Skip to content

v0.25.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 30 Aug 02:07
· 5 commits to main since this release
v0.25.0
102ccbd

HotClip v0.25.0 — 分析看对画面，封面挑对一帧

这一版把“AI 分析看到的画面”与“最终导出的画面”彻底对齐，并把普通抓帧封面从单一响度峰升级为最终成片质量排序。全程复用安装包自带 FFmpeg，不新增模型、依赖、下载、上传或云端费用。

  • 全分析链路同轨：运动/镜头/亮度统计、TransNetV2、表情抽帧、VLM 接触表、候选画面复核、工作台胶片带、去录屏 UI 与人脸取景，全部绑定探测后最终渲染选中的同一条全局视频轨。
  • HDR 分析预览：只有元数据完整且已经真实验证可执行的 PQ/HLG 路径，才在缩放、灰度、人脸或视觉模型分析前复用安全 SDR 预览；普通 SDR、不完整或不支持的 HDR 保持既有 fail-open 行为，不猜色彩路径。
  • 缓存不串画面：Tier-0 信号、镜头边界与视觉扫描的能力身份包含选中视频轨、分析色彩计划和版本。换轨或色彩决定变化会精准重算，不会复用另一张画面的旧证据。
  • 质量排序智能封面：响度峰负责提名有内容的时刻，均匀时刻保留盲区兜底；随后在最终质检/修复后的成片上，本地比较清晰度、信息量、曝光、明暗跨度与 240ms 邻域稳定性。
  • 风险帧先淘汰：黑场、白场与低信息平帧不会赢；模糊或转场不稳的画面相对降权。一片多版按排名取下一张，避免所有版本塌成同一封面。
  • 精确回退与可审计：FFmpeg 指标不可用或没有安全候选时，精确回到此前的响度峰时刻，不影响视频交付；clips.json 记录选择模式、候选数、淘汰数、时间与得分。
  • 真实媒体回归：双视频轨样片证明默认选中的蓝色 PQ 画面经 SDR 预览进入接触表，而显式另一轨保持红色 SDR；四场景封面黄金集证明黑/白场被拒绝、清晰图案稳定胜过同源高斯模糊帧。

HotClip v0.25.0 — Analyze the right picture, publish the right frame

This release aligns every picture-analysis path with the exact video stream used for final render, and upgrades ordinary frame-grab covers from a single loudness pick to deterministic final-render quality ranking. It reuses the bundled FFmpeg with no new model, dependency, download, upload or cloud cost.

  • Selected-picture parity: motion/shot/luma statistics, TransNetV2, facial-emotion frames, VLM contact sheets, candidate review, workbench filmstrips, UI-band detection and face reframing all bind the globally selected final-render video stream.
  • HDR analysis preview: only complete, runtime-verified executable PQ/HLG paths receive the safe SDR preview before resize, grayscale, face or vision analysis. SDR and incomplete/unsupported HDR retain established fail-open behavior.
  • Cache isolation: Tier-0, shot and vision evidence identities include the selected stream, analysis colour plan and version, so a picture or colour-decision change rebuilds only the affected evidence.
  • Quality-ranked smart covers: audio peaks propose content-relevant moments while a uniform reserve covers blind spots; the final post-QA/repair clip is then ranked locally for sharpness, information, exposure, tonal range and 240ms neighborhood stability.
  • Unsafe frames lose first: black, white and low-information flat frames are rejected; blur and unstable transitions rank lower. Multi-version variants take the next ranked frame instead of collapsing onto one cover.
  • Exact fallback and receipt: unavailable metrics or no safe candidate returns exactly to the previous audio-peak timestamp without blocking delivery. clips.json records mode, candidate/rejection counts, selected time and score.
  • Real-media regression: a dual-video fixture proves the default blue PQ picture reaches the SDR contact sheet while the explicitly selected red SDR stream remains distinct; a four-scene cover golden set proves black/white rejection and sharp-over-blurred ranking.
Assets 11
Loading