Making pictures and video from a phone, without wasting money

The prompt gets all the attention. The settings underneath it decide what you actually get, and what it costs — sometimes by a factor of two for the same four seconds.

The five settings that matter

Every image and video model in the catalogue is really a prompt plus a handful of controls: aspect ratio, resolution, length (video only), audio (video only) and reference material. Everything else is decoration. Get those five right and a cheap model beats an expensive one used carelessly.

Aspect ratio: choose it before you write the prompt

This is the one people set last and should set first, because it changes the composition rather than the crop. Ask for 16:9 and the model composes a wide shot; ask for 9:16 and it puts the subject in the middle of a tall frame with room above and below. Generating wide and cropping to tall afterwards gives you a worse picture, every time.

The lists differ by model, and they are longer than you would guess. Nano Banana 2 goes from 1:1 out to 21:9 and even 8:1 for a banner. GPT Image 2 adds 2:1, 3:1 and their inverses. The Kling video models accept exactly three — 16:9, 9:16 and 1:1 — so if you need 4:3 video, Seedance is where to go. On a phone, 9:16 is usually what you want and almost never the default.

Resolution: the setting that moves the bill

Nobody publishes what a resolution does to the price, so we ran the generations and wrote it down. These are real recorded costs, in tokens you pay.

Two things fall out of that table. On Nano Banana Pro, 2K is free — it costs the same 36 tokens as 1K, so there is no reason ever to run that model at 1K. And 4K is not a small upgrade anywhere else: on Nano Banana 2 it is more than double the 1K price, and it takes longer.

For video the same lever is bigger. Seedance 2 Mini costs 30.4 tokens for four seconds at 480p and 65.6 at 720p — the identical clip, more than twice the price. Kling V3 Turbo goes 108 at 720p to 135 at 1080p for three seconds.

Length: linear, and it adds up fast

Video models bill roughly by the second, so doubling the length roughly doubles the price. The ranges are not the same across models and this is worth knowing before you plan a shot: Kling 2.5 Turbo Pro accepts exactly 5 or 10 seconds and nothing in between. Veo 3 Fast takes 4, 6 or 8. Seedance 2 Mini runs from 4 to 15. Seedance 2.5 goes to 30, which is long enough to be genuinely expensive — plan that one.

Where we have measured one length and you ask for another, the app scales the quote and labels it an estimate rather than presenting arithmetic as an observation. Where we have measured your exact combination, the quote is the real number.

Audio: on by default, and not free

The Seedance models generate sound with the picture and have it switched on by default. Seedance’s own documentation is explicit that enabling audio costs more — one of the very few places any provider says so out loud. Veo 3 produces native audio too, and Kling 3.0 has a sound flag.

If the clip is going into an edit with its own music, or onto a feed where it plays muted, turn it off. It is the cheapest saving on this page and it costs you nothing you were going to use.

Reference material: bring your own

This is the part most people never touch, and it is where the quality actually comes from. Nearly every model here accepts input images, and several accept far more than one.

• Nano Banana 2 takes up to 14 input images. Nano Banana Pro takes 8. That is enough for a character sheet, a set of product angles, or a style reference plus the subject.
• Seedance 2 Mini takes a first frame and a last frame, plus up to 9 reference images, 3 reference videos and 3 reference audio clips. Giving it a first frame is the single most reliable way to control what a video looks like.
• Gemini Omni Video takes up to 7 images, 3 audio files, and one video segment with a start and end time — which makes it the one to reach for when you are working from footage you already have.
• Veo 3 supports first-and-last-frame and reference modes as explicit generation types.

Uploads are capped at 25 MB per file, and the accepted formats are the ones a phone actually produces: JPEG, PNG, WebP, HEIC and GIF for stills; MP4 and MOV for video; MP3, M4A, WAV and WebM for audio; plus PDF, plain text, Markdown and CSV for documents. A photo straight out of an iPhone is HEIC and works as-is.

A workflow that works on a phone

You are standing up, on a slow connection, and each video attempt takes one to five minutes. Fighting that with the expensive model is how a bill happens.

• Draft cheap and small. Seedance 2 Mini at 480p, audio off, four seconds. About 30 tokens. Run three variations of the prompt for less than one 720p attempt.
• Fix the prompt, not the model. Most bad output is an under-specified shot. Say what is in frame, where the camera is, and what moves. Kling takes a negative prompt as well — use it for the thing that keeps appearing and should not.
• Lock the frame. Generate a still you like on an image model first, then hand it to the video model as the first frame. A 16-token picture plus one video run beats five video runs.
• Spend once at the top. Only when the composition is settled, run it at 720p or 1080p, audio on if you need it.

The quote screen is the safety net

Every generation is priced before it runs and waits for you to say yes. The number is in tokens, your balance is next to it, and the app tells you whether that price was measured or estimated. Nothing is spent until you agree, and if the job then fails or produces nothing, you are not charged at all.

That is the moment to catch an accidental 4K. Read it — it is there for exactly this. Every model’s price is also on the pricing page, along with the full measured table, without needing an account.