Pick T2V or I2V before the prompt

Pick T2V or I2V before you write the prompt. The first radio button is the job. The paragraph in the box is not.

Search still asks “text-to-video vs image-to-video” as if that were a model review. It is not. On a generator it is a mode: text in, or a still plus text in. The AI video generator exposes both. This week’s decision is which one you press before you spend a sentence describing camera, light, or dialogue. A prompt written for the wrong door is how you pay video credits to invent a face you already photographed, or how you feed a pack shot into a text-only row and watch the label drift.

This is not a roundup of every T2V and I2V slug. It is the radio button you should not skip on Monday.

The radio button is the contract

Text-to-video invents the subject and the motion. Image-to-video takes a picture you already approved and invents motion forward from it. Those are different contracts. Mixing them in one prompt — “use this product but also a new kitchen and a new person” — is how identity becomes a lottery.

If the still exists and the still is the deliverable, I2V is the door. Composition, wardrobe, label, and room are already decided. The prompt should talk about what moves: a slow push, a hand turning the bottle, a camera orbit that does not restage the pack. Kling Video V3 Standard Image to Video is one published I2V row: image required, native audio on, 3–15 seconds, 63 credits. The still is not optional on that row. That is the point of the radio button.

If there is no still, T2V is the door. Veo 3.1 is a text-to-video row: native audio, 4 / 6 / 8 seconds, 4K listed, 40 credits. You are asking it to invent a take from a sentence. That is allowed when the shot does not exist yet — a scene, a street, a room you will never photograph. It is the wrong door when legal already signed a pack photograph and you are hoping the model will “just add motion” from adjectives.

Reference-to-video is a third radio, and it is not this page. R2V treats the still as a casting call: identity travels, the scene is new. I2V treats the still as frame one. If you are unsure, you do not have a prompt problem. You have a mode problem. Stop typing.

Write the prompt after the mode. A T2V prompt names the subject, the room, the camera, and (on audio-native rows) the sound. An I2V prompt assumes the subject is already in the file and should not re-describe the bottle as if it were fiction. Re-describing a locked still is how I2V restages it.

What “this week” actually means on a slate

A week of clips is usually a mix. One hero talking shot. Three pack turns. A b-roll street. A hook poster that must load as that image. The mistake is opening the generator once, leaving the radio on whatever you used yesterday, and pasting every brief into the same box.

Before you generate anything on Monday, sort the slate into two piles:

Still exists and must hold. Pack shots, a face legal approved, a thumbnail that is also the first frame, a room you already lit. These are I2V. Crop the still to the ratio you will ship before you attach it. “9:16” in the prompt does not recrop a 16:9 plate.
Shot does not exist. Dialogue in a place you will not film, a camera move that has no reference, a scene that is allowed to be invented. These are T2V. Write sound into the prompt on rows that generate native audio. Leaving sound out means the model chooses for you.

Do not put both piles through one row because it is convenient. A T2V row will invent a cousin of your SKU. An I2V row given a weak still will animate the weakness for the full duration. The cheap place to reject a still is before you buy seconds.

Duration ladders differ by row. Veo 3.1 does not sell 15 seconds. Kling V3 Standard I2V does. Do not write a fifteen-second I2V brief into a Veo T2V box and call the truncation a model failure. The radio button includes the catalog card: durations, audio flag, whether an image is required.

Retries follow the same split. If I2V comes back with the right pack and the wrong motion, change the motion sentence, not the still. If T2V comes back with the wrong person, you do not have a motion bug. You never locked identity. Either lock a still and switch to I2V, or accept that T2V will keep sampling faces.

Do not prompt around a mode you have not picked

People try to fake I2V inside T2V by writing “exactly this product, same label, same lighting.” The model has not seen the product. It has seen words. Words are not a barcode. The output will be a bottle. It will not be the SKU.

People try to fake T2V inside I2V by attaching a still and asking for a different room, a different person, and a different time of day. That is a restage. Some rows will attempt it. Most will warp the still until it is neither the photograph nor a new shot. If you wanted a new scene with the same face, that is reference-to-video, and you should say so on a row that lists that category. If you wanted a new scene with a new face, put the still away and use T2V.

The test is one sentence, written before the prompt:

“This picture is frame one.” → I2V.
“This picture is the actor; the scene is new.” → R2V, not this radio.
“There is no picture.” → T2V.

If you cannot say one of those three out loud, do not generate yet. Photograph or generate the still first. Image credits are the identity debugger. Video credits are a terrible one.

Hands, labels, and type are the usual I2V failures when the still was never locked. A T2V talking shot can hide a mushy label in motion. An I2V pack turn cannot. If the still would fail a brand audit on a freeze-frame, I2V will publish that failure at 24 or 30 frames a second. Fix the still. Then press the I2V radio.

Silent versus sounded is a second radio, and it is not this page. Some I2V rows carry audio; some T2V rows do not. Read the catalog flag after you pick the mode. Do not assume “video” means dialogue.

The Monday checklist

List the clips you actually owe this week, not the ones you might make if a generate is pretty.
For each clip, answer: does a still already own identity? Yes → I2V. No → T2V.
Crop stills to the output ratio. Attach the cropped file, not the desktop original.
Open the generator. Press the radio. Then write the prompt for that door only.
On I2V, prompt motion and (if the row has audio) sound. Do not re-describe the pack.
On T2V, prompt subject, room, camera, and sound. Do not pretend a paragraph is a photograph.

If the week is mostly pack shots, your default radio is I2V. If the week is mostly invented scenes, your default is T2V. Defaults are not a religion. They are how you stop paying for the other job by accident.

When the still is ready and the radio is I2V, generate. When there is no still and the radio is T2V, generate. Do not open a blank T2V box “to see what happens” on a SKU you already photographed. That experiment is how the cousin bottle ships.

If you need an account to run the two doors against the same brief, sign up and pick the radio before the prompt. The prompt is the second decision. It is not the first.

FAQ

Can I write one prompt and use it on both radios?

No. A T2V prompt has to invent the subject. An I2V prompt should not. Paste the same paragraph into both and you will either restage a locked still or ask a text-only row to honour a photograph it cannot see. Write after you pick.

Is image-to-video always cheaper than text-to-video?

Not as a rule. Credits follow the named row, not the radio label. Kling V3 Standard I2V is 63 credits on the published card; Veo 3.1 T2V is 40. Compare slugs, durations, and audio flags. The reason to pick I2V is identity, not a discount.

What if I only have a messy phone photo?

Then you do not have a locked still. Crop, clean, or regenerate the still until a freeze-frame would pass. I2V will not rescue a bad plate. T2V will not honour it either. Fix the picture, then press I2V.