Generate and iterate

Draft every shot at low resolution or on a cheap tier, change one thing per take, and render at full quality only once you have a keeper. Most of a video's cost is retries, not the final clip, so the way you iterate decides your budget.

Draft cheap, finish onceCopy Anchor Link

Video models charge per second and per resolution. A 480p draft tells you whether the motion, framing and character work. That is all you need to know before you pay for 1080p.

Prices checked on 15 Oct 2026. API rates are official; check your own plan's credit table, because app prices differ.

A creator on X who prompts for a living posted this comparison on 3 Oct 2026 for one 10 second shot that needed 7 tries:

• 1080p every time: 39.80 USD
• 480p for all 7 tries, then one upscale of the keeper: 8.59 USD

The same post roughs out shots on Seedance 2.0 Mini (0.36 USD per 10 s at 480p) before finishing on Seedance 2.5 (1.03 USD per 10 s at 480p). These are one person's numbers, but the method matches what every cost guide we read recommends.

Change one variable per takeCopy Anchor Link

When a take is close but wrong, keep everything that works and change only the part that does not. If you change the camera, the action and the lighting at once, you cannot tell which change helped.

Same as previous take. Keep framing, character, lighting and audio identical.
Change only: camera move from "orbit" to "static, locked-off".

• Pick one variable: camera move, action, lighting, reference strength or duration.
• Say exactly what changes and that everything else stays the same.
• Turn native audio off on the first passes when your tool allows it. One less thing to judge.
• Do not fire three generations at once to test one change. The same creator priced that at 17 USD for a single change at 1080p.
• For a small error in a good take, try an edit model before you regenerate. Google Flow charges 40 credits for an Omni Flash edit (checked 15 Oct 2026). Write short edit prompts and add "Keep everything else the same".

If a shot keeps failing, strip it down: lock the camera, cut the action to one verb, clear the background. Then add things back one at a time.

How many takes to expectCopy Anchor Link

Nobody gets every shot on the first try. Reported numbers vary a lot with how hard the shot is.

Plan around 5 to 10 takes for a typical single character shot. Budget fewer when you have strong reference images, more for action, hands and several people.

To estimate a video, multiply shots by takes by the cost per take. Six shots at six takes each on Veo 3.1 Fast (0.96 USD per 8 s clip at 1080p) is 36 clips, about 35 USD. Draft the same takes on Veo 3.1 Lite at 720p (0.40 USD per 8 s) and only render keepers at 1080p, and the bill drops to a fraction of that. Measure cost per usable second, not cost per generation.

Extend and chain shotsCopy Anchor Link

Most models cap a single clip at 8 to 15 seconds, and quality drifts past about 8 to 10 seconds. To go longer, either extend a clip or chain shots with first and last frames.

ExtendCopy Anchor Link

Extensions follow your prompt, not an exact frame, so the join can show. Plan a pause or a cut point where the extension starts. Veo's docs also warn that a voice may not carry over well if nobody speaks in the last second of the source clip.

Chain with first and last framesCopy Anchor Link

First and last frame mode exists on Veo 3.1, Seedance 2.x, Kling 3.0 and Wan. Kling 4.0 Flash does not support it (checked 15 Oct 2026).

Keep a take logCopy Anchor Link

A take log tells you what you already tried, what each shot really cost and which prompt made the keeper. Keep one row per generation in a spreadsheet.

shot | model/tier | res | dur | seed | ref strength | prompt version | change vs prev | credits | verdict (keep / fix:<what> / kill)

• Save the output file name with the row, so you can find the take later.
• Note the seed when the tool shows one, and lock it when a take has good motion. Most closed models do not promise the same result from the same seed.
• Copy winning phrases into a style file and reuse them word for word in later shots.

When to give up on a shotCopy Anchor Link

Set a limit before you start. A simple rule from the research:

• Draft up to 3 takes at 480p or 720p on a cheap tier.
• If none of them gets the must haves right (identity, the action reads, no artifacts where you will cut), stop rerolling and rewrite the shot.
• Final: 1 or 2 takes at delivery resolution, then upscale only the keeper.

Rewrite instead of retrying when the same failure shows up three times. Ways to rewrite:

• Split one long action into two short shots.
• Replace a group scene with one character per shot and cut between them.
• Frame the hands out, or keep them still.
• Move text and logos to the edit.
• Hide the hard moment with a cut to a reaction or detail shot.
• Try another model for that shot. Kling handles some action content that Seedance rejects.

See Troubleshooting for fixes to specific failures.

Spend your retries on stillsCopy Anchor Link

"Spend your retries on stills" is fal's advice, and the logic holds: a bad keyframe costs one image, a bad clip costs a full video render. Fix the look, the character and the composition in images first, then animate.

• MindStudio budgets 3 to 5 image variations per shot frame before any video.
• A 6 panel storyboard sheet cost 1 credit in one test, against 6 credits for six separate frames, and kept props more consistent.
• Approve the reference carefully. A wrong detail in the image repeats in every take.

SourcesCopy Anchor Link

• BytePlus ModelArk Seedance 2.5 docs and pricing
• Gemini API Veo docs and pricing
• Google Flow credits help
• vfxwaffu on X, cost comparison
• fal on X, retries on stills
• Kapwing, Big Yowie creator interview
• MindStudio, AI short film under 200 USD
• heyuan110, Seedance 2 prompt guide
• JXP, GPT Image 2.5 storyboard test
• buraktuyan on X, spec ad case study