ClipForge v0.8.99 — 可控生成 / Adaptive video conditioning
更稳的人物、商品与镜头连续性
这一版让每个镜头在生成前先编译真实可执行的条件计划，再按供应商和模型能力自动选择最合适的生成方式。
- 关键帧、尾帧、人物定妆、商品图、前镜连续性、动作和声音参考统一进入镜头条件计划。
- 支持多参考的模型会接收人物、商品和连续性锚点；只支持首尾帧的模型会自动保留确定性更高的时序约束。
- 支持原生音频的模型可在生成画面时同步完成逐字对白、口型、动作、旁白、环境声和物体声；其余模型继续使用稳定的后期配音链路。
- 模型预检新增多图参考、参考视频、声音参考与原生音频能力提示，并在超出配额或能力不足时明确显示降级结果。
- 每个云任务和候选素材都会保存非敏感生成策略摘要，刷新后仍可恢复上一镜真实尾帧，生产质量台可查看锚点数量、音频路径和能力降级。
- 质量诊断后的重做会继续携带项目导演意图，不再只发送局部动作提示；任何付费重做或换模型仍不会自动执行。
- README、中文 / 英文官网、FAQ、使用手册和内置 Agent Skill 已同步更新。
验证：112 个测试文件、1172 项测试全部通过；TypeScript、数据库迁移、lint、Next.js 生产构建和手机横竖屏响应式实测通过。
下载
- macOS Apple Silicon：
.dmg（未签名，首次打开请右键“打开”） - Windows x64：
.exe（未签名，SmartScreen 可选择“更多信息 → 仍要运行”） - Linux x64：
.AppImage（下载后执行chmod +x） - Agent Skill：
clipforge-skill.zip
More reliable identity, product fidelity, and shot continuity
Each shot now compiles an executable conditioning plan before generation and adapts it to the selected provider and model.
- Keyframes, end frames, character sheets, product images, cross-shot continuity, motion, and audio references share one provider-neutral control plan.
- Multi-reference models receive identity, product, and continuity anchors. Models limited to start/end frames preserve the strongest deterministic temporal constraints.
- Native-audio models can generate exact dialogue, lip sync, performance, narration, ambience, and object sounds together with the image sequence. Other models keep the stable post-production voice path.
- Model preflight now reports multi-image, reference-video, reference-audio, and native-audio capabilities, including visible quota or compatibility fallbacks.
- Every cloud task and candidate stores a compact non-sensitive strategy summary. Persisted tail frames restore continuity after refresh, while the production quality gate shows anchor count, audio path, and fallbacks.
- Diagnosis retakes retain the project direction, and no paid regeneration or model switch runs without explicit confirmation.
- Chinese and English READMEs, website pages, FAQ, guides, and the bundled Agent Skill are updated.
Verified with 112 test files / 1,172 passing tests, TypeScript, migration checks, lint, a Next.js production build, and responsive browser QA.