First frame and last frame control means you give an AI video model two still images, one for where the shot starts and one for where it ends, and the model invents the motion in between. Your prompt describes the journey between them.
Keep the two frames matched in size, position and light so the model has a believable path to follow, and write the prompt about the journey, not the pictures themselves.
- Two pictures, one clip. The start image and the end image are fixed. Everything between them is the model’s guess, steered by your prompt.
- Several models offer it, including Veo 3.1, Kling, Seedance, Grok Imagine and Luma. Not every model does: Runway’s own Gen-4.5 takes a first frame only in its API.
- Describe the movement, not the frames. The images already show what things look like. Your words say how the camera and the subject get from one to the other.
- Match the two frames. Same aspect ratio, same lens feel, same light direction, same character. Big mismatches come back as a melt or a dissolve.
- Ask for less change than you think. A short clip can only cover a short distance. Split a big change into two clips.
Think of an animator who draws the first and last pose of a movement and leaves the drawings in between to an assistant. You are the animator. The video model is the assistant. It will always land on your last drawing, but how it gets there depends on how clearly you explain the move, and on whether the move is physically possible in the time it has.
Which AI video models let you set a start and an end frame?
The plain point: most of the big video models now take a start and an end image, but the names, limits and status differ, and some features are still labelled preview or early access. This table is what each vendor’s own documentation said on 12 October 2026. Check it again before you build a project on it.
| Model | Start + end frame? | What the docs say |
|---|---|---|
| Veo 3.1 (Google)Gemini API previews; generally available on Google Cloud | Yes, on Veo 3.1 and Veo 3.1 Fast | Google calls it interpolation. The end image is a lastFrame setting that only works together with a start image. In the Gemini API the models were listed as Preview (veo-3.1-generate-preview, veo-3.1-fast-generate-preview), and interpolation was not listed for Veo 3.1 Lite (read 12 Oct 2026). 1 Google says those Gemini API previews were shut down on 22 October 2026, with gemini-omni-1.1-flash as the replacement (read 3 November 2026). 23 On Google Cloud’s Gemini Enterprise Agent Platform, Veo 3.1 and Veo 3.1 Fast are generally available, and Google lists them, plus previews of Veo 3.1 Lite and Gemini Omni 1.1 Flash, as taking a first and a last frame. 4 5 |
| Kling 3.0 / 4.0 (Kuaishou)4.0 launched 15 Oct | Yes; 4.0 adds up to 10 keyframes | Kling lists “Start & End Frames-to-Video” for VIDEO 2.6 and 3.0. Kling says 4.0 rolled out to all users on 15 October 2026; its manual lists Kling 4.0 Flash, which has no start-and-end frames or keyframes, as coming soon. 6 7 |
| Seedance 2.5 (ByteDance)Dreamina, and via Runway’s API | Yes | ByteDance’s Dreamina guide tells you to give the model both frames on long clips. Runway’s API, which also hosts Seedance, calls first and last images “keyframe mode” and says it cannot be mixed with reference images in the same request. 8 9 |
| Grok Imagine Video 1.5 (xAI)API | Yes, on 1.5; not on the classic model | xAI’s docs describe a last_frame setting on grok-imagine-video-1.5 and say the older model rejects it. 10 |
| Ray 2 (Luma)API | Yes | Luma’s API examples set a start image (frame0), an end image (frame1), or both, under the heading “With start and end keyframes”. 11 |
| Gen-4.5 (Runway)API | First frame only, in the API | In Runway’s own API, Gen-4.5 takes a first frame and no last frame. The same API accepts first and last frames for Veo 3.1, but not a last frame on its own. 12 |
What a model can do in a company’s app and what it can do through its API are not always the same, and this table mostly reflects the developer docs. If your tool shows a “start frame” and “end frame” box, that is this feature, whatever it is called. Prompt formulas for each model are on the Veo, Kling, Seedance and Grok Imagine cheat sheets.
Start and end frames are widely available, but they are a newer feature: several of the models that offer it were still marked preview or early access on 12 October 2026. Read the vendor’s page for the exact name of the setting before you start.
How does first and last frame control work?
The plain point: the model treats your two images as fixed points and generates the frames that join them. It is a close cousin of image-to-video, where only the first frame is fixed, and of keyframes in animation.
