本文档记录 FlintStudio 的版本更新内容。
This file documents notable changes to FlintStudio.
-
稳定性增强
- 优化任务超时配置:视频合成从 30 分钟延长至 60 分钟,图片生成从 5 分钟延长至 10 分钟
- 实现熔断器机制:API 调用失败超过阈值时自动熔断,防止级联故障
- 增加指数退避重试:任务失败自动重试 3 次,提高成功率
- 错误分类处理:区分关键错误、可重试错误、可降级错误
-
一致性控制
- 新增角色一致性控制:通过 seed 锁定和描述保持角色形象稳定
- 新增场景一致性控制:保持场景光线、环境元素一致
- 智能提示词增强:自动在提示词中注入一致性参数
-
性能优化
- 数据库索引优化:为常用查询添加 15+ 个索引,提升查询速度
- 批量出图优化:支持并发批量生成,减少总体等待时间
- 进度预估系统:实时计算剩余时间和完成百分比
-
降级处理
- 图片生成失败时使用占位图,不阻塞整个工作流
- 失败面板记录详细错误信息,便于排查
-
Stability Enhancements
- Optimized task timeouts: video composition extended from 30min to 60min, image generation from 5min to 10min
- Implemented circuit breaker: automatic circuit breaking when API failures exceed threshold
- Added exponential backoff retry: automatic retry up to 3 times on failure
- Error classification: distinguish critical, retryable, and degradable errors
-
Consistency Control
- Character consistency: seed locking and description for stable character appearance
- Scene consistency: maintain lighting and environment elements
- Smart prompt enhancement: auto-inject consistency parameters into prompts
-
Performance Optimization
- Database indexes: added 15+ indexes for common queries
- Batch image generation: concurrent batch processing
- Progress estimation: real-time remaining time and percentage calculation
-
Graceful Degradation
- Placeholder images on generation failure without blocking workflow
- Detailed error logging for failed panels
- 桌面版(Electron)
- 新增
desktop/子目录,支持打包为 Windows 安装包(NSIS),一键安装、双击快捷方式即运行 - 内嵌 Redis:构建时自动下载 Windows 便携版并打入安装包,用户无需单独安装 Redis
- 内置 Web UI 与窗口:安装后自动启动 Next.js + Worker,打开本地浏览器窗口
- 支持 OpenClaw 远程控制(连接本机地址即可)
- 提供
docker-compose.desktop.yml用于启动 MySQL(可选 Docker 方式)
- 新增
- 画风(视觉风格)
- 一键成片前可选择画风(写实实拍、3D 虚幻 CG、漫剧、日式动画、3D 国漫、电影感、美漫等)
- 分镜与出图统一采用所选风格;所选画风保存为项目默认
- 代码健康度修复
- advanceRun 并发竞态:Redis 分布式锁保护 phase 切换
- Worker 任务幂等:重试时已完成任务直接跳过
- episodeId 路径校验、OpenClaw 参数校验、Next.js 启动时 validateEnv
- 多步 DB 写入使用事务;API body 校验与视频流式返回;Shell 转义与 SSRF 防护
- Desktop (Electron)
- New
desktop/subdirectory: package as Windows installer (NSIS), one-click install and run from shortcut - Embedded Redis: auto-download Windows portable Redis during build; users do not need to install Redis
- Built-in Web UI and window: starts Next.js + Worker and opens local browser window
- OpenClaw remote control supported (point to localhost)
docker-compose.desktop.ymlfor MySQL (optional Docker)
- New
- Visual style
- Choose visual style before one-click run (live-action, Unreal CG, manhua, anime, 3D donghua, cinematic, American comic)
- Storyboard and image generation follow the selected style; choice saved as project default
- Code health fixes
- advanceRun race: Redis distributed lock for phase transition
- Worker idempotency: skip already completed tasks on retry
- episodeId path validation, OpenClaw param validation, validateEnv on Next.js startup
- DB transactions for multi-step writes; API body validation and video streaming; shell escaping and SSRF protection
- 多模型LLM适配器: 全新轻量级模型适配系统
- 支持 8+ 种主流模型:GPT-4o/Claude/DeepSeek/Moonshot/GLM/本地模型等
- 自动模型适配策略:根据模型特性优化提示词和参数
- 用户可配置默认模型和自定义模型参数
- 自动成本估算和 Token 统计
- 自我修复Agent系统: 类似OpenClaw的自动诊断修复
- 自动检测并修复 6 种常见错误类型
- 智能修复策略:JSON修复、模型降级、任务拆分、指数退避等
