Draft on the cheap models. Finish on the good one.
Nobody gets the shot on the first try. You will render a dozen versions of a scene before one of them is right, and eleven of those go in the bin — which means most of your money is spent on searching for a shot rather than on the shot itself. That is the line item to attack, and the way to attack it is not to render the search at full price.
The order that costs least
Three stages, and the model changes at each one.
| Stage | Setup | Per 5s take | What you're asking |
|---|---|---|---|
| 1 · Find the idea | Mini, 480p | ≈ $0.18 | Does this scene work at all? Framing, movement, timing. |
| 2 · Tighten it | Fast or 2.0, 480p | ≈ $0.28–0.35 | Is the prompt doing what I think? Do the references hold? |
| 3 · Shoot the take | Seedance 2.0, 720p or 1080p | $0.76 / $1.87 | Nothing. You already know the shot — this is the render that goes in the edit. |
Put a realistic number of attempts through that: ten drafts on Mini, three refinements at 480p, one final at 720p comes to roughly $3.60. The same fourteen takes rendered straight on the full model at 720p cost $10.60. Same finished shot, a third of the bill — and the drafts come back faster, which matters more than the money when you're circling an idea.
What a draft can and cannot tell you
This is where the method goes wrong if you're careless: rejecting a good idea because the cheap render of it looked cheap. Keep the two categories separate.
| Trust a draft for | Never judge on a draft |
|---|---|
| Composition and framing | Face detail in close-ups |
| Camera movement and its speed | Hands and complex anatomy |
| Timing — whether the beat fits the duration | Fabric, hair and skin texture |
| Whether the prompt reads as intended | Fine motion and physical plausibility |
| Whether a reference image clears moderation | Anything you would call "polish" |
The right question at the draft stage is "is this the right shot?", not "is this a good shot?". A Mini render that has the blocking right and mushy faces is a pass — the faces are the full model's job. A Mini render with the wrong camera move is a fail no matter how it's rendered.
One consequence worth stating: don't send a draft to a client. They will judge it as a finished thing, because it looks like one. Drafts are for you.
Where upscaling actually fits
Upscaling is a delivery step, not a shortcut past stage three. It enlarges what the model already generated — it does not add the detail a stronger model would have invented, and it cannot fix anatomy, texture or motion artefacts. Upscale a weak draft and you get a larger weak draft, sometimes with the faces smoothed into plastic.
Where it does pay: after the final render, when the take is right and you want more pixels for delivery. Generators bill per pixel while upscalers bill per second of output, so taking a finished 720p take to 1080p costs from about $0.04 — far less than the $1.87 a native 1080p render would have cost. That is a real saving, applied to a shot you have already approved.
| Engine | ≈ Cost per 5s at 1080p | Best for |
|---|---|---|
| ByteDance, AIGC preset | $0.04 | The default — built for generated footage |
| ByteDance, pro tier | $0.36 | Same engine, ten times the compute |
| Topaz Proteus | $0.10 | Conservative, stays close to the frame |
| Topaz Starlight | $0.10 | Diffusion — redraws detail, strongest on faces |
| Topaz Gaia | $0.10 | Gentle, for sources that are already clean |
And a limit worth knowing before you plan around it: upscalers take no reference and no prompt. We checked every engine above — video in, settings in, nothing else. They reconstruct detail from the frame itself and have no idea what your character is supposed to look like. A face the generator got wrong comes back wrong and sharper.
How this looks in Longtake
The model picker carries live per-take prices, so the cost of the stage you're in is on the button before you press it. Takes are numbered per scene, so a scene reads as a column of attempts with what each one cost — the draft-to-final progression is visible rather than something you track in your head. The same logic is wired into character screen tests, which run on the cheapest available model precisely because the question there is "does this portrait pass?", not "is this footage beautiful?".
Questions
- Why not just draft at 480p on the full model?
- You can, and it's still cheaper than 720p. But the model, not the resolution, is where the money is: Mini at 480p is about half the price of the full model at the same resolution, and for "does this composition work" it answers just as well.
- Do drafts and finals come out the same, only sharper?
- No — a different model makes different footage, not the same footage at higher quality. Blocking and timing usually carry over; detail, texture and fine motion change. That is exactly why the final render is a real render and not an upscale.
- How many drafts is normal?
- For a scene you care about, five to fifteen. If you're consistently past twenty, the problem is usually the prompt or the references rather than the model.
- Is Seedance 2.5 worth it for finals?
- It costs more per token and goes up to 30 seconds with far higher reference limits, so it earns its price on long takes and crowded scenes. For a five-second shot the full 2.0 is usually the sensible finish.
- Can I upscale a draft to skip the final render?
- You can, and sometimes for a background shot it's fine. Just know what you're getting: a bigger version of the draft, not the take the full model would have made.
Related
Prices checked in August 2026: BytePlus ModelArk for generation, fal.ai for upscaling. Figures move; the order of operations doesn't.