Every approved key visual turns into a week of resizes: story, feed, square, banner, then the same again in a second language. Luma Variants, live since 1 October, takes one approved static ad and does that step with the logo, headline and CTA locked. That is the slowest, least creative week in a campaign, and the one where errors slip through.
Luma Variants: one approved ad, five placements, several languages
Luma launched Variants on 1 October. You upload one approved static ad, choose placements and languages, and it resizes and translates. Luma says the logo, headline and CTA “stay intact”. The five placements are 9:16, 4:5, 1:1, 6:5 and 16:9. It is live in the Discover tab under the Ad Variants card, aimed at “agency and brand teams”. This first release covers static ads only, and the post gives no price. Its language examples are English and Spanish (Luma, 1 Oct).
My read: A resize tool is only as good as what it refuses to touch. Luma locks the logo, the headline and the CTA. For an FMCG ad the fourth lock is the pack, and nobody has promised that one yet. One Group. Every Shelf. was built vertical-first at 1080×1920, one category per beat. Going from 9:16 to 16:9 is a recomposition, not a crop: the products have to stay whole and the type readable. That is the real test. The second one is script. The examples are English and Spanish. Arabic and Persian headlines run right to left and change the whole layout. Test one before you trust twenty.
Stability AI rebuilds around licensed music
TechCrunch reported on 2 October that Sean Parker is rebuilding Stability AI around music. Sony, Warner and Universal took part in its $76 million round in late August and licensed their catalogues for training. Stability has released three new audio models and music-editing software that makes instrumental tracks or short snippets from text. Humming a melody or beatboxing a pattern as input is coming next. Parker says he is “playing by the rules this time” (TechCrunch, 2 Oct).
My read: The score on Sky Red Grape is my composition, with the track made in Lyria. There is no dialogue and no on-screen text, so the music carries the story. The first question a brand’s legal team asks about that track is where the training data came from. A model trained on licensed label catalogues gives a cleaner answer. The limit: the report gives no commercial terms and no prices. Licensed training does not tell you what you may do with the output. Read the licence before the bed goes on air.
Suno Speech: voice and music generated as one track
Suno opened its Speech beta to everyone on 1 October. You write the text and describe a voice and a music style, and it returns one track: a spoken voice over original background music. Suno warns that “British accents can wander off to Australia and back” and that dramatic pauses may be “very dramatic”. The post lists no languages, no prices and no length limits (Suno, 1 Oct).
My read: Good for a scratch track. Not for a final mix. In a commercial the voice-over and the music bed are separate decisions. The voice gets re-recorded after legal notes, and in the 15-second cutdown it moves against the picture. One baked-in track removes both options, and the post says nothing about separate stems. Use it to time an animatic before the real voice is booked. Then throw it away.
What I’d do today
Take one static ad that is already approved and live. Run it through Luma Variants into 9:16 and 16:9 in the two languages you actually ship. Check three things at 100%: is the pack whole and unstretched, does the headline still sit where the eye lands, and is the translated line one a native speaker would sign off. If all three hold, it saves a week of resizes. If the pack breaks, keep resizing by hand and wait for the next release. If you want the rest of the pipeline set up, here is how I work.
Sources
- Luma AI — Introducing Luma Variants, 1 Oct 2026 — lumalabs.ai
- TechCrunch — Sean Parker is rebuilding Stability AI around music, 2 Oct 2026 — techcrunch.com
- Suno — Introducing Speech (beta), 1 Oct 2026 — suno.com