Creative workflows
First and Last Frame AI Video: How It Works, How to Pick the Two Frames, and What Went Wrong in Our Tests
What first and last frame (start and end frame) AI video does, which tools offer it, how to choose two frames that join cleanly, prompts that work, and what our two recorded Veo 3.1 Fast clips got right and wrong.
Video examples here were generated on Cicadas and keep their model, settings and date. Cicadas earns money when you generate with these models. How we test
- Test window:
- 2026-09-08 – 2026-10-01
- Clips recorded with these models:
- 9
- Models:
- Veo 3.1 Fast
- Their provider cost:
- ≈$8.90
- Selection:
- every clip we keep is listed with its settings, cost and request ID; failed or discarded runs are listed in the test record.
Short answer: first and last frame AI video takes an opening image and a closing image and generates the motion between them. It works best when the two frames show the same subject, in the same light, from almost the same place, and differ by one change: the camera moves closer, the light shifts, one thing moves. On Cicadas the first and last frame video page does this with Veo 3.1 Fast, silent, 4 to 8 seconds, from 100 credits. In our two recorded clips the endpoints held; the path between them did not always follow the prompt.
We make Cicadas, and the page and clips in this guide are ours. The clips were recorded on 8 and 20 September 2026. On 5 November 2026 we cut both into frame sheets, measured how close their first and last frames came to the images we supplied, and read the other tools' help pages. Prices come from the page's pricing in our code on the same day.
What first and last frame generation is
Ordinary image to video gives the model one picture and lets it decide where the clip ends. First and last frame (also called start and end frame, or keyframes) adds a destination: the clip should open on image A and close on image B. Your prompt describes the journey.
Two things follow from that:
- The endpoints are targets, not copies. The model redraws the scene. In our product clip, the first frame scored 0.92 and the last frame 0.86 on a 0-to-1 similarity scale (SSIM) against the images we supplied: clearly the same pictures, not identical ones.
- The middle is invented. Nothing in the two images says how to get from one to the other, so the model chooses. A clear, physical move in the prompt ("the camera pushes in", "the light warms") narrows its choice.
It is not frame interpolation, which adds in-between frames to footage you already have. Here there is no footage, only two stills.
Where you can do it
| Tool | What it calls it | Notes |
|---|---|---|
| Cicadas | First and last frame video | Veo 3.1 Fast, 4, 6 or 8 s, 720p or 1080p, silent, JPG/PNG/WebP up to 8 MB each, price shown first |
| Google Flow | Frames: "Add start frame", "Add end frame" | The end frame can be left out, per Google's Flow help page |
| Kling | Start and End Frames | Described on Kling's own quickstart page |
| Luma Dream Machine | Keyframes | Described in Luma's learning hub |
The three outside tools are listed from their own help pages found on 5 November 2026; we did not run them for this post, and their plans and limits change often. The Veo 3.1 page on Cicadas uses the same model for text and single-image clips, with sound if you want it. If you only have one picture, image to video is the simpler page.
Our two recorded clips
Both were made on the first and last frame page with Veo 3.1 Fast, 8 seconds, 16:9, 720p, silent.
The same picture at both ends: a loop attempt
We used one paper-cut landscape as both the first and the last frame and asked for clouds that "drift gently and return to their starting position" with a locked camera. The result opens and closes on nearly the same picture: its first and last frames score 0.98 against each other. In between, the island, the cottage and the waterfall stay put, but a large cloud drifts across the front of the island from about 2.5 to 5.5 seconds and leaves again. That is motion, which is what we wanted, but played on repeat the clip pulses rather than flowing. Recorded provider cost: US$0.80 (estimated). Open this recipe.
If you want a loop, generate it this way, then run it through the free loop video tool, which can blend the end into the start before it repeats the clip, in your browser.
Two different pictures: a product push-in
The first frame was a wide shot of a green glass bottle on wet stone with water arcing around it; the last was a close-up of the same bottle. The prompt asked for "one continuous physical camera push-in" spread "over the full eight seconds", with the bottle rigid and "no ghost images". What our frame sheet shows:
- 0 to 3 seconds: a wave of water crashes over the bottle, and for a moment you see the bottle through the water. This is the ghosting the page warns about.
- About 3 to 4 seconds: most of the push-in happens here, and the bottle slides from the centre towards the right as it grows.
- 4 to 8 seconds: the close-up holds, with splashes behind.
So the clip starts and ends where we asked, but compresses the camera move into about a second instead of eight. Recorded provider cost: US$0.80. Open this recipe. The same bottle as a single-image clip is in our perfume product video guide.
