ClipForge v0.9.3 — 本地素材库与可靠上传
把自己的图片与视频存进项目，整理、预览、采用到分镜，再直接合成出片。
- 本地素材库：一次最多上传 12 个文件、每个 80MB，支持图片和视频；可编辑名称与标签、搜索、筛选、排序、分页和预览。
- 可靠上传：流式传输、进度与文件检查状态、取消和未完成项重试；相同内容自动复用，保留已有名称与标签。
- 直接用于镜头：选中的素材成为当前合成输入，旧候选继续保留；单镜头上传也支持视频并自动进入素材库。
- 只用本地补齐：只读取当前项目素材，不调用模型或在线素材源；保护已选的完成素材和进行中任务，以及商品原图镜头。
- 状态与缓存修复：失败记录和未选中旧版本不再阻止补齐，页面不再误用旧候选；本地素材检索按项目隔离，名称与标签修改即时生效。
验证：124 个测试文件、1240 项用例通过，类型检查和生产构建成功；已实测上传、重复文件复用、分镜替换、仅本地补齐以及带音轨的 720p 成片合成。移动端 390px 无横向溢出。
下载与运行
- macOS Apple Silicon：
ClipForge-0.9.3-arm64.dmg - Windows x64：
ClipForge.Setup.0.9.3.exe - Linux x64：
ClipForge-0.9.3.AppImage - Agent Skills：
clipforge-skill.zip、clipforge-script-judges.zip - MCP 安装包：
clipforge-mcp-0.1.4.tgz，可直接通过npx --package <安装包下载地址> clipforge-mcp接入已有 ClipForge 实例。 - Docker：
ghcr.io/xixihhhh/clipforge:0.9.3与latest；已通过容器启动、页面及 FFmpeg/drawtext 检查并推送，两个标签对应同一镜像。
桌面安装包由 CI 自动构建，尚未逐个平台进行安装验收。macOS / Windows 包未签名；首次启动可能出现系统安全提示。已有用户数据沿用原目录。
Local material library and reliable uploads
Keep your own images and videos in each project, organize and preview them, then assign them to shots and compose.
- Batch-upload up to 12 files, 80 MB each; edit names and tags, search, filter, sort, paginate, and preview.
- Stream uploads with progress, validation, cancellation, and retry. Identical content reuses the existing file without losing edited metadata.
- Assign images or videos as the active shot input while retaining earlier takes. Per-shot uploads also enter the library.
- Fill empty shots using only the current project's library, with no model or online-stock calls. Ready or in-progress selected assets and product-image shots are protected.
- Failed attempts and inactive historical takes no longer block filling or appear as active footage. Local search caches are project-scoped and metadata edits take effect immediately.
Validation: 124 test files / 1,240 tests, type checking, and production build passed. Real upload-to-compose verification produced a 720p video with audio. The mobile library was checked at 390px without horizontal overflow.
The release includes macOS Apple Silicon, Windows x64 and Linux x64 packages, two Skill bundles, and the versioned MCP tarball. Desktop artifacts are CI-built; installation has not been tested on every platform. macOS and Windows packages are unsigned. Docker tags 0.9.3 and latest are published and resolve to the same image, after container startup, page, and FFmpeg/drawtext checks passed.