IMAGE TO VIDEO
Start from a frame, then move it
Give the model a still and describe only the movement. Thirteen models, an optional last frame to pin where the shot ends, durations up to 30 seconds, and output to 4K. Any image you already generated can be picked without downloading it first.
Updated:
What image to video does
Image to Video animates a still you supply, optionally guided by a prompt. The first frame is fixed before generation starts, so composition, wardrobe, lighting and framing are settled and the model is left with one job. Deciding how the shot moves.
That is the practical argument for generating a still first and animating it second, even when text to video could have produced the same shot in one pass. A still is cheaper to reject than a clip, and it gives you an approval step in the middle of the process rather than at the end of it.
The image to video form, control by control
Seven controls, and the two image blocks at the top account for most of what the result looks like.
Image to Video sits at the top of the generation type list.
Pick the generation type
Image to Video is the first entry in the type dropdown, above Text to Video, Start-End Frame Video, Draw to Video and Video Enhancer. It is the type to reach for whenever the first frame already exists.
Selecting it rebuilds the form with an image source block at the top, which is the structural difference from Text to Video. Everything below that block behaves the same way.
Choose a model, and read its limits
The image to video list adds the Vidu Q3 Drama and Vidu Q3 Turbo variants that text to video does not offer. Vidu Q3 Drama reads 720p to 1080p and 2s to 30s, which is the longest single generation available anywhere in the video workspace.
Some models behave differently here than they do in text to video. Seedance 2.0 Mini reads 480p to 4K in this list against 480p to 1080p in the text to video one, so the same model name is not the same ceiling across generation types.
The image to video roster is longer than the text to video one, and the limits differ per model.
Five ways to supply the first frame. Previous Task is what makes chaining generations practical.
Supply the first frame
The Image Source block takes an upload, a previous task from your own history, a saved AI Actor, a saved AI Character, or a direct image URL. The drop zone accepts JPEG, PNG, WEBP, GIF and AVIF up to 100 MB, with a minimum of 300 by 300 pixels.
Previous Task is the option worth knowing about. It lets a still generated in the image workspace become the first frame here without a download and re-upload in between, which is how a still and a clip end up part of one sequence rather than two separate jobs.
Add a last frame if you have one
A second, optional block accepts a last frame image using the same five input methods. Leaving it empty lets the model decide where the clip lands. Filling it pins the final frame.
This is worth distinguishing from Start-End Frame Video, which is a separate generation type built around the same idea. Here the last frame is an optional constraint on an animation, and there the transition between two images is the entire job.
The optional last frame. Supplying it constrains where the shot ends up.
The prompt is optional here, and it should describe motion rather than the scene.
Describe the movement only
The prompt is optional in image to video, because the model already has the scene. When you do write one, everything spent describing what is visible in the frame is wasted, since the frame is already settled.
Use the space on movement instead. Which elements move, in what direction, at what speed, and what the camera does over the clip. A short prompt naming one clear camera behaviour usually beats a long one restating the picture.
Set the duration
The slider runs between the end points of the selected model, which the picker printed beside its name. Vidu Q3 Pro accepts as little as 1 second, and Vidu Q3 Drama reaches 30.
Animation faults show up early. A movement that looks wrong in the first two seconds will still look wrong at fifteen, so testing at the short end is the cheap way to find out whether the prompt and the frame agree.
Duration, with end points set by the selected model.
Resolution is filtered to what the selected model can produce.
Set the output resolution
Only the resolutions the selected model supports appear here, so a model with a single fixed output shows one option and a model spanning a range shows the full set.
Matching output resolution to the source frame is worth a moment of thought. Animating a small image at a high output resolution asks the model to invent detail that was never in the input, and the result usually reads softer than the still it came from.
Models available for image to video
Thirteen models. This list adds the Vidu Q3 Drama and Turbo variants that text to video does not carry, and several models expose different limits here than they do there.
| Model | Best for | Notes |
|---|---|---|
| Gemini Omni Flash Default | First passes | The default. 720p, 3 to 10 seconds. Fast enough to test whether a frame animates well. |
| Vidu Q3 Drama | Long single takes | 720p to 1080p, 2 to 30 seconds. The longest generation available in the workspace. |
| Vidu Q3 Turbo | Speed on the Vidu family | 540p to 1080p, 1 to 16 seconds. The fast variant beside Q3 Pro. |
| Vidu Q3 Pro | Quality on the Vidu family | 540p to 1080p, 1 to 16 seconds. Accepts clips as short as one second. |
| Seedance 2.0 Mini | Reaching 4K | 480p to 4K, 4 to 15 seconds. Its ceiling here is higher than in text to video. |
| MiniMax H3 | Consistent 2K output | 2K, 5 to 15 seconds. |
| Luma Ray 3.2 | Motion quality | 540p to 1080p, at fixed 5 or 10 second lengths. |
| Happy Horse 1.1 | General animation | 720p to 1080p, 3 to 15 seconds. |
| Happy Horse 1.0 | Comparison against 1.1 | Same range as the newer version, useful when 1.1 misreads a frame. |
| Seedance 2.0 VIP 1080p | Guaranteed 1080p | 1080p fixed, 4 to 15 seconds. |
| Seedance 2.0 Uncensored | Subjects other models refuse | 720p, at 5, 10 or 15 seconds. |
| Google Veo 3.1 | Prompt adherence | Strong on following a motion instruction literally. |
| Kling 3.0 PRO | Prompt control and sound | 3 to 15 seconds, with a negative prompt field and audio generation. |
Resolution and duration figures come from the model picker in the workspace, which prints them beside each model name. Read them in the form you are actually using, because limits are set per generation type rather than per model: Seedance 2.0 Mini reaches 4K in image to video and stops at 1080p in text to video.
