I2V, T2V and V2V name what an AI video model starts from. T2V (text-to-video) makes a clip from a written description alone. I2V (image-to-video) starts from a still image and animates it. V2V (video-to-video) takes an existing clip and changes it: extends it, edits it, restyles it, sharpens it or re-dubs it. You will also see R2V (reference-to-video), first-last frame and keyframe modes, and an “A” on the end (T2VA, I2VA) when the model makes sound along with the picture.
The labels look like jargon, but each one answers a practical question: how much of the shot have you already decided before the model starts? Here is every mode you are likely to meet in a model’s settings, what each is for, and which current models offer it.
The modes at a glance
| Short name | Stands for | You give the model | Use it for |
|---|---|---|---|
| T2V | Text-to-video | A written description | Exploring looks, establishing shots, anything with no reference yet |
| I2V | Image-to-video | A start frame plus a short motion prompt | Most story work, because the look is locked before you pay for motion |
| First-last frame (FL2V) | First and last frame to video | An opening image and a closing image | Reveals, transformations, shots that must end on an exact frame |
| L2V | Last frame to video | Only the closing image | Shots that must land on a picture, when the opening can be invented |
| Keyframes | Several stills to video | A series of images pinned to points in time | Choreographed moments that have to hit specific poses |
| R2V | Reference-to-video | Images of people, places or objects, plus a prompt | Several recurring characters in one new shot |
| V2V | Video-to-video | An existing clip | Extending, editing, restyling, upscaling or re-dubbing footage |
T2V: text-to-video
Text-to-video builds the entire shot from words. The model decides the faces, the clothes, the set, the framing and the light, and it makes those decisions again from scratch every time you press generate. Ask for “a lighthouse keeper climbs a spiral staircase” twice and you get two different lighthouse keepers.
That is a weakness for a story and a strength for everything before the story. When you don’t yet know what a film looks like, ten text-to-video clips of the same description are ten directions to choose from. It is also the honest choice for one-off shots that nobody needs to recognise twice: weather, skylines, crowds. We wrote a longer comparison in text-to-video vs image-to-video.
I2V: image-to-video
Image-to-video starts from a picture you supply, usually as the first frame. Everything you can see in that picture is decided. The model’s job shrinks to one question: what happens next?
That is why I2V is the default for anything with recurring characters. You make (or pick) the frames first, check that the same person looks like the same person across all of them, and only then spend money on motion. It also changes how you write the prompt. The image already says how things look, so the prompt should say only what moves. Describing the picture again tends to make the model change it.
Most people searching “i2v” want exactly this mode, and most of the work in a real AI film happens here.
First-last frame, L2V and keyframes
These are image-to-video with more than one anchor.
First-last frame (sometimes FL2V) takes an opening image and a closing image and generates the motion between them. It is the tool for a reveal, a costume change or a camera move that has to end on a precise composition. Keep the two frames related, the same person in the same room with a similar lens, or the model has to invent too much in the middle and the in-between gets strange.
L2V is the less common sibling: you give only the last frame and the model invents a plausible lead-up. MiniMax’s own prompt guide for H3 lists it as one of four task modes (T2VA, I2VA, FL2VA and L2VA) and describes it as a path “that converges from a plausible preceding state to the last frame”. The abbreviation isn’t standardised across vendors, so if you see L2V somewhere else, check the model’s documentation before assuming it means the same thing.
Keyframe mode goes further and pins several stills to points on the timeline. FLUX 3’s keyframe mode takes up to 10. It is the closest thing to blocking a scene with an actor: the model has to hit each pose in order, in one continuous take.
R2V: reference-to-video
Reference-to-video sits between text and image. You give the model pictures of who and what should appear, but not the frame itself, and the prompt says how to combine them. One image can be a character, another a location, another a prop. The model composes a new shot that contains them.
This is how you get two consistent characters into a shot that neither of their start frames shows. The catch is that references can leak. A character photo shot against a red wall may bring the red wall with it. Good reference prompts give each image one job and say what to ignore, for example “Image 1 is Theo: use only his face, hair and clothes, not the background”. Seedance 2.5 accepts up to 50 references, Wan 3.0 up to 10, and MiniMax H3’s reference mode up to 9.
V2V: video-to-video
Video-to-video is a family rather than one mode. What they share is that the input is footage, either yours or an earlier generated clip. The common jobs:
- Extend: continue a clip past its last frame. FLUX 3’s extend mode needs a source under 15 seconds.
- Edit: change something inside a clip while keeping the motion, such as a prop, a colour or the weather.
- Restyle: keep the movement and the timing but change the look, for example live action to animation.
- Upscale: render the same clip at a higher resolution with sharper detail.
- Re-dub: replace what someone says. Lightricks publishes an open-source “Dub-It” workflow for LTX that does this, validated so far in English, French, Spanish, German and Russian.
V2V is where a lot of finishing work is going. Fixing one wrong detail in a good clip is usually cheaper, and keeps more of what you liked, than generating the whole shot again.
What the “A” means: T2VA, I2VA
An “A” on the end means audio. A T2VA or I2VA model generates the sound together with the picture: footsteps, room tone, music and, on some models, lip-synced dialogue. That matters for the prompt. On a model with sound, “rain on a tin roof, a dog barking down the street” is a direction it will follow. On a silent model the same words do nothing.
Sound is not always optional. Gemini Omni 1.1 Flash and MiniMax H3 always generate it, so if you want a silent clip you have to say so (“no dialogue, no music”) or strip it in the edit. Kling 3.0 does not speak dialogue at all.
Which current models offer which modes
| Model | Extra frames | Reference images | Video-to-video |
|---|---|---|---|
| FLUX 3 | End frame, or up to 10 keyframes | – | Extend, edit, upscale |
| Seedance 2.5 | End frame | Up to 50 | – |
| MiniMax H3 | End frame, or last frame only | Up to 9 | – |
| Wan 3.0 | End frame | Up to 10 | – |
| Gemini Omni 1.1 Flash | End frame | – | – |
| LTX 2.5 | End frame | – | Open-source workflows, such as re-dubbing |
| Kling 3.0 | – | – | – |
All of them do image-to-video. A dash means the mode wasn’t in the documentation we checked on 22 October 2026, not that it can never exist. Capabilities change every few months, so look at a model’s current docs before ruling it out.
Which mode should you use?
Our rule of thumb is to decide as much as you can before the model does. If you are still finding the look, use T2V and generate a lot. Once the look is chosen, make start frames and switch to I2V for every shot that carries the story. Reach for first-last frame or keyframes when the ending of a shot matters, R2V when two or more recurring characters share a shot that no single frame shows, and V2V when a clip is nearly right and you want to fix it instead of starting over.
Whichever mode you pick, the prompt habits are the same. The 10 best practices for AI video prompts apply to all of them.
FAQ
What does I2V mean in AI?
I2V means image-to-video: a model that takes a still image, usually as the first frame, and generates a video clip that moves on from it. You write a short prompt describing the motion and the model fills in the rest.
What is V2V AI?
V2V, or video-to-video, covers models that take an existing clip as input and change it. The common versions extend a clip, edit something inside it, restyle it, upscale it or replace the dialogue.
What does T2V mean?
T2V means text-to-video. The model makes the whole clip from a written description, with no image to start from.
What is L2V?
Last frame to video. You supply the image the clip should end on and the model invents how it gets there. MiniMax H3 offers it under the name L2VA (the A for audio). Other vendors may use the letters differently, so check the model’s own documentation.
Is I2V better than T2V?
For a story with recurring characters, yes, because every shot can start from a frame you have already approved. For exploring ideas, T2V is faster and gives you more variety. Most projects use both: T2V early, I2V for the final shots.
What is the difference between I2V and R2V?
In I2V your image becomes the actual first frame. In R2V your images are references the model draws from to build a new frame, so the people and places appear but the composition is the model’s.