Castaly · Blog · Guides

Image to Video: First Frames, Last Frames and Reference Images by Model

How Castaly's video models use a still as a first frame, a last frame or a reference, which models accept what, what it costs, and three examples from still to clip.

A woman in a cobalt-blue trench coat on the top deck of a car park at dusk, a ribbon of neon-lit film strips coiling around her; this still became the first frame of a Wan 2.7 clip.

The most reliable way to get the video you imagined is to make the opening frame first. A still image is cheap to iterate, and once one is right, a video model only has to decide what happens next.

This guide covers the three ways Castaly's video models use images, which models accept which, what it costs, and three complete examples from still to clip.

Three ways to use an image

First frame. The image becomes the first frame of the video, and the model animates forward from it. Composition, lighting and characters are fixed by your still.

First and last frame. You also supply the closing frame, and the model fills in the motion between them. Useful for transitions and for landing a shot on an exact composition.

Reference images. The model keeps a subject, product or style from your images but composes a new shot. The reference never appears as-is; the character turns up in the scene your prompt describes.

Most models that accept both treat them as separate modes: one request uses frames, or references, not both.

Which models accept what

Model First frame First and last Reference images Reference video Other inputs
seedance-2.5 ✓ ✓ up to 30 up to 10 up to 10 audio clips
seedance-2.0, -fast, -mini ✓ ✓ up to 9 up to 3 up to 3 audio clips
wan3.0-video, -prime ✓ ✓ up to 10 up to 5 up to 5 audio clips, a file or a web page
MiniMax-H3 ✓ ✓ up to 9 up to 3 up to 3 audio clips
kling-3.0-omni, kling-o1 ✓ ✓ up to 7, with elements 1
kling-3.0, kling-2.6 ✓ ✓ 1, with 1 video 1, with 1 image
kling-3.0-turbo, kling-2.5-turbo ✓
wan2.7-i2v ✓ ✓ driving audio, or a first clip instead of a frame
wan2.7-r2v ✓ up to 5 with videos up to 5 with images
happyhorse-1.1-i2v one image, required
happyhorse-1.1-r2v up to 9, required

wan2.7-t2v and happyhorse-1.1-t2v take text only. On Kling 3.0 and 2.6, one reference image plus one reference video is motion control: the character from the image performs the movement from the video.

In the Studio, add an image as a reference and pick its role: First frame, Last frame or Reference image. Anything you generated can go to your Library from Creations and be reused as a reference from there.

What images cost

Nothing extra. A first frame, a last frame or a reference image does not change the price of the clip on any model. A 5-second 720p Seedance 2.5 clip costs $1.44 with or without a first frame.

Reference videos are different. Seedance adds the reference video's length to its token count, and Kling 3.0 Omni and Kling O1 switch to a higher per-second rate when a reference video is attached. The video price comparison has the figures.

So the real cost of image-to-video is the still plus the clip. Flat-price image models make the still almost free: $0.0212 on Wan 2.7 Image, $0.0332 on Seedream 5.0 Lite.

Example: one still, one first frame

The still came from Wan 2.7 Image at 1536 × 1024:

A woman in a cobalt-blue trench coat stands on the open top deck of a car park at dusk, a ribbon of neon-lit film strips coiling around her.
wan2.7-image · 1536 × 1024 — $0.0212 at today's price.

It then became the first frame for Wan 2.7 I2V. The video prompt does not describe the picture again; it says only what happens next:

One continuous take, no cuts. The ribbon of film keeps unwinding and speeds up until every frame is lit, tightening into a neon helix that lifts off the deck and climbs past the parapet into the dusk sky, the woman in the cobalt coat turning to follow it up. …

wan2.7-i2v · 5 s · 1080p · first frame from the still above — estimated $0.56. Still and clip together: about $0.58.

The clip keeps the still's 3:2 shape, because Wan 2.7 I2V follows its first frame's proportions. If you want 16:9 video, make a 16:9 still.

Example: a reference, not a frame

Here the still is a subject to keep, not a frame to start from:

A figure in an acid-green vinyl coat reaches for a bucket in a flower-shop cold room where the tulips along the shelves have turned to glass.
wan2.7-image · 1536 × 1024 — $0.0212 at today's price.

Wan 2.7 R2V took it as a reference image, with a prompt that moves the glass flowers somewhere new:

One continuous take, no cuts. Keep the glass flowers from the reference exactly as they are and carry them into a new place: a wide city crossing at night in the rain. …

wan2.7-r2v · 5 s · 1080p · 16:9 · one reference image — estimated $0.56.

The flowers carried over; the room, the figure and the composition did not. That is the difference between a reference and a first frame.

Example: a long shot from a first frame

For a 10-second hero shot, a Seedream 5.0 Lite still at 2560 × 1440 cost $0.0332 and set the scene: an observatory at dusk, a researcher's hands on a closed star atlas, the first silver flecks lifting off the painted dome. Seedance 2.5 then took it as a first frame:

One continuous take, no cuts. Every remaining star point tears off the dome at once and pours into a single roaring silver river that spirals around the researcher, knocks the brass instruments off the table and blasts out through the observation slit into the valley. …

seedance-2.5 · 10 s · 1080p · 16:9 · first frame from a Seedream still — 488,025 tokens, about $7.05. The first frame added nothing.

Tips that save generations

  1. Describe the change, not the picture. The model can already see the frame. Spend the prompt on motion, camera and what transforms.
  2. Set the shape in the still. Models that follow the first frame, and Seedance's Follow reference ratio, take their proportions from the image.
  3. Iterate on the image, not the video. Ten stills on Wan 2.7 Image cost $0.21, less than one 5-second 720p clip on most video models.
  4. Use a last frame for an exact ending. When the closing composition matters, supply it rather than hoping the model lands there.
  5. Keep text out of the still. Lettering is one more detail the model has to hold steady while everything else moves; the prompts above all end with "no text".

Start from an image in the Studio →

FAQ

Does adding a first frame make a video more expensive?

No. Frames and reference images do not change the price on any Castaly model. Only reference videos do.

Can I use both a first frame and reference images?

Not in the same request on most models. Seedance, Wan 3.0, MiniMax H3 and Kling Omni and O1 treat frames and references as separate modes. Wan 2.7 R2V is the exception, allowing a first frame alongside its references.

Which model accepts the most references?

Seedance 2.5: up to 30 reference images, 10 reference videos and 10 audio clips in one request.

Why does my image-to-video clip have the wrong shape?

Some models follow the first frame's proportions. Make the still in the ratio you want the video to be.