With a first frame alone, the model has to decide where the shot ends up. With both frames, the destination is decided by you, so the model only has to solve the route. That is why this is the tool to reach for when an exact ending matters: a product that must land on its logo, a character who must end in the pose that the next shot starts from, or a transformation from one state to another.
Google’s own Veo 3.1 guide describes the job in one line, and notes that the transition comes with sound. 13
How to write the transition prompt
The plain point: the two images already say what everything looks like. Your prompt says how things move from the first picture to the second: the camera move, the subject’s action, the speed, and the sound. Google’s workflow says the same thing in its instructions for the feature. 13
- Name one camera move. “Slow push-in”, “180-degree arc around her”, “crane up”. One move gives the model a clear path; two moves give it a compromise. The vocabulary is on camera grammar.
- Name the action that connects the frames. If she is seated in frame one and standing in frame two, say “she stands up and turns to the window”. Do not leave the model to guess.
- Say the pace. “Slowly”, “in one smooth motion”, “steady”. Without one, a model with a fixed end point can rush to it and then hold.
- Say what stays still. “The background does not change. Nothing else moves.” This stops the model filling the gap with extra motion.
- Add sound if the model makes it. One line for effects or ambience. See sound and dialogue in AI video.
Write the verbs, not the nouns. The frames hold the look; your prompt holds the camera move, the action, the pace and what must not move.
How to keep scale, position and light matched
The plain point: the closer your two images are in everything except the thing you want to change, the smoother the clip. A model asked to join two frames that do not belong together will often fade one into the other instead of moving.
The reliable way to get two matching frames is to make the end frame from the start frame, with an image editor or an image model told to change one thing only. The prompts for that are on image edit prompts that change only one thing. If a person appears in both frames, lock their identity with the templates on identity lock prompt templates.
Before you generate, put the two frames side by side and check:
- Same aspect ratio and size. A square start and a wide end leaves the model to invent or crop. See aspect ratio and resolution.
- Same scale for the subject, unless the change in scale is the camera move you are asking for. A face that doubles in size needs a push-in in the prompt.
- A position the subject can reach. If the subject is on the left in frame one and on the right in frame two, the prompt must say how they cross, and the clip must be long enough for it.
- Same light direction and colour. Sun from the left in one frame and from the right in the other has no physical path. The model fakes one.
- Same lens feel. A wide-angle start and a telephoto end look like two different cameras. Keep the depth of field and perspective similar.
- Same costume, hair and props, down to small details. Anything that differs will change mid-clip, and that change is usually what gives the clip away.
Reasoning, 12 Oct 2026 — this checklist is this site’s own advice, drawn from how the feature is described above; it is not a vendor rule or a measured test.
Make the end frame from the start frame and change one thing. Two frames that match in everything else give the model one obvious path.
Common first and last frame failures, and the fix
The plain point: most failures come from asking for too much change, or from two frames that do not belong together. The fixes below are this site’s own working advice, not vendor documentation.
| You see | Likely cause | Try this |
|---|---|---|
| A dissolve or meltone image fades into the other | The frames differ too much to connect with real motion | Make the end frame closer to the start (edit one thing), name a concrete camera move, add “no dissolve, no cross-fade, continuous motion”. |
| Rush, then freezeit gets there early and holds | No pace in the prompt, or too little to do | Say “steady, even pace from start to finish”, give the subject an action that fills the time, or pick a shorter clip length. |
| The subject teleportsjumps across frame | The end position is too far for the clip length | Move the subject closer in the end frame, or split the move into two clips and chain them. |
| The face changes mid-clip | The two frames show slightly different faces | Generate the end frame from the start frame with an identity lock; attach the same reference where the model allows it. |
| The camera ignores you | Two or more moves asked for, or the move contradicts the frames | One move only, and check the frames actually show that move (a push-in needs a closer end frame). |
| An extra cut appears | The model read the change as a new shot | Add “one continuous shot, no cuts”. If it persists, reduce the difference between the frames. |
If a clip keeps failing in the same way, the general fault-finding guide is why your generation came out wrong.
