Footage and screens
1. Real footage, not generated
Use real stock footage of real people unless the person asks for generated video. Viewers spot generated faces, and a film about real problems loses its footing when they do.
Where. Pexels (Pexels Licence: free for commercial use, no attribution) through its official API:
scripts/stock_search.py videos "<query>" out_dir saves thumbnails and a list with each clip's largest file,
usually 4K. It needs a free PEXELS_API_KEY. Other libraries the person has an account with are fine too.
Record the id of everything you download.
How to search. Search for the action and the mood, not the plot: "man worried laptop", "woman coffee laptop cafe", "couple planning table". Download thumbnails, tile them into a contact sheet, and choose by looking.
One person, several clips. A character needs more than one shot: calm for the introduction, worried for the problem, at ease for the ending. Find a shoot, not a clip. On a clip's page, the same creator's neighbouring ids are usually the same actor in the same clothes and room. Two or three clips of one person are worth more than ten clips of ten people.
Cast before naming. Look at who the footage actually shows, then choose names, jobs and cities that fit. Mixed casts are fine. Do not put a name on a face it plainly does not suit.
Reject clips with visible brand logos you cannot crop or blur, readable screens of other products, licence plates, text in the wrong language for the story, or a mood that contradicts the beat (a smiling face in the problem).
2. Choosing the seconds
Make a contact sheet of each candidate at one frame a second (scripts/sheet.sh). Note, per clip, the seconds
that are calm, the seconds that are worried, and any second with a logo or a camera bump. A slide or a reveal
at the start of a clip is usually unusable.
If one take must cover a long line, cut it in two with a punch-in: the same clip shown again from a later
second at 1.2 × size. If a take is too short, slow it: down to 80% plays fine; below that, interpolate frames
with ffmpeg minterpolate (mi_mode=mci) and look at the result for warping before using it.
3. Cutting and grading
Write every cut into footage.tsv (name, source file, start, duration) and produce the clips with
scripts/footage.sh footage.tsv grade.vf assets/footage 3840x2160. One pass does the cut, the scale and the
grade, straight from the source file.
- Keep the manifest. It is the only record of where a cut came from. If it is lost,
scripts/match.pycan find an offset again by comparing frames. - Cut at the resolution you will deliver. Upscaled 1080p footage in a 4K film is soft next to sharp text.
- Keep the source downloads the film uses; delete the rest once the cast is settled.
One grade for everything. Clips from five shoots look like five films until they share a grade. grade.vf
is a neutral starting recipe: slightly lower saturation, lifted blacks, held whites, fine grain, a light
vignette. To tie footage to a brand, add a colorbalance step before noise that leans the shadows toward the
brand's dark colour and the highlights toward its light one, a few percent at most, for example
colorbalance=rs=0.045:gs=-0.03:bs=0.06:rh=0.055:gh=0.025:bh=-0.045 for plum shadows and cream highlights.
Look at a full-size crop of a face after grading: skin must still look like skin.
A soft-focus copy of a clip (scale to 480 px, blur, scale back) makes a ground for text.
4. Photos and props
Photos used as props (the thing a person wants, a place) come from the same sources under the same licence
rules. Choose pictures without recognisable faces: a wedding shown as hands and rings, not a couple who are not
your character. Blur a maker's name with ffmpeg delogo. Crop all props to one aspect ratio so they can share
one frame design.
5. The real product
What is shown of the product is the real product. Never redraw an interface, and never pass stock off as it.
| Product | What to capture |
|---|---|
| Website | the live site in a desktop viewport: the hero, key sections, and full-page scrolling captures to pan across |
| Web app or SaaS | each state of the task, signed in to a demo account with believable data |
| Mobile app | the same, at a phone viewport, shown inside a phone frame |
| Physical product or shop | the company's own photos and footage; the product page; ask for files if the site's are small |
| Service | the booking or sign-up flow, and real results the company is allowed to show |
How. Use a headless browser (Playwright) against the live site, or the product run locally in its demo or mock mode, at a device scale factor of 3 so screens stay sharp when enlarged. Capture every state the script walks through, in order, and name the files by step. If you have no access to a signed-in product, say so and ask for a demo login or for screenshots; do not invent the screens.
The demo account. Give it the story's numbers, so the screen says what the narrator says. If the product computes results (dates, amounts), use what it actually computes for those inputs, even if the story would be neater with rounder numbers.
Do not change the product. If the demo data must change, change a throwaway copy of the mock or fixture layer and keep the diff and the capture script in the project. Never edit interface code to make a better picture, and never commit to their repository.
Pin what you use. Copy the captures the film uses into assets/screens/. Tools and reruns overwrite
working folders.
Pulled-out pieces. The strongest way to show a screen is the phone at one side with one piece of that screen (a card, a number, a sentence) enlarged beside it. Capture that element on its own at high resolution; a crop of the full screen goes soft when enlarged, and softer still at 4K.
Say what is not live. Check which branch or environment your captures came from. If a screen in the film is
not in production, write it in REVIEW.md and tell the person.
6. The brand
Take fonts and colours from the product and the website themselves, not from memory: read the site's font
imports and CSS variables, and the app's. They can differ (a marketing site in one pair of faces, the app in
another); ask which the film should follow if it is not obvious, and let screenshots keep their own type.
Download the font files (Google Fonts serves woff2 per subset) into assets/fonts/ and note the licence.
Check that the faces have the glyphs the film needs: currency signs (₹, €), arrows, the multiplication and minus signs. Where a glyph is missing, set it in a fallback face.
7. The QR code
scripts/qr.py makes an SVG with 30% error correction so a logo can sit in the middle. Put UTM tags on the link
(utm_source, utm_medium=qr, utm_campaign) so visits from the film can be counted, and show the short
address as text beside the code. Check that the address answers before you use it. After the end card is built,
decode the code from a rendered frame with scripts/scan.py.
8. Credits
Add each source to CREDITS.md as you take it: file, id or URL, creator if known, licence, where it is used.
Do it at the moment of download; reconstructing it later costs more and misses things.