Skip to content

v2.5.1

Choose a tag to compare

@jeffstric jeffstric released this 22 Oct 03:46
· 310 commits to main since this release

智剧通(ZJT)2.5.0 发布说明 / Release Notes

数据来源:飞书多维表格「GitHub 看板」已发布列,当前共 74 条记录,其中 60 条(GitHub 状态 Open,看板可见)为 2.4.0 快照之后新发布的内容,全部计入本版本。
该列另有 14 条 GitHub 状态为 Closed 的记录,属于此前版本已发布内容,未计入(与 2.4.0 口径一致)。


一、中文

一句话说明

智剧通 2.5.0 迎来数字人全面升级:对接新音频接口并基于 MiniMax H3 生成数字人视频,超长数字人自动多轮拼接,支持音色选择与角色音色描述;同时新增视频提取剧本、抖音图文转视频、声音库与智能体打通、故事板与工作流互切、每日算力统计等功能,并修复数字人、故事板、画布、多人协作权限、RunningHub 槽位泄漏等 30 余个问题。

详细说明

✨ 新功能

  • 视频提取剧本:支持输入视频自动提取剧本。
  • 抖音图文转视频:新增抖音图文转视频能力,图文素材一键成片。
  • 声音库打通:现有声音库与智能体全面打通。
  • 剧本权限管理:剧本拆分师支持更新剧本权限。
  • 故事板 ↔ 工作流互切:工作流页面与故事板支持相互跳转切换。
  • 只看自己:故事板、工作流多人共享模式新增"只看自己创建"选项;故事板支持查看仅自己或指定用户的作品。
  • 画布导出压缩包:画布导出支持类似故事板的压缩包格式。
  • @ 预览:工作流提示词框输入 @ 后,下拉框右侧展示详细预览图(对齐 storyboard)。
  • 算力统计:新增每日各用户算力消耗的统计与表格。
  • 世界管理:管理员可删除或隐藏其他世界(含所有剧本智能体中的世界)。

🤖 模型与供应商(数字人全面升级)

  • 新音频链路:对接新的音频接口,音频生成完毕后自动调用 MiniMax H3 生成数字人视频。
  • 更快更省的数字人:数字人生成切换为h3方案,速度更快、算力更省,并支持音色选择。
  • 超长数字人拼接:数字人时长超过 15 秒时自动多轮 H3 生成并拼接;超长数字人整体从 wan2.2 切换为 MiniMax H3 拼接方案。
  • 角色音色描述:H3 角色新增音色描述,生成视频时可传入,降低声音来源为 TTS 的比例。
  • 说话档案(人设):调整创建角色的提示词,为对话角色增加说话档案,台词更贴合人设。
  • 默认性价比模型:新用户默认选择性价比模型(剧本智能体、营销智能体、故事板、画布)。

🐛 问题修复

  • 数字人:修复画布中数字人未走 MiniMax H3 的问题;修复 H3 首帧/参考生视频在角色不说话时产生无法识别说话内容的问题;修复单角色说话时非说话角色音色未传入视频模型的问题;修复生成堵塞时 4 分钟内密集重试、进程堵死异步卡槽、长度不一致无法生成、分辨率错误与多片段视频劣化、手部不会动等一系列问题。
  • 剧本智能体与协作:修复剧本智能体链接分享给其他用户(链接携带 user_id)后消息无法发送的问题;修复多人协作模式下权限判断不严、直接把 URL 参数中的 user_id 视为真实用户的问题;修复使用他人世界角色后无法修改音色的问题;修复角色行为描述中出现"换人测试"的情况;剧本过短时拆分失败增加友好化提示;剧本拆分默认不再使用对口型模式。
  • 故事板:修复宫格检查方案误杀率过高、宫格图反复失败一直卡住无法生成的问题(失败自动降级为直接生图);修复故事板 render.js path 属性报错;修复分镜视频来源为原声、预览时却播放 TTS 声音的问题;修复修改视频提示词后对话框下方提示词未同步(onblur 同步)的问题。
  • 画布:修复画布滚动选择无记忆、反复询问用户的问题;修复快捷键/滚轮选择弹窗在选择后仍反复弹出的问题。
  • 上传:修复上传角色/道具/场景图片过大时没有明确报错提示的问题;修复上传角色图误触发重新生成、导致用户图片被覆盖的问题。
  • 工作流:修复工作流页面大量请求未携带认证 token、后端 401 响应缺少 detail 的问题。
  • 供应商:修复 RunningHub 槽位泄漏(任务取消后已领取槽位不释放、永久占用并发容量)与 RunningHub 池错误永久占用的问题;修复商业版 lite 不支持音色情感控制的问题。
  • 平台:修复官网 pipeline 步骤永久卡死、宫格生图轮询 401 无限重试、GPT Image 参考图 404/410 被误报为"服务异常"的问题;修复暂存区删除角色/场景/道具的问题;修复开发环境下 CDN 音频无法下载、签到管理页面易误导用户的问题;修复 10.20 测试批次发现的一批问题(含时间轴导出素材包/剪映草稿对非管理员用户 401 失败等)。