Settings reference
Every control in the image to video form, what it changes and the values it accepts.
Image Source
The first frame. Accepts an upload, one of your previous completed tasks, a saved AI Actor, a saved AI Character, or a direct image URL.
JPEG, PNG, WEBP, GIF, AVIF up to 100 MB, minimum 300 by 300 pixels
Last Frame Image
Optional final frame, taking the same five input methods. Supplying it pins where the clip ends instead of leaving that to the model.
Optional, same formats as the first frame
Detailed Prompt
Optional here, since the model already has the scene. Best spent entirely on what moves and what the camera does.
Up to 2,500 or 4,000 characters, depending on the model
Model
Decides the resolution ceiling, the duration range and which extra controls exist. The image to video list is longer than the text to video one.
Thirteen models, including the Vidu Q3 Drama and Turbo variants
Duration
Clip length in seconds, bounded by the selected model. A slider where the model accepts a range, discrete values where it does not.
1 to 30 seconds across the roster
Video resolution
Output resolution, filtered to the selected model. Raising it well above the resolution of the source image asks the model to invent detail.
480p through 4K, model dependent
Aspect ratio
The frame shape of the output. Where it differs from the source image the model has to decide what happens at the edges, so matching the source is the safer default.
Model dependent
Choosing a frame that animates well
Some stills animate cleanly and some fight it, and the difference is usually visible before you generate anything. A frame with clear depth, meaning a foreground subject separated from a background, gives the model somewhere to put parallax when the camera moves. A flat frame with everything on one plane gives it nothing, and the result often slides rather than moves.
Hands, faces at small scale in frame and text on signage are the usual trouble spots. They are stable in a still because nothing is asking them to change, and they tend to deform once motion is applied. Where the shot allows it, framing so those elements are larger and better lit costs nothing and removes most of the problem.
The aspect ratio of the source is worth matching in the output. Where the two differ the model has to decide what exists beyond the edges of your image, and inventing that border is a different task from animating what you gave it.
When to fill the last frame field
Leaving the last frame empty is the right default. The model resolves the ending itself, which is what you want when the shot is exploratory and you are still finding out what the movement should be.
Filling it earns its place when the clip has to hand off to something else. A shot that cuts to a second clip reads better when its final frame is close to that clip's opening frame, and pinning the ending is the direct way to get there. The same applies to a loop, where the last frame needs to return to the first.
Where both end points genuinely matter more than the path between them, the Start-End Frame Video type is the better tool. It is built around that constraint rather than treating it as an option, and it exposes a movement amplitude control that image to video does not.
Model tutorials
Deeper coverage of the models in this roster, with sample output and the frames where each one struggles.
Image to video questions
What image formats and sizes does it accept?
The drop zone takes JPEG, PNG, WEBP, GIF and AVIF up to 100 MB, with a minimum of 300 by 300 pixels. You can also paste a direct image URL, pick one of your previous completed tasks, or select a saved AI Actor or AI Character instead of uploading anything.
How is this different from Start-End Frame Video?
Image to Video animates from a first frame, and the last frame field is an optional constraint on that animation. Start-End Frame Video is built the other way round: two images are the required input and generating the transition between them is the whole job. Use image to video when you care about the movement, and start and end frame when you care about both end points.
Can I animate an image I generated in the image workspace?
Yes, and without downloading it first. Choose Previous Task in the Image Source block and pick the completed generation. That is the normal way to chain a still into a clip, and it keeps both results in the same Artifact Gallery.
Which model gives the longest clip?
Vidu Q3 Drama, at up to 30 seconds. It is the only model in the workspace that reaches that length in a single generation, and it is available in image to video rather than text to video. Most other models stop between 10 and 16 seconds.
Why does the same model offer a different resolution here?
Limits are set per generation type, not per model name. Seedance 2.0 Mini reads 480p to 4K in the image to video list and 480p to 1080p in the text to video one. The picker prints the current range beside the model name, so read it in the form you are actually using.
Should the output resolution match my source image?
Roughly, yes. Asking for an output well above the resolution of the input means the model is inventing detail that was never there, and the clip often reads softer than the still it came from. Where you need both length and resolution, upscale the finished clip afterwards instead.
What does an image to video generation cost?
Most video models charge per second of output, and higher resolutions cost more, so the figure moves with the model, duration and resolution. The exact cost is shown in the form before you submit, and failed tasks are refunded automatically during credit reconciliation.
Other ways to make a video
Image to Video is one of six routes into the video workspace.
Video & LipSync
- Video Generator
- Text to Video
- Image to Video
- Start-End Frame to Video
- Draw to Video
- Motion Control
- Video Enhancer
- Video Upscaler
- Video to Video LipSync
- Audio to Video LipSync
- Image to Video LipSync
- Video FaceSwap
- Seedance 2
- Vidu Q3 Pro
- Gemini Omni
- Google Veo 3.1
- Kling 3.0 Pro
- Luma Ray 3.2
- LTX 2.3
- Happy Horse 1.1
- WAN 2.7
- Kling 3.0 Motion
- ByteDance Upscaler
- InfiniteTalk
- InsightFace