续对话。Characters 于今年 3 月上线，团队正探索可导航环境、多参考图像及记忆功能等后续方向。 ## 正文 SIGGRAPH 2026 is the premier conference in computer graphics, with a history that includes the debut of the first Pixar short film. It is also one of the few research venues where practitioners from the world's largest animation and entertainment studios sit alongside academics. This year, Characters was accepted into Real-Time Live!, the conference’s highly-anticipated session where teams have 6 minutes to demo their work live in front of thousands and are judged by a jury of computer graphics and real-time technology experts. We sat down with members of our research and product team behind the work — Yining Shi, Katie Lewis and Taras Khakhulin — to talk about what it took to get there, what the room responded to and where this research is headed next. Runway’s Characters demo at SIGGRAPH 2026. Let's start with the research itself. What problem were you setting out to solve with Characters, and what was the state of the art before this work? Taras Khakhulin: When we made the work public, there was nothing else like it: a system you could talk to that generates its response frame by frame, in real time. Real-time avatars existed before, but most were controlled talking heads under the hood. They restricted you to a human figure and often required specific formats, like a video upload or a 3D asset. Characters starts from a single image. That can be a photo of you, a pet, an animated character or something completely abstract, and it doesn’t need any fine-tuning or per-style adaptation. There was no existing system to reference when we built this, which meant every design decision had to be proven out from first principles. Generalizing to any style, even highly abstract characters, was something few people realized was possible. Real-time is doing a lot of work here. Why does that latency bar matter so much? Katie Lewis: Latency matters because the product is a natural conversation. By the time you stop talking, the character should already be responding, and even a small delay can break the experience entirely. Yining Shi: Hitting that bar took the rendering engine behind Characters, our first real-time model deployed in production, plus months of engineering work streaming video and audio back to users and squeezing milliseconds out of every layer of the stack. Characters shipped in March, and we've seen great reception from our community of users. SIGGRAPH is a different moment: a peer-reviewed stage in front of the graphics research community. What did bringing the work to SIGGRAPH mean to the team? Lewis: The peer review itself meant a lot. Getting selected is the research community validating that the work is technically novel. Runway builds all of its own foundation models, and this was a chance to talk about that research directly with a research audience. And during rehearsals, the judges pushed us to prove the system is truly spontaneous and reactive. Khakhulin: Real-Time Live! is ultimately proof by format. A paper can work in theory, but any demo that survives that stage is work that actually runs. Runway’s product and research team, demoing Characters live at SIGGRAPH 2026. What was the reception like, in the room and since? Shi: The moment that landed hardest was creating a character live: taking a photo of next year's conference chair on stage and having them talking as a character within seconds. Throughout the conference, attendees sat down with us, took a photo and talked with themselves, asking their character every question they could think of. Lewis: A few things consistently surprised people. Conversations can run 30 minutes or more without degrading, when most assumed the ceiling was a couple of minutes. Researchers came up after the session to dig into the architecture, including whether the model is truly autoregressive, and a group presenting later in the conference even quoted Runway as an inspiration for their own work, which is exactly the kind of exchange you hope for at a venue like this. Khakhulin: The frame generation surprised people, too. We generate every single frame rather than looping segments of video to maintain consistency, and that is what makes those long conversations possible. Yining Shi demoing Characters at SIGGRAPH 2026. Source: SIGGRAPH This took the research and platform teams working as one unit. How did those two sides shape each other? Shi: There was no waterfall. Once the model showed promising results, researchers and engineers met almost daily until it shipped. Khakhulin: Some things can only be optimized at the level of the method and architecture, and past a certain point, the gains have to come from engineering. Both had to happen simultaneously to reach the point where the experience feels natural, just like the way people talk in real life, rather than slightly laggy. Runway has prioritized building interactive AI Characters responsibly from the start. What did conversations around safety and moderation look like on the ground at SIGGRAPH? Lewis: Because the model works with any character, people immediately imagine experiences built for children, and they ask about guardrails. Others ask how a character with a specific function stays on task: if it is meant to read bedtime stories and the conversation drifts, how does it redirect appropriately? Moderation and the certainty of moderation came up repeatedly, and it’s something we are actively prioritizing. What are you looking ahead to now? What does this work look like a year from now? Khakhulin: The next step is giving characters somewhere to live. We are working toward navigable environments, so a character can exist in a world you can move through and interact with differently. We are also exploring how you might interact with characters through text and control their expressivity and emotions. Lewis: We’re also exploring making Characters more personal, whether through multiple reference images or a memory and objects of their own. We’re also working on making Characters faster and more accessible to broader audiences. The underlying model keeps improving, and each version of it makes more of this possible out of the box. News Customer Stories How “House of David” Used Runway to Become Amazon’s Latest Hit Series News