Generative Media · Personal experiment

Football anime

A small episodic series testing recurring characters, scene control and visual continuity across separate video generations.

The point was not the story itself, but repeatability. Each short scene creates another chance for identity, timing, composition or style to drift. That makes the sequence useful for seeing what current tools can hold consistently and where review and iteration are still required.

Selected clips

Generated with Google Veo. These three clips are selected outputs from iterative personal experiments with recurring characters and scene continuity. The clips include audio and are shown as examples, not controlled benchmark results.

EP02

What is this?

EP03

Why is he walking?

EP04

You’re late.

What I was testing

The useful signal is not whether one clip looks good. It is which parts of intent remain stable when the task is repeated.

01 · IdentityCharacter continuityKeeping recurring characters recognisable across separate generations.
02 · CompositionScene controlMaintaining intended action, relative placement and scene details.
03 · TimingDialogue timingKeeping short spoken lines aligned with pacing and visual focus.
04 · Visual languageStyle consistencyPreserving the same anime treatment, tone and overall visual identity.

What I observed

Across these three selected clips, Veo was notably consistent at the level of overall character identity and visual language, while smaller details and exact composition still required review.

01 · IdentityCharacter continuity heldThe two recurring characters stayed recognisable across separate generations, even though smaller facial, hair and uniform details were not identical.
02 · Scene controlCore scene intent heldThe intended situation and character roles remained readable, while framing, background actors and small scene details stayed generative rather than fixed.
03 · TimingShort dialogue remained usableThe short spoken lines worked with the pacing of these selected clips, but timing still needs qualitative review rather than being assumed from the prompt.
04 · Visual languageOverall anime treatment stayed consistentThe same anime-like visual language and tone carried across the three clips, while lighting, proportions and costume details still varied.
05 · EvaluationExperiment depth has a costRepeatability only becomes visible across multiple generations. Access, credit and pricing limits therefore constrain how confidently different tools can be compared.
Evidence boundary: these are personal generative-media experiments, not a production pipeline or a measured quality benchmark. The evidence is narrower: repeated generations make control failures visible and show where manual review, reference management and iteration still matter.
Tooling boundary: I also explored Seedance, Kling, Google Flow and HeyGen. Access, credit and pricing constraints limited how many repeated generations I could run with each tool, so I did not have enough comparable evidence for a controlled model comparison. The observations on this page therefore relate only to the Google Veo outputs shown here. Other paid workflows may provide stronger consistency or control, but I did not test them deeply enough to make that claim.
Football anime · 2026Three selected clips from a continuing personal experiment.
Back to Generative Media →