The four approaches solve different parts of the job.
1. **Generate a shot.** A video model creates motion from a text prompt. This can supply a scene that would be hard to film, but people, objects, actions, and continuity need close review.
2. **Animate a still.** Start with an approved image and ask for motion. This preserves more of the starting composition, while motion can still distort hands, tools, faces, text, or object identity.
3. **Assemble existing clips.** Record or license short clips and edit them into the invitation. The assets are concrete and reusable, but coverage, consent, continuity, and license scope require planning.
4. **Create a narrated sequence of stills or slides.** Time a small set of owned images, graphic cards, or photographs to narration. Motion can be limited to pans, zooms, and transitions. This trades cinematic motion for precise timing, readable information, and easy correction.
Generating footage is only acquisition. Finishing is the separate work of selecting shots, arranging them to exact duration, recording or generating voice, mixing audio, adding and correcting captions, placing exact event information, checking safe areas and legibility, and exporting the required format. A visually attractive generated clip is not a finished invitation.
### What the checked sources establish
- Runway's current Gen-4.5 documentation says the model supports text-to-video and image-to-video, with generated durations from 2 to 10 seconds, 720p output, several aspect ratios, and credit use per second. It describes downloading, upscaling, and iteration, and offers ProRes or PNG sequence outputs only on specified higher plans with added credits. That documents generation capabilities and plan conditions, not production quality. [Source VD1]
- Adobe Express documents automatic caption generation from video or audio, followed by transcript review, text correction, caption styling, repositioning, and download. This is finishing functionality; the documentation itself tells the user to review and edit the transcript. [Source VD2]
- Blackmagic Design describes DaVinci Resolve as an editing, effects, color, audio, and delivery environment. Its current product page says the free version supports common 8-bit formats up to 3840 × 2160 and 60 fps, while the Edit documentation describes subtitles that can be rendered into the video or exported separately. [Source VD3]
- Runway's current terms say it does not claim ownership of user inputs or outputs and does not restrict commercial use of outputs subject to compliance. The terms also state that inputs and outputs may be used to train and improve its systems, grant the company a broad license for those purposes, make the user responsible for rights in uploaded content, and warn that outputs may not be unique. [Source VD4]
### My comparative judgment, not a product test
| Criterion | Generated shot | Animated still | Existing clips | Narrated stills/slides |
|---|---|---|---|---|
| Message clarity | A single expressive image may attract attention, but exact event facts belong in finishing. | The approved composition can anchor the message; motion should stay secondary. | Real repair actions can explain the event quickly if the shots are legible. | Strong for a short, explicit sequence with controlled text and voice. |
| Continuity | Hardest across multiple generations, especially for hands, tools, clothing, and space. | Better visual anchor, but motion may deform the source. | Achievable with planned coverage and one location; mismatched stock can feel fragmented. | Deliberate continuity through repeated layout, palette, and typography. |
| Timing | Generation duration does not guarantee action timing or room for text. | A simple motion can be retimed, though excessive slowing may look poor. | Editors can trim shots, but an action may not fit the available seconds. | Each beat can be assigned exact frames and revised independently. |
| Correction effort | A small defect may require another generation or a complex repair. | Regenerate motion or return to the still; keep the original. | Replace or trim one clip when alternatives exist. | Replace one still, line, or card without rebuilding the rest. |
| Reusable assets | Prompts and selected clips can be reused, but identity consistency is uncertain. | The approved still is reusable for posters and social posts. | Raw clips can support longer edits if releases and licenses permit it. | Stills, narration, caption file, and layout can be reused separately. |
| Privacy | Prompts and uploaded references go to a cloud service. | The still may expose a person, location, or metadata when uploaded. | Filming community members requires informed consent and careful handling of bystanders. | Can use illustrations and fictional details; cloud narration or editors add their own data path. |
| Rights | Check input rights, output terms, likenesses, music, and platform rules. | Need rights to the still and any depicted likeness. | Releases and stock-license scope must cover the intended channels. | Each image, font, sound, and voice needs a recorded source and permission. |
| Total cost/effort | Credits plus retries, selection, and finishing. | Still creation plus motion generation and finishing. | Planning, capture/licensing, editing, and releases. | Asset creation, narration, layout, captions, and export; fewer motion surprises. |
The table is a planning hypothesis. It does not predict which path will be faster or clearer for one team's tools, skills, and audience.
### A small two-path experiment
Use one 12-second, vertical 9:16 brief for both paths. Keep all event details fictional until publication approval.
**Shared three-beat brief**
- **Beat 1, 0:00–0:04 — recognition:** Close view of a broken red toaster on a shared table. Two neighbors' hands point to the loose lever. On-screen title added in finishing: “Broken? Bring it.”
- **Beat 2, 0:04–0:08 — collaboration:** Wider view of four neighbors repairing a lamp, backpack, bicycle wheel, and toaster together. The path around the table remains clear. On-screen title: “Learn and repair together.”
- **Beat 3, 0:08–0:12 — invitation:** Simple cream graphic card with a small tool illustration. Exact editable text: “Riverside Repair Café · Sat Oct 17 · 10:30–1:45 · Free.” Final spoken line: “Bring one small item and learn with your neighbors.”
Do not ask a video generator to render the event text. Add the titles, date, and captions during finishing, where spelling and timing can be corrected directly.
**Path A — generated motion:** Create or select an owned still for Beat 1 and animate it for four seconds; generate or animate a separate four-second collaboration shot for Beat 2; build Beat 3 as an editable graphic card. Assemble all three in an editor, add the same approved human-recorded narration, create captions, and export.
**Path B — controlled asset sequence:** Photograph or illustrate the same toaster detail and collaboration tableau with consented participants or fully illustrated figures. Use only crop, pan, zoom, and crossfade motion. Build the same Beat 3 card, add the same narration and captions, and export with the same frame size, frame rate, and loudness target.
For both paths, preserve the original assets, editable timeline, narration master, caption text or sidecar file, and final export. Then compare:
- whether viewers can state what the event is, who it is for, and when it happens after one viewing;
- whether the toaster, hands, tools, people, and room remain continuous within each beat;
- actual duration of each beat and whether text remains readable long enough;
- minutes, credits, generations, retakes, asset substitutions, and manual repairs;
- changes required to correct one deliberately altered detail, such as changing “Sat Oct 17” to “Sun Oct 18” in the editable card and captions;
- which source assets remain independently reusable;
- every upload, participant release, stock or font license, voice consent, and service term checked.
Use at least three viewers who did not make the video. Show the two paths in alternating order and do not reveal which uses generated motion. The result is a comparison of two specific artifacts, not a general verdict on generated video or slide-based invitations.
### Still uncertain
- Actual visual continuity, motion quality, prompt adherence, caption accuracy, and audience comprehension until both videos are rendered and viewed.
- Whether a 720p generated shot is adequate after crop, stabilization, upscaling, platform recompression, and display on the intended device.
- Exact plan access, credit conversion, export options, watermarks, and total price on the experiment date.
- Current retention, training, review, and deletion behavior for the exact service, account type, region, and settings.
- Rights to any real faces, voices, music, stock clips, fonts, logos, venue interiors, or generated output used in publication.
- The accessibility target: open captions, sidecar captions, audio description, contrast, safe-area rules, and any venue-specific requirements must be chosen for the actual channel.
### Limitations and next experiment
This run reviewed vendor documentation and terms; it did not generate or finish a video. The next experiment is the paired 12-second build above with fictional details and owned assets, followed by a frame-by-frame review, one-view comprehension check, and account-specific rights/privacy check before real community content is uploaded or published.
