全栈 AI 应用Full-stack AI app· 2026
Vivid AI
沉浸式 AI 角色聊天应用Immersive AI character chat app
用两阶段多模态架构解决图片理解后角色一致性被稀释的问题。A multimodal AI chat app focused on preserving persona consistency after image understanding.
案例摘要Case summary
目标用户Audience
希望体验可自定义人格、多模态对话和沉浸式聊天的用户。Users who want to experience customizable personas, multimodal conversation, and immersive chat.
我的角色Role
负责角色管理、多模态对话、流式响应和部署方式的产品决策与验收；全部代码由 AI 编程工具完成。Owned product decisions and acceptance for persona management, multimodal chat, streaming responses, and deployment choices; all code was produced with AI coding tools.
交付结果Outcome
形成可本地运行的全栈聊天应用，并沉淀访问控制、限流、启动检测等演示环境经验。Delivered a full-stack chat app that runs locally, plus demo-environment experience on access control, rate limiting, and startup detection.
产品洞察Product insight
单一模型端到端处理图片时，视觉任务的客观描述会冲淡角色人格。把视觉识别和角色化回复拆成两阶段，可以让多模态对话更稳定。When a single model processes images end-to-end, objective visual descriptions dilute the character persona. Splitting visual recognition and persona-grounded replies into two stages makes multimodal conversation more stable.
项目简介Overview
Vivid AI 支持自定义 AI 人设、多模态图文对话、流式实时响应和会话标题生成。项目重点不是“能聊天”，而是处理多模态任务与角色设定之间的冲突。Vivid AI supports custom AI personas, multimodal text-and-image chat, streaming real-time responses, and conversation title generation. The focus is not being able to chat — it is handling the conflict between multimodal tasks and persona settings.
技术栈Tech stack
结果证据Evidence
README 记录两阶段图片处理、SSE 解析、标题生成和部署踩坑。README documents two-stage image processing, SSE parsing, title generation, and deployment pitfalls.
scripts/screenshot.js 用于自动生成项目截图。scripts/screenshot.js automatically generates project screenshots.
start.ps1 增加后端就绪检测，避免前后端并行启动竞态。start.ps1 adds backend readiness detection to avoid a frontend/backend parallel-start race.
关键决策Key decisions
AI 处理图片后容易忘记角色设定。The AI tends to forget the persona setting after processing an image.
GLM-4V 只负责客观图片描述，DeepSeek 基于完整人设生成角色化回复。GLM-4V only produces objective image descriptions; DeepSeek generates persona-grounded replies from the full character profile.
演示环境公开后可能被爬虫刷接口。An open demo environment can be scraped by bots hitting the API.
增加后端限流和前端访问密码，降低滥用风险。Add backend rate limiting and a frontend access password to reduce abuse risk.
想看完整的代码和实现细节？Want the full code and implementation details?
在 GitHub 上查看View on GitHub