您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：上线仅三周，《不问凡尘》在Steam拿到87%好评率，预估近千万流水，超26万玩家走进这个AI生成的修仙世界。
上线仅三周，《不问凡尘》就在 Steam 拿到87%的好评率，预估近千万流水，超26万玩家走进这个AI生成的修仙世界。支撑这款超写实互动作品研发与上线的，是一支约20人的团队。
AI在这里不是简单的提效工具，而是贯穿游戏资产生产、研发协作和交互实现的核心生产力。依托火山引擎的模型和云服务，《不问凡尘》用9个月完成开发上线，整体制作成本约是传统制作的1/3。
它的成绩至少验证了两件事：一是AI正在把游戏制作从“大厂专属”变成“小团队可及”，二是越来越多玩家愿意为AI游戏付费。
一座由AI“长”出来的修仙世界
《不问凡尘》是一款“真人”超写实互动作品：玩家可以探索海量支线剧情、与 NPC 互动、修炼法术，自由选择不同修行路线，像在玩一部修仙小说。
与传统游戏资产生产方式不同，它的核心资产是由豆包视频生成模型 Seedance、豆包图像创作模型 Seedream 和豆包语音合成模型各自承担一类任务，再由项目组完成设定、筛选、修改和最终审核。
Seedance：NPC不再只会站着说话
在这个 Demo 中，玩家以第一人称走进店铺：门外是挂着灯笼的古风街道，室内摆满陶罐、卷轴；男猎户说话时微微转头，女店主从整理货架到回身迎客，眨眼、微笑和摆手自然连在一起。镜头从街道进入室内，再切到人物近景，角色、场景和光影没有断裂感。
这类连续表演正是游戏叙事管线的难点。项目组根据角色设定、动作需求和镜头节奏准备参考素材，使用 Seedance 2.0/2.5生产主要视频资产。模型对动作连续性、人物表情和场景光影的处理，让AI生成内容不只是一张会动的图，而是能够承载剧情的完整镜头。
Seedream：妆发细节到位，角色才立住
角色设计 Demo：薄纱、刺绣、碎发与侧逆光细节清晰，切换界面后人物特征保持一致
角色设计 Demo 里的女性人物手持书卷，豆绿色外袍带着半透明薄纱质感，金色刺绣、鬓角碎发和纸张纹理清晰可见；暖光从侧后方落下，勾出发丝和人物轮廓。切换不同界面，角色五官、妆发细节始终一致。
Seedream 5.0 Pro 参与人物、场景等图像资产生产。团队提供人物设定、服饰材质和场景参考，再通过多轮生成与编辑确定效果。模型对光影、人物肌理和物体材质的表现，帮助团队把角色设定落到可见的细节中，也为后续视频生成建立统一的视觉基准。
Seed TTS + Character：一开口，人物性格就出来了
在角色对话 Demo 中，古风叙事没有被读成平直的旁白，人物语速舒缓，饮酒动作间还留出呼吸和吞咽的节奏；声音与口型、动作相互贴合，让角色更像是在交流，而不是播放台词。
项目组结合人物身份和剧情语境准备台词与声音要求，再由 Seed TTS 结合 Seed Character（豆包角色扮演模型）完成语音生产，而且模型能够根据上下文，演绎出自然、富有表现力的语音，让宗门天骄、市井百姓各有其声，NPC 对话不再是机械念白。
过去，打造一款这样的互动作品，可能需要上百人团队、两三年开发周期、大量美术和实拍成本；如今，AI负责批量生成，团队负责世界观、审美、玩法和最终质量，不到一年就可以用更低的成本实现。
一条会持续运转的AI生产管线
AI资产有了，怎么才能玩起来
要让AI资产真正“玩起来”，还要解决另一个问题：普通AI视频是沿时间轴展开的画面，本身不能响应交互，直接嵌入游戏，往往只是一段不能响应玩家操作的“播片”。
《不问凡尘》项目组把 Seedance 生成的视频拆成不同图层，再用状态机连接剧情节点、角色动作和玩家输入。玩家作出选择或触发 QTE 时，系统会调用相应内容、切换动作状态。
在战斗等玩法上，团队采用了“二三维实时合成技术”，进一步把视频分解为多个图层，在同一时间轴上由不同逻辑控制，并结合3D 渲染、光影合成，让预生成内容能够随玩家操作被调用、切换和组合。AI视频由此不只是“可看”，还真正进入了玩法。
TRAE与Agent：让流程run起来的AI搭子
单个模型不会自然形成生产线，项目组背后是一套由火山引擎承载的自动化生产体系：火山方舟提供模型调用与管理，云平台和 GPU 算力支撑规模化生产，TRAE 与 Agent 工作流把能力带入日常研发。
美术团队借助 TRAE 搭建自动化生产管线和工具，固化素材生成、筛选和版本管理等重复步骤；Agent 串联多模态模型来实现内容的自动化生产。模型负责生成，Agent 负责衔接，人始终负责判断。AI由此从单点工具变成可以反复运行的生产系统。
项目组介绍，这套能力将持续延展：大模型进一步进入智能 NPC、动态剧情、角色生成和 UGC 共创，推动玩法升级；AI管线持续生产新玩法与新内容，稳定内容供给；《不问凡尘》版权资产接入火山方舟版权中心，面向企业客户提供合规授权，延展IP的商业生命周期。
Steam“特别好评”和上线三周的市场表现，不仅说明用户愿意体验并为AI深度参与生产的互动作品付费，更表明《不问凡尘》走通了一条新的路径：模型、Agent 工具和人组成的生产管线，让AI游戏找到一套能够支撑产品上线和商业反馈的新生产方式。
雷峰网(公众号：雷峰网) 雷峰网
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: Only three weeks on the line, No Question Dust received 87 per cent of the rating in Steam, predicting that nearly millions of water flowed, and over 260,000 players entered this AI-generated cuisine world.
