Thumbnails as a system
Scripted composition, variants and model ranking
- final export
- 2K
- of variants, v1 to v3
- 3 batches
- crop, exposure and type saved
- JSON
- predicted CTR across 5 variants
- 5.46 → 5.60%
- The problem
- The thumbnail decides the click: you need several variants true to the series, and a way to choose other than gut feeling.
- For whom
- The Chute du Monarque series and its spin-offs, which must stay recognisable from one episode to the next.
- The outcome
- 2K exports, three variant batches, settings saved as JSON and a predicted-CTR ranking that knows when to abstain.
Context
The thumbnail is not generated in one go. It starts from film frames or generated references, then gets composed against a fixed identity, so the series stays recognisable whatever the source image.
What I built
- Composition scripts (variants, hooks, 2K export) with settings saved as JSON.
- Generated references, then recomposed, ivory and gold type, fleur-de-lis.
- Variants ranked by The Oracle’s CTR model, abstaining when the gap is too small.
Architecture 7 blocks
- 01 · InputSourcesFilm frames or generated reference images
- 02 · ToolGradingExposure and tone curve by Python script
- 03 · ToolCompositionCrop, ivory and gold type, fleur-de-lis
- 04 · DataSettingsCrop, exposure and type saved as JSON
- 05 · ModelRankingPredicted CTR and uncertainty from The Oracle
- 06 · HumanA/BThe audience decides when the model hesitates
- 07 · Output2K exportFinal thumbnail in the series identity
Pipeline 5 steps
- SourcesFilm frames or generated reference images.
- CompositionCrop, exposure and type, by script.
- VariantsSeveral hooks and framings.
- RankingPredicted CTR and uncertainty.
- A/BThe audience decides when the model hesitates.
Decisions the choice, the reason, the alternative ruled out
- D1Compose the thumbnail rather than generate it in one go: source image, then a fixed identity.
The series stays recognisable whatever the starting image, and the text stays exact.
Ruled outA fully generated thumbnail, text included.
- D2Save every setting (crop, exposure, type) in a JSON file.
A variant can be redone or fixed identically, with no manual retouching to redo.
Ruled outUntracked manual retouching.
- D3Produce variants by script, in successive batches (v1 to v3).
Several hooks and framings can be compared under the same identity.
- D4Let the ranking model abstain when the gap is below its error margin.
On episode 2 the predicted gap (5.46 → 5.60%) was smaller than the uncertainty: answer “undecided, settle with A/B”.
Ruled outPicking the variant with the best raw score.
Deliverables 4
- Final thumbnails for episodes 1 and 2 and “Braquage National”.
- Episode 2 variants (“Bugatti”, “PS5”, “decoration”…) as 2K exports.
- Composition scripts and JSON settings files.
- The Oracle’s ranking of the 5 episode 2 hook variants.
Gallery 6
What this project means for you
The same know-how, on your problem.
A thumbnail system
A fixed identity and composition scripts: your thumbnails stay recognisable and spin off into variants without starting over for every video.
Choosing with an error margin
Variants ranked by a model trained on your own data, which also says when it does not know and then points to an A/B test.
Traceable campaign visuals
Every visual keeps its source and settings: a brand can find, fix or adapt it months later.