What is First / last frame? Definition, key points and related tools
Images that pin a clip's opening or closing frame. First frames give continuity between segments, but on some models (e.g. H3) they exclude character references — continuity costs you the cast lock, the mechanical cause of blocking drift.
What pinning a frame buys you
A first frame is reproduced at the pixel level; Alibaba describes Wan 3.0's frame modes in exactly those terms. Add a last frame and the model generates the motion in between, which Google's Veo 3.1 documentation calls interpolation. In a drama pipeline that gives three uses:
- Chaining. The last frame of segment N becomes the first frame of segment N+1, so position and motion continue across the join. BytePlus documents a switch that serves this: Seedance's return_last_frame returns the final frame as a PNG at the video's own pixel size, without a watermark.
- Planned end states. A cliffhanger freeze or an exact final pose is easier to hit when the frame is supplied rather than described.
- Transitions. A controlled move between two views of one place, the case Google's Veo 3.1 prompting guide demonstrates with a 180-degree arc around a singer.
The trade-off, model by model
On several models, pinning a frame switches off the reference inputs that hold the cast together. What each vendor documents (checked 2026-09-12):
| Model | Frame modes | Combine with references? |
|---|---|---|
| MiniMax H3 | first, last, or both | No. Any reference image, video or audio role forbids frames, so a frame-pinned segment also loses its bound voice sample |
| Seedance 2.0 / 2.5 (BytePlus) | first, or first + last | No. First frame, first + last and multimodal reference are mutually exclusive scenarios |
| Wan 3.0 | one first, one last | No. Frames cannot be mixed with reference images, video or audio |
| Kling VIDEO 3.0 | start and end frames | Yes. Up to 3 additional Element Library elements can be bound |
| Veo 3.1 (Gemini API) | image plus lastFrame | Documented as a separate feature from referenceImages (up to 3), with no combined example; in both, person generation is limited to adults |
| Vidu | start-end-to-video endpoint, two images | Separate endpoint from reference-to-video |
The H3 row is easy to miss: the exclusion covers reference audio as well as images, so continuity by frame costs both the face lock and the voice lock (see reference audio).
Using frames without losing the cast
- Chain the action, then re-anchor. Use a frame where one continuous action is split across two segments, and return to reference mode for the next one; the character consistency explainer covers why error accumulates along a chain.
- Take the frame at full size, not as a screenshot. A player screenshot is often scaled and adds its own compression, which the next segment then reproduces. Use the model's own last-frame output where offered (Seedance's PNG at the video's pixel size) or extract the final frame from the delivered file at its native resolution.
- Match the geometry. In frame modes the image decides the output shape: H3 treats the ratio as adaptive and ignores other values, Wan recommends adaptive, and Seedance 2.5 keeps the first frame's aspect ratio, so a 16:9 frame in a 9:16 series yields a 16:9 clip. Where a different ratio is requested, BytePlus documents a centred crop of your image, which can cut off a character standing near the edge. Vidu also requires the two frames' aspect ratios to be within 0.8 to 1.25 of each other.
- Keep the two frames close. Kling advises two similar images on the same theme for transitions within about 5 seconds and warns that large differences may trigger a shot switch. The movement between the frames also has to be physically possible in the clip's seconds.
Common mistakes and their causes
- A frame on every segment for continuity. Identity then rests on the previous clip's last frame, so small drifts compound from segment to segment.
- A character sheet as the first frame. The clip opens on a white-background studio pose and takes the sheet's aspect ratio. Sheets belong in reference slots (see character sheet).
- Cuts inside an interpolation. A last frame asks for continuous motion toward one image. Write a frame-pinned segment as one continuous move and keep multi-shot storyboards in reference-mode segments.
- Losing the frame you need. Veo keeps generated videos for 2 days and Seedream image URLs expire after 24 hours; save the frame before the next segment needs it.
Key figures and sources
| Figure | Value | Source | Date |
|---|---|---|---|
| MiniMax H3 frames vs references | first_frame / last_frame not allowed with any reference_image, reference_video or reference_audio role | platform.minimax.io | 2026-09-12 |
| Seedance 2.x frame modes (BytePlus ModelArk) | exclusive with multimodal reference; return_last_frame gives a PNG at video size, no watermark | docs.byteplus.com | 2026-09-12 |
| Wan 3.0 frame modes | 1 first and 1 last frame; cannot be mixed with reference media | www.alibabacloud.com | 2026-09-12 |
| Kling VIDEO 3.0 start/end frames | up to 3 additional elements can be bound | kling.ai | 2026-09-12 |
| Vidu start-end-to-video | two images; their aspect ratios within 0.8–1.25 of each other | platform.vidu.com | 2026-09-12 |
FAQ
Can I pin a first frame and still attach character references?
It depends on the model. Kling VIDEO 3.0 binds up to 3 elements alongside start and end frames. MiniMax H3, Seedance 2.x and Wan 3.0 reject the combination outright, and Veo 3.1 documents frames and reference images as separate features.
Why did my first-and-last-frame clip cut instead of moving?
The two frames were too far apart for a continuous move in the time available. Kling notes that large differences may trigger a shot switch. Bring the frames closer (same location, similar framing and lighting), lengthen the clip, or split the change into two segments.
Is a last frame on its own useful?
Yes, where supported. MiniMax H3 accepts a last frame without a first, which fixes where a segment ends, for example a cliffhanger pose, while the model chooses the opening. The same exclusion applies: no reference images, video or audio in that call.
Sources
- https://platform.minimax.io/docs/api-reference/video-generation-v2-create
- https://docs.byteplus.com/en/docs/ModelArk/1520757
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-guide
- https://www.alibabacloud.com/help/en/model-studio/wan3-video-generation-api-reference
- https://kling.ai/quickstart/klingai-element-library-3-user-guide
- https://kling.ai/quickstart/ai-video-start-end-frames
- https://ai.google.dev/gemini-api/docs/veo
- https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1
- https://platform.vidu.com/docs/start-end-to-video