Camera control and motion — what separates slop from craft — step 5 of 8

The image-to-video keyframe pattern

Image-to-video is not just "cheaper than text-to-video." It's a different workflow that puts control where it belongs.

The two-step pattern

Step 1: Compose the first frame with gemini-3.1-flash-image via generate_content. (gemini-3-pro-image is the other live still ID.) A 1K still is ~$0.067. Iterate — try 3-5 variants until you have a frame you actually want. Total spend: ~$0.20-0.35 to lock the frame.

Step 2: Animate the chosen frame with a video model in I2V mode.

Hand the frame to Veo 3.1 (generate_videos / veo-3.1-lite-generate-preview or Standard) or Omni (client.interactions.create, video_config.task = image_to_video) with a motion-only prompt. The model already has the composition. It just has to solve motion. Acceptance rates jump, retake counts drop, and the video model spends its capacity on the thing it's good at.

You can skip step 1. One-shot Omni I2V is a valid call: no separate still in the pipeline.

Why this is the production default

Three reasons it beats text-to-video for most real work:

1. Brand asset control. Your client's product can be authored into the frame deterministically (image models are better at product fidelity than video models). The video model never gets to invent the product.
2. Sequence consistency. Use the same anchor character / product / style across multiple I2V generations and the shots cut together. T2V invents fresh every time; sequences fall apart.
3. Predictable budget. I2V acceptance rates run 2-3x retakes to one usable shot. T2V runs 5-10x. The cost gap is the retake drop, not the still.

An 8-second Veo 3.1 Lite clip is $0.40. Four 1K stills are ~$0.27. The seconds dominate. You stack a still because you want the frame. Imagen 4 (imagen-4.0-*) on the Gemini API died 2026-08-17; gemini-2.5-flash-image retires 2026-10-02.

The prompt structure changes

In T2V, your prompt has to describe everything — composition, lighting, subject, action, style, mood. The model fills in the gaps and the gaps are where the failures live.

In I2V, the frame already specifies composition, lighting, subject, and style. Your prompt only has to specify:

• The motion (shot type + camera move from previous reads).
• The duration.
• Any subject animation (eyes blink, hand moves, water pours, etc.).

A clean I2V motion prompt is short:

"Slow 3-second push-in. Subject's gaze drifts toward the window. Background stays static."

That's it. The frame already said everything else.

When I2V breaks down

I2V is the default but not universal. Three failure modes:

1. Camera move incompatible with the frame. If the frame is a tight extreme-close-up on a face, you can't ask for a wide parallax move — there's no scene depth in the frame to parallax against. Match the move to what's in the frame.
2. Motion the model can't infer. If the frame is a static pose with no implied motion (a person standing perfectly still), and you ask for "they wave their hand," the model has to invent the hand position over time and may break the character's identity. T2V is sometimes better when the action is complex.
3. Long durations. I2V's first-frame anchor weakens over time. By second 8-10, the model has drifted significantly from the frame. If you need 25 seconds of a specific character, cut multiple I2V clips. Sora 2's 16-20s jobs (three extends for a minute) go away 2026-09-24.

The mix-and-match angle

Different vendors are best at different parts of the I2V pipeline. You can mix:

• Frame in gemini-3.1-flash-image (~$0.067) for product fidelity.
• Animate in Veo 3.1 Lite ($0.05/sec × 8 = $0.40) or Runway gen4_turbo I2V at $0.05/sec.
• Or skip the still and one-shot Omni I2V at ~$0.10/sec.

Same shot, Veo 3.1 Standard T2V path:

• Frame implicitly in the video model (no separate frame spend, but the model picks the composition).
• 8 sec × $0.40 = $3.20 before retakes.

The T2V path looks simpler on paper. But the retake rate is 4-5x higher in T2V mode, so the real cost is $12-$16 per usable 8-second shot. Lite I2V at $0.40 with 2.5x retakes is $1.00 of video plus ~$0.07 of frame — and the frame is why you did it.

The mental shift

People coming from image generation prompt video models the same way: one big sentence with all the keywords. It doesn't work.

People who get good at AI video stop writing "scene descriptions" and start writing shot specifications: a frame

• a move + a duration, or one Omni I2V call with the image already in hand. That's the working pattern in 2026 and it gets more true every quarter.