- 学习机制:记录修复成功率,优先使用高效策略
- 支持自动和手动触发修复
- OpenClaw深度集成: 完整的远程控制能力
- 新增
/api/openclaw/control控制接口 - 支持工作流暂停/恢复/取消/重试
- 支持队列管理:暂停/恢复/清空
- 支持模型切换和连接测试
- 支持系统诊断和自动修复触发
- 新增
- 数据库扩展: 新增
HealingIssue表记录修复历史
- Multi-Model LLM Adapter: New lightweight model adaptation system
- Support 8+ mainstream models: GPT-4o/Claude/DeepSeek/Moonshot/GLM/Local
- Automatic model adaptation: optimize prompts based on model characteristics
- User-configurable default models and custom parameters
- Automatic cost estimation and token tracking
- Self-Healing Agent System: Auto-diagnosis and repair like OpenClaw
- Auto-detect and fix 6 common error types
- Smart healing strategies: JSON repair, model fallback, task splitting, exponential backoff
- Learning mechanism: track success rates, prioritize effective strategies
- Support automatic and manual healing triggers
- OpenClaw Deep Integration: Complete remote control capabilities
- New
/api/openclaw/controlAPI endpoint - Workflow pause/resume/cancel/retry support
- Queue management: pause/resume/clear
- Model switching and connection testing
- System diagnosis and auto-fix triggering
- New
- Database Extension: Added
HealingIssuetable for repair history
- 构建优化: 修复模块路径引用问题,优化 Docker 多阶段构建
- 健康检查: 新增
/api/health端点,支持数据库和 Redis 健康检测 - 依赖清理: 移除未使用的
openai和react-hot-toast依赖 - 日志系统: 新增结构化日志工具
src/lib/logger.ts - 错误处理: 新增统一错误处理工具
src/lib/utils/error-handler.ts - Docker 优化: 启用 standalone 输出模式,减小镜像体积
- Build Optimization: Fixed module path references, optimized Docker multi-stage builds
- Health Check: Added
/api/healthendpoint with database and Redis health checks - Dependency Cleanup: Removed unused
openaiandreact-hot-toastdependencies - Logging System: Added structured logging utility
src/lib/logger.ts - Error Handling: Added unified error handling utility
src/lib/utils/error-handler.ts - Docker Optimization: Enabled standalone output mode for smaller image size
- 提示词系统全面优化: 参考 AI 导演工作流优化四个阶段的 System Prompt
- 剧本分析:强化角色视觉设定、场景空间信息提取
- 分场:明确分场原则,规范摘要和场景命名格式
- 分镜:增加表演调度深化(面部肌肉组、三锚点肢体法)、光影色彩转译、景别运镜词典
- 配音提取:增加情绪标注、语气提示、音效标注
- 自定义提示词: 支持在设置中心自定义四个阶段的 System Prompt
- 本地模型优化: 本地端点(localhost/127.0.0.1)无需配置 API Key
- 自动创建项目修复: 改为 POST 接口,防止浏览器预加载意外创建项目
- 数据模型扩展: Panel 新增 metadata 字段,VoiceLine 新增 lineType/emotion/tone/audioNote 字段
- Prompt System Optimization: Enhanced System Prompts for all 4 workflow phases referencing AI director workflow best practices
- Custom Prompts: Support customizing System Prompts in settings
- Local Model Improvements: Local endpoints don't require API Key
- Auto-create Project Fix: Changed to POST endpoint to prevent accidental project creation
- Data Model Extensions: Added metadata to Panel, added lineType/emotion/tone/audioNote to VoiceLine
- 本地模型支持: 新增对接 Ollama 和 ComfyUI 本地模型
- Ollama: 本地 LLM 支持(llama3.2, qwen2.5 等)
- ComfyUI: 本地 Stable Diffusion 图像生成
- 零成本、隐私保护、无限生成
- OpenClaw 远程控制: 新增
flintstudio-controlSkill,支持通过 Telegram/飞书等 IM 平台远程控制 FlintStudio- 创建项目、启动工作流、检查状态、获取结果
- 提供 Telegram Bot 和飞书 Bot 完整示例代码
- OpenClaw API 端点: 新增 7 个 REST API 端点供 OpenClaw 调用
- 文档精简: 删除所有冗余文档,仅保留 CHANGELOG.md 和 README.md
- Local Model Support: Added support for Ollama and ComfyUI local models
- Ollama: Local LLM support (llama3.2, qwen2.5, etc.)
