What is the difference between text-to-video and image-to-video?

Text-to-video generates a clip from a written prompt alone, so the model decides composition, subject appearance, colour and motion. Image-to-video starts from a supplied first frame, so composition and appearance are fixed and only motion is generated. Image-to-video needs meaningfully fewer attempts to produce a usable shot, which usually makes it cheaper in practice despite identical per-second pricing.

What is scene extension in AI video?

Scene extension continues an already-generated clip rather than producing a new one, letting you build sequences longer than a model's single-generation limit. Gemini Omni 1.1 Flash extends in 10-second increments up to 40 seconds, and its August 2026 update analyses up to ten seconds of prior footage when extending instead of only the final second, which keeps characters and motion consistent across joins.

Which video generation mode is cheapest?

All three usually bill at the same per-second rate, so the cost difference comes entirely from how many attempts each needs. Image-to-video and keyframe-locked generation typically need two to three attempts per usable shot against three to five for open text-to-video, making constrained modes roughly 40% cheaper for the same output.

When should I use scene extension instead of generating a longer clip?

Use extension when a sequence is built and revised shot by shot, because fixing one segment does not make you re-pay for the good ones. Use a model with a long native generation, such as Wan 3.0's 30-second single pass, when you need one genuinely continuous take and cannot tolerate visible joins.

Quick Answer

Text-to-Video vs Image-to-Video vs Scene Extension

Published: August 28, 2026

The Short Answer

These are three points on one axis: how much you constrain the model.

Mode	You supply	Model decides	Attempts per usable shot
Text-to-video	a prompt	everything	3-5×
Image-to-video	prompt + first frame	motion only	2-3×
Keyframe-locked	prompt + first + last frame	the path between	2×
Scene extension	prompt + existing footage	continuation	1-2× per segment

Since all four typically bill at the same per-second rate, the constraint level is what actually determines your bill. More constraint, fewer attempts, lower cost — and less creative surprise.

Text-to-Video — Maximum Freedom, Maximum Variance

You write a prompt. The model chooses composition, subject appearance, lighting, colour grade, camera movement and motion.

Use it for: exploration, mood-finding, concept work, anything where you do not yet know what you want.

The cost trap: every attribute the model chooses is an attribute that can be wrong. A prompt specifying four things — subject, action, setting, camera move — will typically get at least one wrong per attempt. That is where 3-5× re-roll factors come from, and why teams are surprised by video invoices.

The consistency trap: two text-to-video generations from the same prompt produce two different-looking subjects. Any multi-shot sequence built purely this way will not hold a character.

Image-to-Video — the Default for Production Work

Supply the first frame. Composition, subject appearance and colour are now fixed inputs, not model decisions. Only motion is generated.

Use it for: most commercial output. Product shots, character sequences, anything with a brand or a specific look.

Why it is cheaper without a discount: you have removed the majority of failure modes. Re-roll factors typically drop to 2-3× — a 40% cost reduction against text-to-video at identical per-second pricing.

The prerequisite: you need a first frame. Generating one with an image model is fast and cheap relative to video, and gives you a still you can iterate on at a fraction of video cost before any video is generated at all.

Keyframe-Locked — Deterministic Camera Moves

Supply both a start and an end frame; the model generates the transition. Gemini Omni 1.1 Flash added this in its August 27, 2026 release.

Use it for: specified camera movements, before/after reveals, transitions between known states, anything where the destination is not negotiable.

This is the tightest constraint available on a single generation and produces the lowest attempt counts — often 2× or better. The trade is obvious: you have converted a creative decision into an input, which is exactly what you want for a client deliverable and exactly what you do not want during exploration.

Scene Extension — Length Without Re-Paying

Extension continues an existing clip rather than generating a new one, letting sequences exceed the model’s single-generation ceiling.

The August 2026 change that made this usable: Gemini Omni 1.1 Flash’s extension step analyses up to ten seconds of existing footage rather than only the final second, extending in 10-second increments to a 40-second ceiling. The longer look-back is why extended footage now stays consistent instead of drifting away from the characters and motion you started with.

The economic argument is the real one. Compare two ways to fix segment two of a three-segment sequence:

Approach	What you pay for
Regenerate the whole sequence	segments 1, 2 and 3 again
Extend / re-extend from segment 1	segment 2 onward only

On long sequences with normal revision cycles, that difference compounds quickly.

Its limit: joins. Extension is stitching, however good the stitching is. For one genuinely continuous take with no join risk, a model with a long native generation is the better answer — Wan 3.0 produces 30 seconds with audio in a single pass.

Choosing, in Order

Do you know what the shot should look like? No → text-to-video at draft resolution to explore. Yes → keep going.
Do you have or can you make a first frame? Yes → image-to-video. This should be your default for delivery work.
Is the ending also specified? Yes → lock both keyframes.
Is it longer than one generation allows? Yes → extension if it will be revised, native long generation if it must be seamless.

Cost Worked Example

A 30-second 1080p sequence on Gemini Omni 1.1 Flash at $0.15/sec delivery, $0.03/sec draft:

Approach	Attempts	Cost
Text-to-video, all at 1080p	4×	$18.00
Text-to-video, draft-first	3 draft + 1 final	$7.20
Image-to-video, draft-first	2 draft + 1 final	$6.30
Keyframe-locked + extension	1 draft + 1 final per 10s	~$5.40

Same deliverable, 3.3× spread. None of it came from a vendor discount.