Everyone who edits brand images with AI knows the fifth edit: the colour has drifted, the edges have gone soft, and you start again from the original. Ideogram 4.5, released yesterday, is built against exactly that. If its claim holds, iterating on an approved key visual stops being a gamble.
Ideogram 4.5 keeps an image clean across many edits
On 30 September Ideogram released 4.5, which it calls “the most precise edit model”: each edit changes what you ask for, without the pixel shifts, colour changes and artifacts that build up over several rounds. It is live in the Ideogram app, the API and with launch partners, with open weights “soon” (Ideogram, 30 Sep). The model page shows recolouring and relighting a product while the composition stays the same, editing a crop of a 24.2 MP image without downsizing it, and translating the text in a design while keeping the layout — its examples are Japanese, French and English (Ideogram model page). The comparison against GPT Image and Nano Banana is Ideogram’s own test.
My read: On TeaTime Rose Jelly the palette stays in pinks and warm gold from the first frame to the packshot, so the pack feels like the last scene of the story, not a cut to an ad. That only works if the pink is the pack’s pink in every frame. Drift across edits is what kills it, which is why I go back to the original instead of editing an edit. Translation in place is the bigger promise for anyone running one pack in two languages. The limit: the page shows no Persian or Arabic. Right-to-left lettering is where I would test first, before trusting it on a pack.
HeyGen’s first video model starts at one cent a second
HeyGen launched HeyGen Video on 30 September, its first general-purpose video model: text, image or reference to video, with dialogue, ambience and effects generated in the same call. It is built on MiniMax H3 and post-trained by HeyGen; launch pricing is $0.01 a second through October, half the standard $0.02. HeyGen’s own table lists Kling 3.0 Pro at $0.168 and Veo 3.1 at $0.40 a second for image-to-video at 768p (HeyGen, 30 Sep; OpenRouter listing). Clips run 5 to 15 seconds at 480p or 768p, and the docs say it is weakest at long on-screen text, petals, paper, hair and close hand work (HeyGen docs).
My read: Read that weakness list again: petals, hair, hands. That is half of food advertising. TeaTime is a rose garden with kids picking biscuits off branches. A cheap model that fails exactly there is not a hero-film tool. It is an animatic tool, and that is still useful. On Mooo Milk — From 1 to 1000 the camera stays locked on one court at one hour of golden light, and the film is 29 seconds of keeping that consistent. At a cent a second I can test the blocking of every beat before I spend Kling credits on the frames the client will see. 768p is a ceiling for previz, not for delivery.
ElevenLabs doubles to a $22bn valuation
ElevenLabs said on 30 September that a $300m employee tender led by Wellington and T. Rowe Price values it at $22bn, double its February Series D. Enterprise is 55% of revenue, its agents handle more than 15 million conversations a week — three times February — and it covers 90+ languages. It names a Hasbro partnership for its Voice Library, and says its customers’ voice agents resolve issues 31% faster than chat agents (ElevenLabs, 30 Sep).
My read: The money is in voices that answer customers, not voices that read scripts. That moves a brand character from the ad into the service channel. The tiger in a leather jacket in POFAN BBQ is a character I designed for a 54-second film. If a brand ever lets a character like that talk to customers, the voice is no longer a production line item. It is a contract: who owns it, where it may speak, what it must never say. The 31% figure is ElevenLabs’ own analysis. Write the voice brief before the agent brief.
What I’d do today
Take one approved key visual with the pack in it. Run five edits in a row in Ideogram 4.5 — background colour, light, one copy line, then a translation of the claim. Lay the fifth result over the original at 100% and look only at the pack. If the logo and the colour hold, it goes into your resize workflow. If not, it stays in moodboards. If you want the rest of the pipeline set up, here is how I work.
Sources
- Ideogram — Introducing Ideogram 4.5, 30 Sep 2026 — x.com/ideogram_ai
- Ideogram — Ideogram 4.5 model page, accessed 1 Oct 2026 — ideogram.ai
- HeyGen — HeyGen Video launch, 30 Sep 2026 — developers.heygen.com
- HeyGen — Video 1 model docs, accessed 1 Oct 2026 — developers.heygen.com
- OpenRouter — HeyGen: Video 1 listing, 30 Sep 2026 — openrouter.ai
- ElevenLabs — $22bn employee tender, 30 Sep 2026 — elevenlabs.io