⚙️ 稳定性与内部改进

  • 进程模型升级:进程池从 fork 切换为 spawn,根治 logging 锁死锁、EMFILE 管道泄漏、七牛线程池假超时等同根因隐患。
  • 测试基建:修复 GitLab e2e 测试经常挂的问题。
  • 体验打磨:优化故事板左下角对话框样式、grid 模式时间轴按钮更显眼;优化画布音频生成节点与提取剧本节点的功能和易用性;角色参考图由三视角改为去头图片;数字人放弃使用视频尾帧、改为直接并发生成。

二、English

One-line Summary

ZJT 2.5.0 delivers a major digital-human upgrade — a new audio pipeline with MiniMax H3 video generation, automatic multi-round stitching for long avatars, voice selection and per-role voice descriptions — plus video-to-script extraction, Douyin image-text-to-video, voice-library integration with agents, Storyboard↔Workflow switching, daily compute-usage statistics, and 30+ bug fixes across digital humans, storyboard, canvas, collaboration permissions, and RunningHub slot leaks.

Detailed Notes

✨ New Features

  • Video-to-Script Extraction: Scripts can now be automatically extracted from an input video.
  • Douyin Image-Text-to-Video: Added the ability to turn Douyin image-text posts into videos in one click.
  • Voice Library Integration: The existing voice library is now fully connected with agents.
  • Script Permission Management: Script splitters can now update script permissions.
  • Storyboard ↔ Workflow Switching: The workflow page and storyboard can now jump between each other.
  • "Only Mine" View: Storyboard and workflow gain an "only show my own creations" option in shared mode; storyboard supports filtering works by specific users.
  • Canvas Export Package: Canvas export now supports a storyboard-like compressed package.
  • @ Preview: In the workflow prompt box, the @ dropdown now shows detailed preview images on the right (aligned with storyboard).
  • Compute Usage Statistics: Added daily per-user compute consumption statistics with tables.
  • World Management: Admins can delete or hide other worlds (including worlds in all script agents).

🤖 Models & Providers (Major Digital-Human Upgrade)

  • New Audio Pipeline: Integrated a new audio API; after audio generation completes, MiniMax H3 is invoked automatically to generate the digital-human video.
  • Faster, Cheaper Digital Humans: Digital-human generation switches to the h3 approach — faster and more compute-efficient — with voice selection support.
  • Long-Avatar Stitching: Avatars over 15 seconds are automatically generated in multiple H3 rounds and stitched; long avatars overall move from wan2.2 to the MiniMax H3 stitching approach.
  • Role Voice Descriptions: H3 roles gain voice descriptions that can be passed into video generation, reducing the share of TTS-sounding output.
  • Speaking Profile (Persona): Role-creation prompts now include a speaking profile for dialogue roles, making lines fit the character.
  • Default Cost-Effective Models: New users default to cost-effective models (script agent, marketing agent, storyboard, canvas).

