Start and end frames in AI video, by date
30 entries on start, end or middle frames from 14 of the 17 lines, 29 of them dated 29 July 2024 to 31 August 2026. The table lists each with its line and version; the chart counts them per line. As of 2026-09-25.
1Reading the notes
Frame notes let a user fix how a clip starts, ends or passes through. Runway's Gen-3 Alpha took a first frame in July 2024 and a last frame a week later; Gen-3 Alpha Turbo added both together and then a middle keyframe. Firefly, Kling, Wan, Veo, Midjourney and Vidu add start and end frames during 2025, often on one tier or mode first, as with Kling 2.1's pro mode only. In 2026 Kling allowed set frames inside multi-shot videos and pulled keyframes out of existing footage, and PixVerse's V6 note in March lists first and last frame among the inputs it supports.
| Line | Entries |
|---|---|
| Adobe Firefly Video | 2 |
| Gemini Omni | 2 |
| Google Veo | 3 |
| HappyHorse | 1 |
| Kling AI | 5 |
| LTX | 1 |
| Luma Ray | 2 |
| Midjourney video | 1 |
| MiniMax Hailuo | 1 |
| PixVerse | 2 |
| Runway | 5 |
| Seedance | 1 |
| Vidu | 1 |
| Wan | 3 |
Inclusion rule. Lines with at least one entry of this kind. Order. Alphabetical by line.
2Every entry
| Date | Line | Version | What changed |
|---|---|---|---|
| 29 July 2024 | Runway | Gen-3 Alpha | Gen-3 Alpha takes an image as the first frame |
| 5 August 2024 | Runway | Gen-3 Alpha | The image can be the first or the last frame |
| 8 October 2024 | Runway | Gen-3 Alpha Turbo | First and last frame input for Gen-3 Alpha Turbo |
| 19 December 2024 | Runway | Gen-3 Alpha Turbo | A middle keyframe between first and last |
| 20 January 2025 | Wan | Wan 2.1 | Wan 2.1 image-to-video Plus turns a still into a clip |
| 12 February 2025 | Adobe Firefly Video | Firefly Video Model | Keyframes can be set at the start and at the end of a clip |
| 31 March 2025 | Kling AI | Kling 1.6 | Kling 1.6 can generate from an end frame |
| 7 April 2025 | Wan | Wan 2.1 | Wan 2.1 start-and-end-frame model makes a transition between two images |
| 23 June 2025 | Google Veo | Veo 2 | Veo 2 on Vertex AI takes a last frame as well as a first, or a video to extend |
| 25 July 2025 | Midjourney video | Midjourney V1 video | Looping, set start and end images, and video in the Discord bot |
| 31 July 2025 | Google Veo | Veo 3 | Image-to-video opens for the Veo 3 preview and a Veo 3 Fast preview is released |
| 5 September 2025 | Kling AI | Kling 2.1 | Kling 2.1 takes a start and an end frame, in pro mode only |
| 12 September 2025 | Wan | Wan 2.2 | A Wan 2.2 first-and-last-frame model for transitions between two images |
| 24 September 2025 | Vidu | Vidu Q2 | Vidu Q2 Pro and Q2 Turbo for image-to-video and start-and-end frames |
| 15 October 2025 | Google Veo | Veo 3.1 | Veo 3.1 preview adds extension, up to three reference images and first-and-last-frame input |
| 9 November 2025 | PixVerse | Not named in the note | Multi-transition video and Video Edit are released |
| 17 November 2025 | Kling AI | Kling 2.5 Turbo | Kling 2.5 Turbo pro mode takes first and last frames |
| 16 December 2025 | Adobe Firefly Video | Firefly Video Model | A start frame can be paired with a reference video whose camera motion the generation recreates |
| 18 December 2025 | Luma Ray | Ray3 | Ray3 Modify adds keyframe and character reference controls |
| 21 January 2026 | Runway | Gen-4.5 | Gen-4.5 takes a first frame image with the prompt |
| 30 January 2026 | Kling AI | Not named in the note | Keyframes pulled from a video can serve as start and end frames; .mov export added |
| 31 March 2026 | PixVerse | PixVerse V6 | V6 covers text-to-video, image-to-video, first and last frame and extend |
| 7 May 2026 | Kling AI | Not named in the note | Multi-shot videos can have set first and last frames |
| 9 June 2026 | Luma Ray | Ray3.2 | Ray3.2 adds frame-level control over how action moves |
| 16 June 2026 | HappyHorse | HappyHorse 1.1 | HappyHorse 1.1 image-to-video names consistency across clips |
| 19 August 2026 | LTX | Not named in the note | Audio-to-video takes a frame rate, a last frame and camera motion |
| 27 August 2026 | Gemini Omni | Gemini Omni 1.1 Flash | Gemini Omni 1.1 Flash goes GA with extension, first-and-last-frame input and 360p to 4K output |
| 28 August 2026 | Gemini Omni | Gemini Omni 1.1 Flash | Gemini Omni 1.1 Flash on Higgsfield, up to 10 seconds and 4K |
| 31 August 2026 | MiniMax Hailuo | MiniMax H3 Max | MiniMax H3 Max, a variant tuned for prompt adherence, on Higgsfield |
| Undated version row | Seedance | Seedance 1.0 Pro | Seedance 1.0 Pro: 2 to 12 seconds, up to 1080p, first and last frames |
Inclusion rule. All entries tagged with this kind of change, including platform arrivals and undated rows. Order. By date, oldest first; undated rows last.
3One line at a time
4Using the models, not reading about them
Release notes such as “Start, end and middle frames” matter once a pipeline runs the models; SceneMixer's guides describe how it does:
The shots add up to more than the selected video tier allows for one segment.
SceneMixer, Shot Seconds: Changing Timing, and the Limits — “What does 'over cap' mean on the segment card?” (read 2026-10-13)
Wan 3.0 · 480P is the tier every new project starts on.
SceneMixer, Wan 3.0 · 480P: the default video tier (read 2026-10-13)