It's only three weeks on the line, and no matter what, Steam gets 87% of the rating, predicts that almost millions of water flows, and over 260,000 players walk into this AI-generated cuisine world. It's a team of about 20 people that supports the development and up-to-date of this overwritten interactive work.
AI is not a simple efficiency tool here, but a core productivity that cuts across the production of game assets, R & D collaboration and interaction. Based on the model and cloud service of Volcengine, Unquestioned Dust was developed online for nine months at an overall cost of about one third of traditional production.
Its results confirm at least two things: first, AI is moving the game production from "big-plant exclusive" to "small-team-accessible" and, second, more and more players are willing to play for the AI game paid.
A world of cuisine from AI.
"The Unquestioned Dust" is a "real person" overwritten interactive work: players can explore big-wire dramas, interact with the NPC, refine the arts, and freely choose different paths of practice, like playing a fairy tale.
Unlike the traditional way in which game assets are produced, its core assets are the Doubao Video Generation Model, Seedream, Doubao Image Creation Model, and Doubao Speech Synthesis Model, each with a set of tasks, which are then set, screened, modified and finalized by the project group.
Seedance: NPC no longer just stands and speaks.
In this Demo, the player entered the shop with the first person: Outside the door are the ancient wind streets with lanterns, with pots of pots and scrolls all over the house; when the male hunter speaks, the owner of the store turns a little, and from the shelf to the back of the house, the blinking, smiling and swinging are naturally connected. shot enters the street into the room and cuts into the person's view, without any sense of fragmentation of the role, scene or light.
Such continuous performances are the hard part of the narrative line of the game. The project team prepares the main video assets for reference to asset based on role setting, action needs and shot rhythm, using Seedance 2.0/2.5. The model's treatment of continuity of action, personal expression and scenery allows AI to generate not just a moving picture, but a complete shot that can carry the story.
Seedream: Make-up details are in place.
Role Design Demo: The details of the veil, embroidery, fragmentation and side-reverse are clear, and the identity of the person after the interface is consistent
The female characters in the role design Demo carry a roll of books, green bean robes with a translucent veil, gold embroidery, horny hair and paper texture clearly visible; warm light falls from the back side, tweaking the threads and profile of the person. Toggle different interfaces, the roles are consistent, and the details of hair are consistent.
Seedream 5.0 Pro is involved in the production of image assets such as figures, scenes, etc. The team provides character-setting, clothing material and scenario reference, which is then determined by multiple-wheel generation and editing. The model's performance of light shadows, human muscles and object materials helps teams to set roles in visible details and to create a unified visual benchmark for subsequent video generation.
Seed TTS + Character: The character comes out as soon as it opens.
In the role dialogue, Demo, the ancient narrative was not read as a plain bystander, the character's speech was smooth, and the drinking action left the rhythm of breathing and swallowing; the sound and the mouth, the action, matched each other, making the role more like communication than playing dialogue line.
The project team prepared dialogue line and voice requests in the context of the identity of the person and the situation, and Seed TTS combined with Seed Character (Doubao role-play model), and the model was able to produce a natural, representative voice based on context, so that Zongmen Tiancheng and the city people could speak, and the NPC dialogue was no longer mechanical.
In the past, the creation of an interactive work of this kind could cost hundreds of people, two or three years of development cycles, a lot of art and real-time costs; now AI is responsible for bulk generation, and teams are responsible for worldviews, aesthetics, games and final quality, which can be achieved at a lower cost in less than a year.
An AI production line that will continue to function.
How do I get to play with AI's assets?
To make AI assets truly “play”, another problem is: Normal AI videos are images that go along a time axis and cannot by themselves respond to interactive, directly embedded games, often just a “show” that cannot respond to player operations.
The Unquestioned Dust project group splits the video generated by Seedance into different layers, using the status machine to connect the drama point, role action and player input. When a player makes a choice or triggers a QTE, the system calls the corresponding content, switches the action state.
For games such as combat, the team uses “two-to-three-dimensional real-time synthesis” to further decompose the video into layers, controlled by different logic on the same axis of time, and combined with 3D rendering, photo synthesis, so that the pre-generated content can be called, switched and combined with the player operation. It's not just "seeable," it's actually going into play.
TRAE AND Agent: The AI set that keeps the process running
A single model does not naturally form a production line, and the project group is backed up by an automated production system carried by Volcengine: the volcanic ark provides model call and management, cloud platforms and GPU algorithms support scale production, and TRAE and Agent Z52Z bring their capabilities into daily R & D.
The art team builds automated production lines and tools through TRAE, which consolidates duplicate steps such as asset generation, screening and version management; and Agent's serial multi-modular model to achieve automated content production. Models are generated, Agent connects, and people always judge. AI thus moved from a single-point tool to a production system that can be operated over time.
The project team presented that this set of capabilities would be continuously extended: LLM would move further into smart NPC, dynamic drama, role generation and UGC co-creation, facilitating the upgrading of play practices; AI pipelines would continue to produce new play and content, stabilizing the supply of content; and Unquestioned Dust copyright assets would access the Volcanic Ark Copyright Centre, providing compliance authorization to business clients and extending the IP business life cycle.
Steam's “specially good” and three-week online market performance not only demonstrates the willingness of users to experience and engage in the production of interactive paid for AI's deep participation, but also shows that Unquestioned Dust has a new path: models, Agent tools and a human-made line of production, allowing the AI game to find a new production method that supports up-line and business feedback.
Thunder Peak Network
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.