Chaining shots: the last frame becomes the next first frame
The plain point: take the final frame of one clip and use it as the first frame of the next, and the two clips join without a jump. This is how a longer, continuous move is built from short generations.
Some models offer this as a built-in “extend” feature instead. Google’s Gemini API documents it for Veo 3.1. 1
Planning the frames in advance makes chaining much easier: if your storyboard already has the frame where shot one ends, it is also the frame where shot two begins. The storyboard and shot-list template are on plan before you generate. The same trick, with one image used as both the first and last frame, is the starting point for seamless loops.
For a long move, generate it in short pieces and hand each piece’s last frame to the next one, or use the model’s extend feature where it has one.
Questions people ask
Can I use only a last frame, with no start frame?
It depends on the model. Runway’s API says its hosted Veo 3.1 “does not support generating with only a last frame”, and Google’s own lastFrame setting must be used with a start image. 14
Is first and last frame the same as keyframes?
It is the simplest case of keyframes: two fixed points. Some newer models take more than two, such as Kling 4.0, which Kling says supports up to 10 keyframe images and rolled out to all users on 15 October 2026. 7
Do first and last frame clips have sound?
On models that generate audio, yes. Google describes Veo 3.1’s first and last frame transitions as “complete with audio”, and its Gemini API docs list audio as always on for the Veo 3 models. 13 1
How long should a first and last frame clip be?
Long enough for the move you describe, and no longer. Clip lengths are set per model: in the Gemini API, Veo 3.1 offers 4, 6 or 8 seconds, and must be 8 at 1080p or 4K. 1
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. In a first and last frame prompt, what should your words describe?
- How things move between the two pictures: camera move, action and pace
- What everything in the two pictures looks like, in detail
- A list of styles the model can choose from
Show the answer
Answer: How things move between the two pictures: camera move, action and pace. The page says the two images already show the look, so the prompt holds the verbs: the camera move, the action, the pace and what stays still.
2. Your clip comes back as one picture fading into the other. What is the likely cause?
- The prompt named too few colours
- The clip was set too long for the model
- The two frames differ too much to join with real motion
Show the answer
Answer: The two frames differ too much to join with real motion. The page says a dissolve usually means the frames do not belong together, so make the end frame closer to the start by changing one thing.
3. How does the page suggest building a longer, continuous move?
- Ask for the whole move in one very long prompt
- Use the last frame of one clip as the first frame of the next
- Pick two unrelated pictures so the model has more to do
Show the answer
Answer: Use the last frame of one clip as the first frame of the next. The page says chaining short clips this way joins them without a jump, or you can use a model’s extend feature where it has one.
Related on this site: what image-to-video means · what a keyframe is · anatomy of an AI video prompt · character consistency · text-to-video
- 1Google AI for Developers, Generate videos with Veo 3.1 in Gemini API, last updated 8 Oct 2026, read at source 12 Oct 2026, re-read 13 Oct 2026“Veo 3.1 lets you create videos using interpolation, or specifying the first and last frames of the video.” and, for
lastFrame: “The final image for an interpolation video to transition. Must be used in combination with the image parameter.” ↑a“Use Veo 3.1 to extend videos that you previously generated with Veo by 7 seconds and up to 20 times.” The same page lists extension for Veo 3.1 and Veo 3.1 Fast, “not Veo 3.1 Lite”. ↑b“Natively generates audio with video.”, listed as “Always on”. ↑cdurations “4”, “6”, “8”; “Must be “8” when using extension, reference images or with 1080p and 4k resolutions.” ↑d - 2Google, Gemini API deprecations, last updated 12 Oct 2026, read at source 13 Oct 2026veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview, shutdown “October 22, 2026”, recommended replacement “gemini-omni-1.1-flash”; the page notes that its dates “indicate the earliest possible dates on which a model might be retired”. ↑
