Why are AI videos only a few seconds long?
Video length: a clip is many frames, so many patches; attention compares every pair, so the work grows with the square of the length, and keeping a character the same across time is hard, so tools make short clips and extend or stitch them · Beginner · 5 min · attention, sampling · after Diffusion (denoising image models), Self-attention
A video model works on patches of every frame, and attention compares every patch with every other, so doubling a clip is about four times the work. Characters also drift, so tools make short clips and extend or stitch them.
1 Guess first
2 Look
Why each extra second costs more
1 / 4One picture: 1 frame
about 2,300 patches (calculated)
A 16-second clip: 32 frames
73K patches
Movie Gen: 73K tokens = 16 s (Polyak et al.); per-frame count calculated
A video is many frames, so many patches. In Movie Gen, 16 seconds is 32 compressed frames of about 2,300 patches each: 73K tokens, where one picture needs one frame’s worth.
A clip is a stack of frames, each full of patches; every extra second adds patches, every patch must be compared with all the others, and separate pieces forget what the character looked like.
A video model sees a clip the way an image model sees a picture: as patches, small squares of each frame. Many frames mean many patches. To keep the scene coherent, attention lets every patch look at every other patch, across space and time, so the work grows with the square of the patch count, not in step with it.
4 You are the model
You are the video model. Someone wants a longer clip, then a second clip from the same text, then an extension of the first one.
round 1 of 3 · You made a 16-second clip in one go. Now someone wants 32 seconds, still in one go.
How much more attention work is that?
pairs calculated from Movie Gen; clips illustrative
You are the video model. Pick what most likely happens.5 Check
Length also costs consistency. A character stays the same only if every detail stays tied across hundreds of frames. Clips made separately share nothing but the prompt; even one long clip can drift. Movie Gen’s own team, upscaling video in overlapping windows, saw “noticeable inconsistencies at the boundaries”.
So tools cap one generation at seconds: Veo 3.1 at 8, Runway Gen-4.5 at 10, Kling 3.0 at 15, Wan 3.0 at 30. Longer videos are built from pieces: extend a clip from its last second, start the next clip from the last frame, or make short shots and edit them together, as films do.
8 Say it back
9 Mission · 5 min
On Runway’s free plan, make a 3-second clip: “a woman in a red coat and yellow scarf walks past a blue door”. Save its last frame as an image and start a second 3-second clip from it. Make a third from the text alone. Play them in order: where do the coat, scarf or face change?
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Go deeper
Sources
- Polyak et al., “Movie Gen: A Cast of Media Foundation Models” (arXiv:2410.13720, v2 2025-02-26, https://arxiv.org/abs/2410.13720, checked 2026-10-23): “trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second”; §3.1.1 the temporal autoencoder compresses “8×” across each spatio-temporal dimension; §3.1.6 “Due to the quadratic nature of self-attention”; §3.1.5 upsampling “in a sliding window fashion resulting in noticeable inconsistencies at the boundaries”
- Runway API OpenAPI spec (https://docs.dev.runwayml.com/openapi.json, checked 2026-10-23), `duration` per model: gen4.5 2–10 s, veo3.1 4, 6 or 8 s, kling3.0_pro 3–15 s, hailuo3 5–15 s, wan3 2–30 s
- Google, Gemini API “Generate videos with Veo” (https://ai.google.dev/gemini-api/docs/veo, checked 2026-10-23): Veo 3.1 `durationSeconds` 4, 6 or 8; “Extend finalizes the final second or 24 frames of your video and continues the action”; extend “by 7 seconds and up to 20 times”, “for up to 148 seconds of video”
- Runway pricing (https://runway.com/pricing, checked 2026-10-23): Free plan “125 one-time credits”; “Gen-4.5 uses 12 credits per second of generated video”
- Calculated, not measured: 16 s × 16 fps = 256 frames, ÷ 8 = 32 compressed frames, 73,000 ÷ 32 ≈ 2,300 patches each; pairs = patches²: 73,000² ≈ 5.3 billion, 146,000² ≈ 21 billion (32 s), 273,750² ≈ 75 billion (60 s), (60 ÷ 16)² ≈ 14; real models also compress and restrict attention, so these show the growth, not their bills
- The four clips and their looks in the picture, the role wording, the ad example and the mission prompt are illustrative, not measured
- aifree.dev catalog, runway-free-plan (verified 2026-10-12, read-only API check 2026-10-23): 125 one-time credits, a selection of models, watermark on free exports
Published Oct 23, 2026 · updated Oct 23, 2026