How to choose the two frames
- Same subject, same light, same lens. The more the two pictures share, the less the model has to invent.
- One difference. Camera distance, a lighting change, a door that opens. Two or three differences at once is where morphing and ghosting start.
- Same shape as the output. Pick 16:9 or 9:16 frames to match the clip you will make, so nothing is cropped differently at each end.
- Pull both frames from footage when you can. If you already have a video and want a new move that starts and ends on its stills ("match video to a still"), export exact frames with the free video to frames tool and use them as the two endpoints.
- Keep text out of both frames. Labels and lettering are the first thing to warp in the middle.
Two paintings of the same scene also work as endpoints; for a single painting that should move without changing, the animate a painting page is built for that.
Prompts that work
The page ships with three starting prompts, one for each kind of change. Write the move, then what must stay still:
- Camera move: "Move the camera slowly from the opening wide composition toward the closing close-up. Keep the same subject rigid and stable; no cuts, duplicates or morphing."
- Light change: "Keep the camera locked while lighting changes gradually from the opening image toward the closing image. Preserve scene geometry and subject shape."
- Loop attempt: "For matching endpoint images, keep the landscape stable while clouds drift gently and return toward their initial positions; inspect the join before treating it as a loop."
The page also appends a sentence of its own, asking the model to use your first image as the opening frame and the second as the closing target, with no cuts, text or logos.
What goes wrong, and what to do
| What you see | Likely cause | Try |
|---|---|---|
| Objects pass through each other or show double | The frames differ too much | Bring the frames closer: same angle, one change |
| The move happens in one burst | The model chose its own timing | Use 4 or 6 seconds, or a smaller change between frames |
| Something new appears mid-clip (a splash, a cloud) | The model fills the middle with what fits the scene | Name what must stay still; remove busy elements from the frames |
| The ending is close but not exact | Endpoints are targets, not copies | Cut the last few frames in an editor if you need a hard match |
A generation that fails, or a file that fails our check, returns its credits; a delivered clip that misses is still charged, so test with short, small changes first.
Price and one thing to watch
| Length | Credits | With the 1,500-credit pack (US$15) |
|---|---|---|
| 4 seconds | 100 | US$1.00 |
| 6 seconds | 150 | US$1.50 |
| 8 seconds | 200 | US$2.00 |
Resolution (720p or 1080p) does not change the price on this page. When we read them on 15 October 2026, fal's pages for the Veo 3.1 Fast endpoints said they would be deprecated on 17 November 2026; the first and last frame page will say what replaces them on Cicadas.
Questions
What is first and last frame AI video?
You give the model two still images, one for the opening and one for the ending, plus a sentence about the move between them, and it draws the frames in between. The two images guide the clip; they are not copied pixel for pixel, so the first and last frames of the result come close to yours without matching exactly.
Which AI video generators support first and last frame?
On Cicadas, the first and last frame page runs Veo 3.1 Fast. Google Flow has start and end frames, Kling has a Start and End Frames mode, and Luma Dream Machine calls it keyframes, according to each one's own help pages that we found on 5 November 2026. We only tested Cicadas.
How much does a first and last frame video cost on Cicadas?
A silent Veo 3.1 Fast clip costs 100 credits for 4 seconds, 150 for 6 and 200 for 8, at 720p or 1080p. With the 1,500-credit pack at US$15, an 8-second clip is US$2.00. The quote shows before you generate.
Can first and last frame make a seamless loop?
Not reliably. Using the same picture at both ends brought our clip back close to where it started, a similarity score of 0.98 out of 1 between its first and last frames, but a cloud still crossed the scene in the middle and the join is visible if you look for it. Blend the end into the start with the free loop video tool before you repeat it.
Why does my first and last frame video show ghosting?
Usually because the two frames are too different: another distance, another angle or new objects. The model has to invent the path and may cross-fade or pass things through each other. Keep the same subject, light and lens in both frames and change only one thing between them.
Is first and last frame the same as frame interpolation?
No. Frame interpolation adds in-between frames to footage that already exists, to make it smoother or slower. First and last frame generation creates a new clip from two stills, and the model decides everything that happens between them.
How we checked
paper-loop (20 September 2026) and glass-transition (8 September 2026) are the two clips recorded on Cicadas through fal.ai with the Veo 3.1 Fast first-last-frame endpoint, each sent once, US$0.80 each. On 5 November 2026 we exported frames at two per second (eight per second around the push-in) and compared the first and last frames with the supplied images using SSIM on 1280 by 720 frames. The outside tools' features come from their help pages found the same day. For a model that does not take an ending frame, see the MiniMax H3 Max guide; for how we record clips, how we test.