🐛 Bug Fixes

  • Digital Human: Fixed canvas digital humans not using MiniMax H3; fixed H3 first-frame/reference video generation producing unrecognized speech when a character is silent; fixed non-speaking roles' voices not being passed to the video model in single-speaker scenes; fixed dense retries within 4 minutes on pipeline blockage, processes clogging async worker slots, length-mismatch generation failures, resolution errors with multi-clip quality degradation, and motionless hands.
  • Script Agent & Collaboration: Fixed shared script-agent links (carrying user_id) preventing other users from sending messages; fixed loose permission checks in multi-user collaboration that trusted the user_id in URL params; fixed being unable to change a role's voice after borrowing it from another user's world; fixed "character swap" cases in role behavior descriptions; friendlier prompts when splitting fails on scripts that are too short; lip-sync mode is no longer the default when splitting scripts.
  • Storyboard: Fixed the grid-check scheme's excessive false-positive rate and endless stuck grid-image generation (now degrades to direct image generation on repeated failure); fixed a storyboard render.js <path> attribute error; fixed shot previews playing TTS audio when the clip source is the original sound; fixed the prompt under the chat box not syncing after video-prompt edits (synced on blur).
  • Canvas: Fixed canvas scroll-selection preferences not being remembered and repeatedly asking the user; fixed the shortcut/scroll-wheel selection popup reappearing after a choice was made.
  • Upload: Fixed missing error messages when uploading oversized role/prop/scene images; fixed uploads of role images wrongly triggering regeneration and overwriting the user's image.
  • Workflow: Fixed workflow-page requests missing the auth token and backend 401 responses lacking detail.
  • Providers: Fixed the RunningHub slot leak (slots not released after task cancellation, permanently occupying concurrency) and erroneous permanent occupation of RunningHub pool slots; fixed voice-emotion control being unsupported in the commercial Lite edition.
  • Platform: Fixed official-website pipeline steps getting stuck forever, grid-image polling retrying 401s infinitely, and GPT Image reference 404/410s being misreported as "service abnormal"; fixed deleting roles/scenes/props in the staging area; fixed CDN audio being undownloadable in dev and the misleading sign-in admin page; fixed a batch of issues found in the 10.20 test pass (including timeline export-package/Jianying-draft 401 failures for non-admin users).

⚙️ Stability & Internal Improvements

  • Process Model Upgrade: The process pool switched from fork to spawn, eliminating same-root-cause risks such as logging lock deadlocks, EMFILE pipe leaks, and fake Qiniu thread-pool timeouts.
  • Testing Infrastructure: Fixed frequently failing GitLab e2e tests.
  • UX Polish: Optimized the storyboard bottom-left chat box styling and made grid-mode timeline buttons more prominent; improved canvas audio-node and video-to-script node usability; role reference images changed from three-view to headless images; digital-human generation no longer uses video tail frames and generates concurrently instead.

附:2.5.0 已发布原始记录(60 条)

