Aller au contenu

Covers and thumbnails

Cette page n’est pas encore traduite : voici la version anglaise.

You get a cover for every platform you post to, at that platform’s exact size, with a preview of how its feed crops it. The same cover can replace the dead first second of your video, so the thumbnail and frame 0 match.

  • The cut video, and optionally a face photo or a few slide images.
  • A headline. The cover headline works best as the same sentence as the hook of your script.
  • ./install.sh done: it downloads the fonts and the MediaPipe face and segmentation models the cover tools use.
  • Chrome, Chromium or Edge for the HTML templates (or pip install playwright && playwright install chromium).
Layout Use it when What it looks like
Collage You want to show the range of the content; the face isn’t the point Four stills in a diagonal collage, headline and tags
Face on quadrants Personal brand: your face drives the click Your cut-out face over a 2×2 grid of slides
Split cover A horizontal post (小红书 4:3, YouTube 16:9) with a quote or a “tested it myself” angle Retouched photo on the left, a dark panel with quote, title, thumbnail and chips on the right
Notes cover Notes-style talking heads and course clips A 16:9 notes board, plus a 4:3 centre crop
Framed cover Screen-heavy content Dark plate, big text, a tilted framed screenshot

A new channel usually gets more from a face. An established one can lean on the collage.

  1. Covers come with most video projects: talking-head and vlog projects stop at Pick the cover in your Inbox with ranked frame candidates and the cover text.
  2. For a cover on its own, describe it on Home: which video, which platforms, which layout.
  3. Pick a frame and edit the text. Each platform gets its own cover file next to its video, and the feed-crop preview shows what survives in the grid.

Vlog, talking-head and photo-story projects design their covers from one picked frame, per platform:

  • Frame picking. Frames are ranked by sharpness, exposure, a clear subject, the face expression when there are faces, and an aesthetic score (optional download). Frames inside a crossfade are skipped.
  • Subject cutout. BiRefNet separates the subject from the background (MediaPipe when it isn’t downloaded), so the title can stay clear of it, or sit just behind it in the subject variant.
  • Upscale. A crop smaller than the cover size goes through Real-ESRGAN instead of a plain resize.
  • Title placement. The title goes in the emptiest part of the safe area, clear of faces, the subject and the platform’s own UI: YouTube’s duration badge, Bilibili’s play-count band, Xiaohongshu’s video mark. No stroke, no box; a soft local shade only as dark as the contrast needs.
  • One language per platform. YouTube and Shorts covers are in English; Bilibili and Xiaohongshu covers in Chinese. A bilingual title is split, never stacked.
  • Three YouTube variants for A/B testing: the full scene, a tighter shot on the subject, and a text-led version. The top-ranked one is the cover; the other two are saved as youtube-horizontal.cover.ab1..3.jpg, with a note in the YouTube post to use them in YouTube Studio’s Test & Compare.
  • Small-size check. Every variant is rendered at YouTube’s small thumbnail sizes (168×94 and 320×180); a variant whose headline is too small or low in contrast there is never the chosen cover while a legible one exists.

Ask for a cover and the skill follows workflows/cover:

  • Frame picking. For a talking head, Claude ranks frames by smile, open eyes and a centred face (cover.score_frames) and shows you the top picks on a contact sheet. A talking frame (mid-word, eyes on camera) reads better than an idle, posed one.

  • Retouch before the cover is built: slimming, eye size and a light makeup preset.

    Terminal window
    python3 -m vstudio.retouch face.png face_retouched.png --slim .05 --eye .04 --preset natural

    Presets are none, natural, daily and glam. --faces all handles group shots.

  • Cut-out face for the quadrant layout, with a halo check against a checkerboard.

  • Render at every platform size in one go:

    Terminal window
    python3 workflows/cover/scripts/render_cover.py work/cover.html -o work/cover.png \
    --platform xiaohongshu --platform douyin --platform youtube

workflows/polish puts the cover on the first second of picture. The audio is untouched and the length doesn’t change:

Terminal window
python3 workflows/polish/scripts/polish.py export.mp4 -o final.mp4 --cover cover.png --check

--check writes the first frame as a PNG so you can confirm frame 0 is the cover. If your first second already carries a hook shot, pick a frame from the hook as the cover instead.

Platform Cover What the feed shows
YouTube 1280×720, up to 2 MB the whole cover; keep text left of the timestamp
YouTube Shorts 1080×1920 the whole cover
TikTok, 抖音 1080×1920 profile grid: the centre 3:4
小红书 (3:4 or 9:16 post) 1080×1440 the whole cover
小红书 horizontal 1920×1080 the centre 4:3
B站 1146×717 (16:10) also cropped to 4:3 and 16:9

Covers for the other platforms are in the platform reference. Wherever a feed crops the cover, the export writes a .feed.jpg preview and warns when the headline is probably cut.

Option Default What it does
Cover accent teal #2dd4bf Colour of the collage and quadrant templates (persona cover.accent). The split cover follows your brand colours instead.
Retouch preset natural none, natural, daily, glam.
Matting engine MediaPipe --engine rvm gives cleaner hair; it is GPL-3.0 and downloaded only if you choose it.
Cover length on video 1.0 s How much of the opening picture the cover replaces (--cover-sec).
  • “Make a cover for this video for Xiaohongshu and Douyin. Use my face, pick a frame where I’m mid-sentence and looking at the camera.”
  • “Collage cover with four different moments from the talk. Headline: ‘RAG isn’t the hard part.’”
  • “Split cover for a horizontal Xiaohongshu post: retouched photo on the left, the quote on the right, a ‘tested’ stamp.”
  • “Put the cover on the first frame and give me a YouTube thumbnail under 2 MB.”
  • “Give me three YouTube thumbnails to A/B test.”