Text-to-video vs image-to-video
Text-to-video vs image-to-video: text invents the shot, image-to-video animates a locked still. Same catalog, opposite jobs.
The differences between text-to-video and image-to-video AI models start with the input. Text-to-video invents framing from a prompt. Image-to-video keeps your still and adds motion.
Text-to-video is generation from a written prompt alone. No image goes in. The model decides the subject, the set, the lens, the light. Image-to-video takes a picture you supply as the opening frame. The prompt is only what happens next. If the face, the SKU, or the label has to survive, start with a still. If you need a version of an idea and nothing has to match yesterday, start with text.
Invent versus lock
With nothing to copy, a text-to-video prompt does two jobs at once: it picks what appears and it picks how the thing is shot. A prompt that only names a subject has quietly handed the second job back to the model. Three words return a different film every time you press go. Iteration is the tool: hold the seed if the UI shows one, change one clause, compare the two takes. You cannot guarantee that a face, a product label, or a brand colour survives from one generation to the next, because there is no reference for it to survive from.
That is why T2V wins on establishing shots, abstract b-roll, and anything where a version of the idea is good enough. It loses the moment the clip has to match something that already exists — a real product, a real face, a still you generated yesterday. Those are I2V jobs, or reference-to-video if the row takes a pack.
Image-to-video fixes everything the model would otherwise guess at frame zero: the face, the product, the palette, the framing. Two clips built from the same still look like they came from the same shoot, because they did. The prompt changes job. In T2V it describes the picture. Here the picture already exists, so the prompt is stage direction: what moves, in which direction, how the camera behaves, what the shot lands on. Re-describing the image tends to hurt. The model works to satisfy a description that is already true and starts drifting the frame to match your wording.
A large share of I2V models will not run any other way. When a catalog record is marked as requiring an image, text alone is rejected rather than silently ignored. That is not a UI bug. It is the contract.
Native audio does not change the split. Veo 3.1 is a text-to-video row with audio on, 4 / 6 / 8 seconds, 4K tier, 40 credits on the published card. A talking clip from a prompt is still an invented face unless you used a different row that actually takes a reference. Seedance 2.0 lists text-to-video and reference-to-video; phoneme-level lip-sync in 8+ languages is a T2V/reference capability in vendor materials (fal, Atlas Cloud, Morphic; as of 4 September 2026), not a reason to skip the still when the still is the product.
When the still is the brief
Lock the still first whenever legal, brand, or the listing page has already said yes to a frame. Mood boards, pack shots, founder headshots, the hero of a PDP — those files are expensive to rediscover inside a video model. Animate the approved file. Do not prompt "the same bottle as last week" and hope.
The still can come from a camera or from a text-to-image row. Z Image Turbo is a 1-credit stills card with a 4K ceiling, built for drafts and A/B. Freeze a winner there, then buy motion once. Re-rolling a 6-second video to hunt a label is the expensive way to do a stills job.
After the still is frozen:
- Write only the motion. "Slow orbit left, label holds centre, no new hands."
- Match aspect to the cut. A 1:1 lock that you later pillarbox for 9:16 is a crop, not a generate.
- Keep type, fingers, and SKU out of the prompt if they are already in the pixels. Naming them again is how they mutate.
- Run more than one I2V take from the same still if you need variants. That is multiple clips from one photo, not one generate asked to be a film.
If one object on the still is wrong, edit the still. Qwen Image Edit 2511 is an instruction on a locked file (3 credits list). Then animate. Do not ask the video model to delete a cup while it is also inventing time.
Motion control is a third object, not a better I2V. It needs the still and a driving clip of a body. Use it when the choreography is the ad. Use ordinary I2V when "a little life" is enough.
When text is the brief
Start from words when there is no plate to honour.
Establishing weather, a city at dusk, an abstract texture, a crowd that does not have to be your crowd — T2V is cheaper than commissioning a still you will immediately throw away. The prompt has to carry subject, action, camera, light, and the state the shot ends in. Clip length is chosen up front from a fixed set of options, not trimmed afterwards like an NLE. The same prompt at a different seed is a different take, not a correction.
Talking generates live here when the face is not a legal asset. A presenter who explains a feature in eight seconds on Veo is a T2V job with native audio. A founder who already sat for a still is an I2V or lipsync job. Confusing those two is how a campaign gets a new actor every ad set.
Do not use T2V as a shortcut around a missing still. "Make me the bottle from the listing" without attaching the listing is a new bottle. If the SKU has to match, you do not have a T2V job yet. You have a photography or text-to-image job, then I2V.
Reference-to-video sits beside both. It is the mode for identity packs — several stills, sometimes a clip — when one frame is not enough to hold a character. It is not "T2V with extra adjectives." If the row asks for references, attach them. If it does not, you are back on invent-or-lock.
The two public doors on Versely are the AI video generator for the T2V side of the catalog and the AI image-to-video tool for the locked still. Pick the door that matches the files you have, not the door whose name you searched.
The test on the last ten clips
Look at what you actually shipped.
If eight clips had to match a pack shot, a face, or a frame from last Tuesday, you should have been on I2V (or lipsync, or motion control) and T2V was how the label drifted. If eight clips were weather, vibe, and a city that does not exist, I2V was how you paid for a still you did not need.
Hybrid is normal. Still on Turbo or a camera. Motion on I2V. Talking line on a lipsync or native-audio row if we see a mouth. Captions burned in after the picture is locked, because silent autoplay does not care which generate invented the frames.
Kling 3.0 (February 2026) and the rest of the 2026 roster did not erase the split. They made both sides better. Audio-native T2V (Seedance 2.0, Veo 3.1, Kling 3.0 — per Magic Hour's release tracker and WaveSpeed's 2026 roundup, read 4 September 2026) means the invented talking shot no longer needs a second mouth pass if you accept the invented face. I2V with a locked still still needs the still. The model family is not the job.
Run T2V when you are inventing. Run I2V when you are animating a decision you already made. The vs is that sentence.
Pick the door from the video generator or the image-to-video tool, and sign up when the still is approved and the next click should be motion, not another prompt.
FAQ
Can I use the same prompt on both?
You can paste it. You should not. On T2V the prompt is the picture and the camera. On I2V the picture exists; a full scene description fights the still. Write motion, light change, and ending state. Leave the bottle out of the sentence if the bottle is in the file.
Which one should I use for a talking founder?
Neither, if you already have the face. A locked still plus a lipsync or image-to-lipsync row keeps the person. T2V with native audio invents a speaker. I2V without a mouth pass animates a closed jaw. Match the row to the mouth.
Does native audio mean I should always start from text?
No. Native audio means the video model can write a mix in the same pass. It does not lock identity. A SKU still plus I2V plus a VO laid later is still the right stack when the product cannot change.