First and Last Frame Video: Planning a Controlled Shot

You've got a clean product shot of a coffee bag on a counter, and a second image of the same bag with a hand lifting it. You load both as start and end frames, hit generate, and get a clip where the counter melts, the bag grows a second label, and the hand appears out of the wood grain. Or worse: the clip simply cuts from one scene to the other halfway through.

Neither is a random glitch. The model was asked to bridge two images that were too different, with a prompt that didn't say how. A first last frame ai video is a plan, not a lucky draw. This post covers how to build that plan.

What first and last frame actually does

Normally, image-to-video gives the model one still and asks it to invent where the motion goes. With first and last frame, you pin both ends of the clip. The model fills in the middle.

Google's documentation describes it for Veo 3.1 as specifying the first and last frames of the video, and the Veo 3.1 announcement says it can generate the transition between them, complete with accompanying audio. Kling has the same idea, which it calls Start and End Frames. Seedance models take it too: in the Runware docs for Seedance 2.5, one frame animates forward, and two frames interpolate between them.

What you gain is control over where the shot lands. What you give up is some of the model's freedom, which is exactly why mismatched frames go wrong.

Rule 1: The two frames must share a scene

Kling's own guidance is blunt. It says to choose two similar images with the same theme, and warns that significant differences can cause a lens switch, meaning a cut instead of a smooth move. The Seedance docs say something similar: frames should share their composition so the model sees genuine motion rather than a morph.

In practice, keep these constant between frames:

• The location and background
• The lighting direction and time of day (unless the lighting change is the point)
• The subject's identity, outfit and colour
• The lens look, meaning how wide or tight the framing feels

Change only one or two things: the subject's position, the camera distance, the light, a product's state.

Good pairs:

• A wide shot of a bakery counter, then a closer shot of the same counter with a croissant in the foreground.
• A candle unlit, then the same candle lit, same framing.
• A character facing left, then the same character turned toward camera.

Bad pairs: a studio portrait, then the same person on a beach. The model has no continuous path between those, so it cuts, dissolves or morphs.

Rule 2: Make the frames in one session

The easiest way to get matching frames is to generate the second from the first. Make your start image, then ask an image model to edit it ("same kitchen, same lighting, the mug is now in her hand"). Kling's guide notes that many users generate several images first, then pick two comparable ones.

Generating each frame from a fresh prompt almost guarantees small differences in the background, colour and proportions. Those differences show up as shimmer or a slow morph.

If you want the same character across both, use a saved character or a Seed (a set of reference images) so the face doesn't drift. In Mediafy Studio you can type @ in the prompt to bring one in. It helps, but it doesn't guarantee the two frames match in every detail, so check them side by side.

Rule 3: Match the frame format

Aspect ratio and size need to agree. The Runware Seedance docs say the output dimensions come from the first frame, and that the first and last frames should match in dimensions to avoid stretching. Veo 3.1's documented aspect ratios for this are 16:9 (the default) and 9:16, so crop both frames to one of those before you upload.

A few other constraints I read in the docs, so you can plan around them:

• Seedance requires a first frame. A last frame on its own is rejected, and a request takes at most two frames.
• With Veo 3.1, the Gemini API docs say duration must be 8 seconds when you use 1080p or 4k resolution (and with extension or reference images).
• Kling's guide frames the feature around smooth transitions within 5 seconds.

These specs change between model versions, so check the current limits for whichever model you pick.

Rule 4: Write a prompt for the middle, not the ends

The two images already show the beginning and end. Your prompt should describe what happens between them. Google's Veo docs advise detailed prompts with descriptive adjectives and adverbs, covering camera movement, action and mood.

Compare:

• Weak: a woman in a kitchen
• Better: She slowly reaches across the counter, picks up the white mug, and brings it toward her face with a small smile. The camera stays locked off. Soft morning window light, steady throughout.

Notice what the better prompt does:

1. It names one action with a clear path from frame A to frame B.
2. It states what the camera does (here, nothing), so the model doesn't invent a move.
3. It tells the model what should not change: the light.

If you're stuck on wording, Mediafy Chat can help you brainstorm and draft the motion description, and the Enhance button in the Studio can expand a short prompt. Read the result before you run it, since an enhanced prompt sometimes adds camera moves you didn't want.

Rule 5: Check that the move is physically plausible

Models interpolate better when there's an obvious straight-line path. Trouble spots:

• Big pose changes. A person sitting to standing with a full turn gives the model many ways to get it wrong. Hands and limbs may bend strangely.
• Objects that must appear from nowhere. If the cup isn't in frame A but is in frame B, the model has to invent it. Show it at the edge of frame A instead.
• Text and logos. They tend to warp in the middle of a transition even when both end frames are sharp. Keep them out of the moving area where you can.
• Cuts disguised as moves. If you really want a scene change, make two clips and join them in an editor. Don't force it through one generation.

Results also vary run to run. The same two frames and prompt can give a clean clip once and an odd one the next time. Plan to generate more than once.

A cheap testing routine

Video costs more than images, so don't find your problems at full price.

1. Draft frame A with a fast, cheap image model. In Mediafy Studio a Nano Banana image is 3 credits.
2. Edit frame A into frame B and place them side by side. Ask: could a camera and a person plausibly get from A to B in a few seconds?
3. Pick a video model that offers an end-frame option. Models such as Veo 3.1 and Kling have been documented with it, though what's available in any app changes, so check the model's options in the Studio.
4. Look at the credit cost on the Generate button before you run it. Failed generations are refunded automatically.
5. Run one test, watch the middle of the clip (the ends will look right by design), and adjust the prompt or the frames, not both at once.

If the middle morphs, the usual fix is a closer second frame. If the model cuts, the frames are too different. If the motion is right but stiff, tweak the verbs and adverbs in the prompt.

When not to use it

First and last frame is a good fit for product reveals, simple character moves, light changes, and loops (use the same image as both ends, with care, since the model may freeze or drift). It's a poor fit for long, multi-step action, where you'll get better results chaining shorter clips. If you build repeatable sequences, a Flow on the Studio canvas can chain image then video steps in one run.

New to the Studio? The first image tutorial takes five minutes, and the AI Creator Lab course goes deeper on workflows.

Ready to try a controlled shot of your own? Sign in to Mediafy Studio.

Sources

• Google AI for Developers, "Generate videos with Veo 3.1 in Gemini API": https://ai.google.dev/gemini-api/docs/veo
• Google Developers Blog, "Introducing Veo 3.1 and new creative capabilities in the Gemini API": https://developers.googleblog.com/en/introducing-veo-3-1-and-new-creative-capabilities-in-the-gemini-api/
• Kling AI, "Start and End Frames": https://kling.ai/quickstart/ai-video-start-end-frames
• Runware, "Seedance 2.5 image to video" guide: https://runware.ai/docs/models/bytedance-seedance-2-5/guides/image-to-video