飞书编号 标题 创建日期 分类
f_0854 进程池切换为spawn 2026-09-14 内部
f_0888 画布中数字人不是minimaxh3的数字人,而是h3数字人 2026-09-18 修复
f_0928 实现输入视频,提取剧本 2026-09-27 新功能
f_0934 官网发现 1. pipeline 步骤永久卡死 2. 宫格生图轮询 401 无限重试 3. GPT Image 参考图 404/410 被误报成“服务异常” 2026-09-29 修复
f_0944 数字人需要换一个 更快的方式,支持音色选择 2026-09-30 新功能
f_0949 角色参考图 三视角修改为去头图片 2026-10-05 改进
f_0970 调整创建角色的提示词,增加对话 角色的 说话档案(人设) 2026-10-10 新功能
f_0971 对接新的音频接口 ,生成完毕后,调用minimax h3生成数字人视频 2026-10-11 新功能
f_0974 新用户默认选择性价比模型(剧本智能体,营销智能体,以及故事板,画布) 2026-10-11 改进
f_0975 暂存区 删除角色、场景、道具 有问题 2026-10-11 修复
f_0977 宫格检查 方案优化误杀率过高的问题 2026-10-11 修复
f_0978 修复工作流页面 大量请求没有带上 认证token,并修复后端接口,401没有detail的问题 2026-10-11 修复
f_0982 抖音图文转视频 2026-10-12 新功能
f_0983 用户将它的 剧本智能体链接发给其他用户(带上自身的user_id),其他用户打开后,发现消息无法发生。 2026-10-12 修复
f_0985 工作流页面和故事板可以支持相互切换 2026-10-12 新功能
f_0989 故事板和工作流中,如果开启的是多人共享模式,那么需要新增 只看自己创建的选项 2026-10-13 新功能
f_0990 用户上传角色图片、道具图片、场景图片时,发生图片过大,却没有提示报错信息的问题。 用户上传了一个9.8mb的图片,系统只是显示图片上传失败 2026-10-13 修复
f_0991 发现用户上传角色图时,会触发 错误地重新生成角色图,导致用户上传的图片被覆盖 2026-10-13 修复
f_0992 gitlab e2e测试经常挂 2026-10-13 内部
f_0993 优化故事板 左下角对话框样式 2026-10-13 改进
f_1002 1.现有的声音库和智能体打通 2026-10-14 新功能
f_1004 RunningHub 槽位泄漏(真实的资源泄漏)。 调度循环在「已领取槽位、尚未提交上游」的窗口内如果任务被取消,提交闸门会正确拒绝调用上游,但已领用的槽位不会再随任务生命周期释放,会永久占用并发容量。 2026-10-14 修复
f_1006 minimax h3首帧生视频/参考生视频 不说话时,经常会有 无法识别的说话内容 2026-10-14 修复
f_1008 出现数字人时长超过15秒时,自动调用多轮h3数字人,并进行 拼接 2026-10-14 新功能
f_1012 h3 角色新增角色音色描述,支持生成视频时,添加角色音色描述 2026-10-15 新功能
f_1013 修复 工作流故事板,h3模型 只有一个角色说话,确保非说话的角色的音色传入视频模型 2026-10-15 修复
f_1014 修复runninghub 池会出现错误永久占用的问题 2026-10-15 修复
f_1015 修复故事板 render.js:4211 Error: attribute d: Expected number, "….51 0 1.1-.9 2-2s.9-2 2-2c.18 0 …". 报错 2026-10-15 修复
f_1017 修复商业版lite里 不支持音色情感控制的问题 2026-10-15 修复
f_1018 针对剧本过短情况,拆分失败的友好化提示 2026-10-15 改进
f_1019 超长时间的数字人 从wan2.2切换为minimax h3 拼接的方式 2026-10-15 改进
f_1024 剧本拆分师支持更新剧本权限 2026-10-16 新功能
f_1025 需要支持每日 哪些用户消耗多少算力的统计和表格功能 2026-10-16 新功能
f_1026 开发环境下,发现cdn的音频无法下载 2026-10-17 修复
f_1032 S1 分镜的视频来源是 视频原声,但是它在播放时,预览的时候,放出来的声音是TTS的 2026-10-17 修复
f_1036 用户在画布中滚动的选择,似乎没有没有记忆,每次进去,每隔一段时间,这个记忆就消失了,导致又需要询问用户 2026-10-18 修复
f_1037 画布导出需要支持导出类似故事板那样的压缩包 2026-10-18 新功能
f_1039 修复签到管理页面 容易让用户误解 2026-10-18 修复
f_1045 剧本拆分时,默认不再使用对口型模式 2026-10-19 改进
f_1046 故事版里,当用户修改了视频提示词后,onblur时需要 同步更改 对话框下方的提示词 2026-10-19 修复
f_1048 故事板在grid模式下,时间轴的按钮需要显眼点 2026-10-19 改进
f_1049 修复剧本智能体中 角色音色 一旦用别人的世界的 角色,就无法修改音色的问题 2026-10-19 修复
f_1052 画布快捷键 弹窗,哪怕之前填过,第二次进入页面,也会弹出 2026-10-20 修复
f_1053 故事板需要支持查看 仅自己的,或者指定某个用户的 作品 2026-10-20 新功能
f_1057 故事板如果一直宫格图生成失败,就降级为 直接生图,而不要一直卡住让用户无法生成 2026-10-20 修复
f_1058 支持管理员 删除 或者隐藏其他世界 2026-10-20 新功能
f_1061 修复生成数字人时,遇到堵塞时,会在4分钟之内密集尝试的问题 2026-10-20 修复
f_1062 修复画布页面,频繁出现滚轮快捷键选择弹窗,即便选择过之后,下次还是会弹出 2026-10-20 修复
f_1063 管理员可以删除 所有 剧本智能体中的世界 2026-10-20 新功能
f_1064 剧本智能体 多人协作模式下,判断用户是否有权限的判断逻辑不严,直接把url参数中的user_id视为真的用户id 2026-10-20 修复
f_1065 数字人问题,分辨率错误,多个视频片段后出现 视频劣化 2026-10-20 修复
f_1066 修复10.20测试发现的问题 2026-10-20 修复
f_1070 提取剧本优化 2026-10-21 改进
f_1071 优化画布音频生成节点功能和易用性 2026-10-21 改进
f_1072 数字人的手不会动 2026-10-21 修复
f_1075 工作流页面中,在提示词的框中,输入@后 ,出现的 下拉框,需要在下拉框右侧,有详细的预览图,类似 @storyboard.html 2026-10-21 新功能
f_1077 数字人放弃使用视频尾帧,而是直接并发生成 2026-10-21 改进
f_1078 修复数字人长度不一致,导致数字人无法生成的问题 2026-10-21 修复
f_1079 修复角色的行为描述中存在换人测试这种情况 2026-10-21 修复
f_1083 修复数字人进程 会把异步进程的卡槽 堵死的问题 2026-10-21 修复