# Reference Images for AI Video: First Frames, Last Frames and Character References

How to make the still images that control an AI video: the **first frame** it starts from, the **last frame** it has to land on, and the **reference images** that keep a character, product or location consistent. The advice works with any image generator (ChatGPT, Nano Banana / Gemini, Midjourney, Seedream / Dreamina, Flux and others) and any video model that accepts images.

## Three kinds of images, three different jobs

| Image | What it controls | Use it when |
| --- | --- | --- |
| **First frame** | Exactly how the clip opens: composition, identity, lighting, setting | You want the shot to start on a specific look |
| **Last frame** | Where the clip has to end up | The ending matters: a reveal, a finished product, a pose, a transformation, a hand-off to the next clip |
| **Reference image** | *What* things look like, not *where* the clip starts or ends | A character, outfit, product, location or style must stay consistent across many shots |

First and last frames are pinned: the model treats them as actual frames of the video. Reference images are guidance: the model borrows identity or style from them but builds its own shot.

**Where you'll find these (as of October 2026, so check your tool's current docs):**

- **Google Veo 3.1 (Flow, Gemini API):** *Frames to Video* takes a start and an end image. *Ingredients to Video* takes up to three reference images.
- **Gemini Omni 1.1 Flash:** first/last frame control.
- **Kling 3.0:** start and end frames, plus *Elements* (reusable characters and objects built from several images).
- **Seedance 2.0 / 2.5:** first/last frame generation, plus multi-image reference mode.
- **Grok Imagine:** image-to-video, plus *Extend from Frame* for continuing a clip.

## Rules for every image you feed a video model

1. **Match the aspect ratio** of the video you'll make (16:9, 9:16, 1:1). A cropped or letterboxed image wastes resolution and can confuse the framing.
2. **Leave room for the camera.** Keep about 10–15% of breathing space around the subject so the model can push in, drift or reframe without inventing a cut-off shoulder or a new edge of the set.
3. **One clear subject, readable silhouette.** Busy, cluttered images give the model too many things to animate.
4. **Light it on purpose.** Name the light source and direction (for example "warm window light from frame left") and keep it the same in every image of the same scene.
5. **No baked-in motion blur, text, captions, logos or watermarks**, unless you actually want them in the video.
6. **High resolution, sharp focus on the face or product.** Identity is pulled from the sharpest part of the image.
7. **Neutral but alive.** Pick a pose and expression the action can grow out of. A character mid-sneeze is a hard starting point.

## First frames

The first frame is your establishing shot. Build the prompt around the **visual facts that must persist**, and save motion, camera moves and sound for the video prompt.

```text
Create a [16:9 / 9:16] still frame to be used as the FIRST FRAME of an AI video.
[Shot size and camera angle, e.g. "medium shot, eye level, 35mm lens"].

Subject: [who or what, with age range, distinctive features, hair, wardrobe, pose, expression].
Setting: [specific location and time of day], [foreground / midground / background layers].
Light: [source, direction, color temperature, quality — e.g. "soft overcast daylight from the left"].
Style: [photoreal / animation / illustration medium], [palette], [lens and depth-of-field character].
Composition: subject placed [left third / centered], with breathing room around them for camera movement.

No text, no watermark, no motion blur.
```

## Last frames

The last frame is a **destination**. The model has to get there in the length of the clip, so the gap between the first and last frames is everything.

**Make the last frame by editing the first one, not by starting over.** Upload your first frame to the image tool and ask for a specific change. Starting from scratch almost always shifts the face, outfit, lens or lighting, and the video then "morphs" between two different people instead of moving.

```text
Edit this image to create the LAST FRAME of the same shot, [N] seconds later.
Change only: [the end state, e.g. "she has turned to face the camera and is smiling,
holding the open box at chest height"].
Keep exactly the same: face and identity, hair, wardrobe, props, location, lighting
direction, color grade, lens and aspect ratio.
[Optional camera change, e.g. "the camera has pushed in from a medium shot to a close-up"].
```

**What makes a good first/last pair:**

- **The change fits the time.** A 5-second clip can turn a head, open a door or push the camera in. It can't plausibly cross a city. Big changes need a longer clip or an extra keyframe in the middle.
- **One main change, maybe two.** For example, the subject moves *or* the camera moves a lot, not everything at once.
- **Same world.** Identical wardrobe, hairstyle, props, light direction and time of day, unless the change *is* the point (a costume transformation, a sunset).
- **Physically reachable.** If the subject's hand is empty in the first frame and holding a cup in the last, the cup has to come from somewhere visible.
- **For a seamless loop,** make the last frame the same as, or nearly identical to, the first.