- ComfyUI: Local Stable Diffusion image generation
- Zero cost, privacy protection, unlimited generation
- OpenClaw Remote Control: Added
flintstudio-controlSkill for remote control via Telegram/Lark IM platforms - OpenClaw API Endpoints: Added 7 REST API endpoints for OpenClaw integration
- Documentation Cleanup: Removed all redundant docs, keeping only CHANGELOG.md and README.md
- OpenClaw Skill 增强: 新增完整的 AI 控制功能,支持通过 OpenClaw 龙虾 AI 部署、启动、停止、更新、备份 FlintStudio;新增 config、shell、clean、doctor、port、restart、restore 等命令
- API 中转支持: 新增 Comfly 和云雾 API 中转服务支持,在 .env.example 中添加配置示例
- 版本控制: 项目进入 Beta 0.20 版本,建立版本发布流程
- 系统 Prompt 全面优化:
- OpenClaw Skill: 增强上下文记忆、风险预判、场景示例、持续学习能力
- Workflow Prompts: 添加 Few-shot 示例、结构化输出、约束条件
- 新增
review-analysis.ts: 复查分析专用 prompt,含评分细则、错误模式识别、改进建议模板 - 新增
analyze-novel.ts: 剧本分析专用 prompt,含角色/场景提取、分集策略
- 项目信息: 在 README 底部添加 GitHub 仓库地址 https://github.com/flintcore/FlintStudio/ 和作者邮箱 qihuanteam@gmail.com
- README 优化: 补充 OpenClaw Skill 完整命令列表、子 Agent 能力说明、项目结构更新
- Enhanced OpenClaw Skill: Added full AI control features for deploying, starting, stopping, updating, and backing up FlintStudio; new commands: config, shell, clean, doctor, port, restart, restore
- API Proxy Support: Added support for Comfly and Yunwu API proxy services with configuration examples
- Version Control: Project enters Beta 0.20 with established release process
- Comprehensive Prompt Optimization:
- OpenClaw Skill: Enhanced context memory, risk prediction, scenario examples, continuous learning
- Workflow Prompts: Added Few-shot examples, structured output, constraint conditions
- Added
review-analysis.ts: Dedicated review analysis prompt with scoring criteria, error pattern recognition, improvement templates - Added
analyze-novel.ts: Novel analysis prompt with character/scene extraction, episode strategy
- Project Info: Added GitHub repository https://github.com/flintcore/FlintStudio/ and author email qihuanteam@gmail.com to README footer
- README Optimization: Added complete OpenClaw Skill command list, sub-agent capabilities, updated project structure
- 画风(视觉风格):一键成片前可选择画风(写实实拍、3D 虚幻 CG、漫剧、日式动画、3D 国漫、电影感、美漫等),分镜与出图将统一采用该风格;所选画风会保存为项目默认,下次自动回填。
- API 对接:统一 OpenAI 兼容路径(
/v1/chat/completions、/v1/images/generations、/v1/audio/speech),Base URL 自动补/v1;设置中支持为 LLM/图像/视频配置可选模型名,便于多模型、多端点对接。 - 多 Agent 与复查:剧本分析完成后自动执行「复查 Agent」对集数/角色/场景做质检,结果写入步骤;运行详情展示每步简要结果(集数、复查通过/问题),避免黑盒。
- 初始开源版本:小说 → 剧本分析 → 分场 → 分镜 → 出图 → 配音 → 视频合成,全流程可配置 API。
- Visual style: Before one-click run, you can select a visual style (e.g. live-action, Unreal CG, manhua, anime, 3D donghua, cinematic, American comic); storyboard and image generation will follow it. The choice is saved as the project default for next time.
- API: OpenAI-compatible paths with auto
/v1; optional model names per type (LLM/image/video) in settings for multi-model and multi-endpoint support. - Multi-agent & review: After script analysis, a review agent validates episodes/characters/locations; step results are stored and shown in run details (no black box).
- Initial release: novel → analysis → scene split → storyboard → images → voice → video; all APIs configurable.
格式说明 / Format: 版本号 [x.y.z] 下列出该版本的变更条目。
Add new entries under ## [x.y.z] - YYYY-MM-DD when releasing.