- 3Google, Video generation in the Gemini API, read at source 3 Nov 2026“Veo 3.1 Preview (veo-3.1-generate-preview, veo-3.1-fast-generate-preview, and veo-3.1-lite-generate-preview) in the Gemini API was shut down on October 22, 2026; migrate to Gemini Omni Flash (gemini-omni-1.1-flash) or the GA models available through the Gemini Enterprise Agent Platform to avoid service interruptions.” ↑
- 4Google Cloud, Veo 3.1, Gemini Enterprise Agent Platform documentation, last updated 10 Oct 2026, read at source 13 Oct 2026veo-3.1-generate-001 and veo-3.1-fast-generate-001, “Launch stage: GA”, retirement date “November 17, 2026 or later”; veo-3.1-lite-generate-001, “Launch stage: Preview”. ↑
- 5Google Cloud, Generate videos using first and last video frames, Gemini Enterprise Agent Platform documentation, last updated 12 Oct 2026, read at source 13 Oct 2026“The following models support generating videos from first and last video frames”: gemini-omni-1.1-flash-preview (preview), veo-3.1-lite-generate-001 (preview), veo-3.1-generate-001 and veo-3.1-fast-generate-001. ↑
- 6Kling AI, Kling VIDEO 3.0 Model User Guide, 6 Feb 2026, read at source 12 Oct 2026capability row “Start & End Frames-to-Video”. ↑
- 7Kling AI, Kling 4.0 vs 3.0, 30 Sep 2026, read at source 12 Oct 2026“Kling 3.0 supports Start & End Frames, giving creators control over how a video begins and finishes.” and “Kling 4.0 adds Multiple Keyframes, supporting up to 10 keyframe images.” · Kling AI, Kling 4.0 at Work, 27 Oct 2026, read at source 3 Nov 2026: “Since Kling 4.0 rolled out to all users on October 15, creative teams across film, advertising, and animation have been putting these capabilities to work.” · Kling AI API documentation, Kling 4.0 Video User Manual, read in a browser 3 Nov 2026: “4.5 Kling 4.0 Flash (Coming Soon)”; “First-and-last-frame and multi-keyframe inputs are not currently supported.” ↑a“Kling 4.0 adds Multiple Keyframes, supporting up to 10 keyframe images.” · Kling AI, Kling 4.0 at Work, 27 Oct 2026, read at source 3 Nov 2026: “Since Kling 4.0 rolled out to all users on October 15, creative teams across film, advertising, and animation have been putting these capabilities to work.” ↑b
- 8Dreamina (ByteDance), Seedance 2.5 prompt guide, 18 Aug 2026, read at source 12 Oct 2026“Start/end frames: give the model both so the full 30 seconds doesn’t wander off compositionally.” ↑
- 9Runway, API reference, image to video, model
seedance2_5, read at source 12 Oct 2026“Use position `first`/`last` for keyframe mode, or omit position for reference images. The two modes cannot be mixed.” ↑ - 10xAI, Video generation, read at source 12 Oct 2026“Classic `grok-imagine-video` still rejects `last_frame` and `keyframes` and rejects combining `image` with reference inputs.” ↑
- 11Luma AI, Video Generation (Python), read at source 12 Oct 2026example headings “With start frame”, “With ending frame” and “With start and end keyframes”, using
frame0andframe1. ↑ - 12Runway, API reference, image to video, read at source 12 Oct 2026for
gen4.5, “An array of objects representing images to be used in the output video. Only a `first` frame is supported.”; forveo3.1, “You may specify an image to use as the first frame of the output video, or an array with a first frame and optionally a last frame. This model does not support generating with only a last frame.” ↑ - 13Google Cloud Blog, The ultimate prompting guide for Veo 3.1, 16 Oct 2025, read at source 12 Oct 2026“Generate a natural video transition between a provided start image and end image, complete with audio.” ↑a“Input both images into Veo using the First and Last Frame feature. In your prompt, describe the transition and the audio you want.” ↑b“Generate a natural video transition between a provided start image and end image, complete with audio.” ↑c
- 14Runway, API reference, image to video, model
veo3.1, read at source 12 Oct 2026“This model does not support generating with only a last frame.” ↑
Kling AI, Kling VIDEO 3.0 Model User Guide, 6 Feb 2026, and Kling 4.0 vs 3.0, 30 Sep 2026, both read at source 12 Oct 2026.
Runway, API reference (image to video), read at source 12 Oct 2026. Runway’s help-centre guides refused an automated fetch (HTTP 403) on 12 Oct 2026, so this page uses only its API reference.
The animator comparison, the prompt templates, the matching checklist and the failure table are this site’s own, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK