Video
Text-to-video vs. image-to-video: which should you use?
Key takeaways
Text-to-video builds the whole shot from your words; image-to-video starts from your picture, usually as the exact first frame.
Use image-to-video when the clip has to look a specific way: a particular face, product, composition, or style. Use text-to-video to explore ideas and get B-roll fast.
An image-to-video prompt should describe motion and camera only. Re-describing the image tends to produce stiff, low-motion clips.
Major video APIs usually charge the same per second for both modes, so the real extra cost of image-to-video is making the still first.
Some models won't take text at all, and some cap image-led clips shorter than text-led ones, so check the mode before you plan a shot.
Should you use text-to-video or image-to-video?
Start from an image when you already know what the shot should look like. Start from text when you don't, or when it doesn't matter much.
The two modes trade control against speed. Text-to-video asks the model to invent everything: the subject, the framing, the light, and the motion. Image-to-video hands it a finished frame and asks only for the motion. Generating a still first adds a step, but it's the most reliable way to get a clip that matches a picture in your head, and the only approach that holds up for keeping the same character or product across several clips.
How the two modes work
In text-to-video, the model starts from noise and shapes every frame from your prompt. Nothing's on screen until the model draws it.
In image-to-video, your picture becomes the starting point. The Wan video model's technical report describes the usual method: the source image is placed as the first frame, and the model generates the frames that follow while also reading the image's features. Black Forest Labs' FLUX 3 Video docs say a start image appears "pixel for pixel, as the opening frame."
The same machinery also runs three related modes:
First and last frame. You supply where the shot starts and where it ends, and the model fills in the motion between them. Google's Veo 3.1 and Gemini Omni Flash support this, and FLUX 3 Video reads three or more images as a timeline of keyframes.
Reference images. The picture defines who or what appears, but it never shows up on screen as a frame. Google's reference mode accepts up to three images "of a single person, character, or product." BFL's docs say: if the image must appear exactly as shot, use it as a frame; if it only needs to define identity, use it as a reference.
Video-to-video. You feed in a clip and the model restyles or edits it. It's usually priced higher than the other modes.
What image-to-video gives you
Control is the main reason to start from a still. Google's Veo prompting guide says the source image supplies the subject, the scene, and the style, and serves as the basis for character detail and lighting. You can spend as long as you like getting a single frame right in an image generator, then hand that frame to a video model.
That control pays off most with a recurring character. Text-to-video redraws a face from scratch on every run, so the same prompt gives a slightly different person each time. So creators keep one face across a series by starting each clip from an image of the same character, or by giving the model reference images.
It helps with real products too, whose packaging, colors, and proportions have to match the actual thing (Google's reference mode names products as a use case). An image also locks in a composition that came from a storyboard or a client brief, and a painted, anime, or film-grain look is easier to hold when the first frame already has it.
Where text-to-video is the better call
Text-to-video is faster when you're still deciding what you want. There's no image to make first, so one prompt gets you a finished clip, and five prompts get you five different takes on an idea.
BFL's text-to-video guide says that with a short prompt "the model fills in framing, motion, and mood," and recommends one-line prompts for quick ideas and B-roll. That's the right use: establishing shots, abstract backgrounds, crowd scenes, and anything where no particular face or object has to come back later.
It's also the only option when you don't have a source image and don't want to make one. A text prompt can describe a camera move through a space that would take several stills to plan.
Where each mode goes wrong
Image-to-video's best-known failure is a clip that barely moves. Researchers who studied the problem found that image-to-video models "tend to generate videos with less motion than expected" because they lean too hard on the source image, an effect they call conditional image leakage.
Re-describing the picture in your prompt is the other common cause. BFL warns that re-describing what's in the frame "invites the model to re-imagine it instead of animating it," and Google's guide says redundant prompts "confuse the model and lead to poor results." A weak source image hurts as well: Google notes that a sharp, well-composed photograph gives a more coherent clip.
First-and-last-frame shots have their own weak spot. BFL says its API marks end-frame pinning as experimental, warns of more variance between runs, and advises keeping the two frames related in subject, scene, or camera setup.
Text-to-video's main failure is consistency. A character described in words drifts from clip to clip, which is why BFL advises keeping a subject's description word-for-word identical across shots. Even then, it's a guess each time. Prompt adherence isn't great either: the model card for LTX-2.3, an open video model, says it may not match prompts perfectly.
How prompting changes between the two
The modes differ most in the prompt. For text-to-video, it carries everything: subject, setting, light, style, motion, and camera. Longer, more specific prompts give you more say over the result, though BFL notes that overstuffed prompts make motion less coherent.
For image-to-video, the image already answers most of those questions. Google's guide says: "Focus your prompt on the motion you want to see." Refer to subjects generically ("the woman," "they"), and spend the words on what should change: who moves, how fast, and what the camera does.
What each mode costs
On the major video APIs, the per-second price is usually the same whichever way you start. Black Forest Labs prices FLUX 3 Video identically for text-to-video and image-to-video, and video-to-video costs more:
Tier | Text or image to video, per second | Video to video, per second |
|---|---|---|
Draft | $0.06 | $0.12 |
HD | $0.17 | $0.41 |
FHD | $0.29 | $0.53 |
On Runway's API price list, most models also charge one rate regardless of input, while a few add a small fee for a start frame or reference image: one credit for an image-to-video start frame on Gemini Omni Flash, for example. So the real extra cost of image-to-video is the still you make first, which at BFL's API rates costs under five cents for a small FLUX 3 Image still. That's small next to a video clip, and it's often cheaper than generating three text-to-video takes to find one with the right look.
Clip length by mode
Clip length can differ by mode:
Model | Clip length |
|---|---|
FLUX 3 Video | 5 to 20 seconds in every mode |
Kling 3.0 | 3 to 15 seconds |
Runway Gen-4.5 | 2 to 10 seconds, text or image |
Veo 3.1 (Vertex AI) | 4, 6, or 8 seconds with first and last frames; 8 seconds with reference images on the standard model |
Luma Ray 3.2 | 5 seconds; 10 seconds isn't available with a start or end frame |
Some models skip text-to-video entirely. Runway lists its Gen-4 and Gen-4 Turbo video models as image-input only, while Gen-4.5 takes text or an image.
A quick way to decide
If you... | Use |
|---|---|
Have a face or product that has to look right | Image-to-video or reference images, starting from a still you've made or picked |
Need the same character in several clips | Images of that character for every clip |
Are exploring ideas or need filler footage | Text-to-video, with several takes |
Know exactly how the shot starts and ends | First and last frames, kept close in pose and framing |
For the image side of this workflow, the FLUX 3 explainer covers one current model family that does both stills and video.
Text-to-video vs. image-to-video FAQ
Is image-to-video better than text-to-video?
Neither is better overall. Image-to-video gives more control over how the clip looks, and text-to-video is faster when the exact look doesn't matter.
Why does my image-to-video clip barely move?
Image-to-video models tend to over-rely on the source image, which suppresses motion. Describe only the motion and camera in your prompt, and avoid re-describing what's already in the image.
Does image-to-video cost more?
Usually not per second. Most major video APIs charge the same rate for both modes, and a few add a small fee for a start frame. The extra cost is generating the source image, which is typically a few cents.
What's the difference between a start frame and a reference image?
A start frame appears on screen as the first frame of the clip. A reference image tells the model what a person, character, or product looks like, but it isn't shown as a frame. Use a start frame to lock a shot's composition, and reference images to keep an identity consistent across different shots.
Can I keep the same character with text-to-video alone?
Not reliably. Even a word-for-word identical description gives a slightly different person on each run. Lock the face in an image first, as a start frame or a reference, and reuse that image for every clip.