## Reference images and character sheets

References keep things consistent across many clips. Make them once and reuse them for the whole project.

**Character turnaround:** one character, same outfit, neutral background, consistent light.

```text
Character reference sheet of [character description], same outfit in every view:
front, three-quarter front, profile, three-quarter back, and back, full body,
plus a head-and-shoulders close-up. Plain light-gray background, even soft studio
lighting, consistent proportions, no text labels, no props unless listed: [props].
```

**Expression sheet:** the same face showing neutral, smiling, angry, worried and surprised. This helps when a model struggles with emotion.

**Location plates:** the same set from two or three angles (wide, reverse, close detail) with the same light, so every shot in that scene feels like one place. Generate the characters for that scene *from* the location plate, so the lighting already matches.

**Product references:** clean front, side and three-quarter views, with true color and any labels sharp and legible.

Give each reference a single job in the video prompt, for example *"Image 1 controls her face and outfit only. Image 2 controls the kitchen layout and lighting only."* Say what should **not** carry over when two references could conflict.

## Writing the video prompt between the frames

When the frames carry the look, the prompt should carry the **motion**, so keep it short.

- Describe what happens between the frames: the action, its pace, and the camera move.
- Don't re-describe what's already visible in the images.
- Add sound if the model generates audio (Veo, Kling, Seedance, Grok Imagine and Gemini Omni all can).

```text
Over [N] seconds, [subject] [action, with pace and direction] while the camera
[movement, e.g. "slowly pushes in from medium shot to close-up"].
[Secondary motion: hair, fabric, background people, weather].
Sound: [ambience, effects, music, or dialogue in quotes].
Keep identity, wardrobe and lighting consistent. No cuts.
```

## Chaining clips into a longer scene

1. Use the **last frame of clip 1** as the **first frame of clip 2**. Export it at full resolution, not as a screenshot.
2. When you edit them together, **trim one frame** at each join. Otherwise the repeated frame shows up as a tiny freeze.
3. Every few clips, go back to your original references. Chaining many generations in a row slowly drifts faces and colors.

## Troubleshooting

| Problem | Likely cause | Fix |
| --- | --- | --- |
| The clip looks like a crossfade or morph | The first and last frames are too different | Make the last frame by editing the first; shrink the change; lengthen the clip; add a middle keyframe |
| The face changes partway through | Identity differs between images | Edit from one master image; use a sharper close-up reference |
| The camera jumps or cuts | The framing change is too large for the duration | Describe the camera move explicitly, or split it into two shots |
| Things pop into existence | An object is in one frame but not the other | Show where it comes from, or include it in both frames |
| The style drifts | No style anchor | Add a style reference, and state the medium ("keep the watercolor look") in the prompt |

## Sources

- [Images in ChatGPT — OpenAI Help Center](https://help.openai.com/en/articles/11084440-images-in-chatgpt)
- [Introducing Veo 3.1 and new creative capabilities in the Gemini API — Google Developers Blog](https://developers.googleblog.com/introducing-veo-3-1-and-new-creative-capabilities-in-the-gemini-api/)
- [Bringing new Veo 3.1 updates into Flow — Google](https://blog.google/innovation-and-ai/products/veo-updates-flow/)
- [Kling VIDEO 3.0 Model Guide — Kling AI](https://kling.ai/quickstart/klingai-video-3-model-user-guide)
- [Gemini Omni 1.1 Flash announcement — Google AI](https://x.com/GoogleAI/status/2093008998987403303)
- [Seedance first and last frame guide — seedance.tv](https://www.seedance.tv/blog/seedance-first-and-last-frame)

*Last updated October 2026. Feature names and limits change often, so check your tool's current documentation.*