Durations, aspect ratios, and settings
Settings feel like the boring part of a generator UI, and that's exactly why they're worth six minutes: most wasted credits die on settings chosen by habit rather than intent.
5 seconds or 8?
The question isn't "how much video do I want" — it's how many beats does this shot contain.
- 5 seconds — one beat: a punch, a pour, a stare, a reveal-and-hold. Single-beat shots padded to 8 seconds don't get better; the model fills the extra frames with drift, and drift is where subjects warp. The boxer and dog side-eye prompts are deliberately 5s.
- 8 seconds — a develop-and-payoff structure: a tracking walk, a drone reveal, a two-line exchange, a light sweep that completes. If you can name what changes between second 2 and second 6, you've earned the 8.
When in doubt, take 5. Shorter clips also mean cheaper iterations while you're still tuning the prompt — burn the 8-second budget on keeper takes, not experiments.
Aspect ratio: generate native, never crop
Each ratio is a different composition, not a different crop of the same one. The model composes for the frame you request — a 9:16 request puts the subject in a vertical-friendly framing from the first frame. Cropping a 16:9 render into a 9:16 post throws away both edges of a composition designed to use them.
| Ratio | Where it lives | Composition notes |
|---|---|---|
| 16:9 | YouTube, web embeds, presentations | The cinematic default; camera moves read best here |
| 9:16 | TikTok, Reels, Shorts, Stories | Frame subjects waist-up; lock the camera or use handheld micro-movement |
| 1:1 | Feed posts, product tiles, memes | Center-weighted; great for lock-off product shots and reaction clips |
One production habit worth stealing: if a concept must ship in multiple ratios, generate it separately in each ratio with the same prompt. Two generations cost less than one re-shoot after the vertical crop kills your composition.
Resolution and upscaling
1080p covers every feed and most client work. The upscale question only matters for large-screen delivery — and upscaling is a finishing step, applied to the final edit, not to every raw take. Generate at the native resolution, select takes, edit, then upscale once.
Thinking in credits
Whatever platform you use — HappyHorse online, WeShop AI, or the API — generation costs scale with duration and resolution. A sane budget split for a finished deliverable looks like:
- ~60% exploration — 5-second, default-res takes while the prompt stabilizes
- ~30% coverage — 2–3 variants of the locked prompt at delivery settings
- ~10% finishing — the final take, plus upscale if needed
The anti-pattern is inverting the pyramid: one maxed-out 8-second 1080p roll per prompt idea, judged, discarded, repeated. That's paying keeper-take prices for exploration rolls. Generation speed (~38 seconds per clip) makes iteration fast — settings discipline is what makes it cheap.
The pre-flight checklist
- Beats counted → duration chosen (5 unless the shot develops)
- Target platform known → ratio generated natively
- Still exploring → default res, 5s, iterate
- Keeper take → delivery settings, then stop rolling
Settings sorted. If you're ready to move past the web editors entirely, the last stop is running HappyHorse via API and open source.