The pack is the one thing in a key visual nobody is allowed to reinterpret. FLUX 3 Image, released on 1 October, lets you draw a box, edit what is inside it and leave everything else alone. If that holds on real packaging, it changes how variants and resizes get made.
FLUX 3 Image: edit by coordinates, keep the rest
Black Forest Labs released FLUX 3 Image on 1 October. One endpoint generates and edits; you place or change elements with bounding boxes, recolour, move or resize them, and combine up to ten reference images. Output runs from a 768px preview to 4K, about 16 megapixels. API prices per image: $0.041 at 768px, $0.048 at 1K, $0.10 at 2K and $0.607 at 4K (BFL release notes, 1 Oct). The model page describes the canvas as a 0–1000 grid on both axes and promises “multiple edits at once without changing anything else”; companies can license commercial weights to run it in-house (BFL model page). The API is 50% off through 8 October, and open weights are expected “in the coming weeks” (The Decoder, 2 Oct).
My read: In KRAX — Happy World Potato Day the chips land in the pack at the end of one continuous run, and golden-hour light has to hold from the field to the packshot. The packshot is where AI work usually breaks: the logo softens, the colours shift a shade, and you composite the real pack back in by hand. A locked box is the right idea. A colour script is the other use. On Hit Wafers the two flavours on the pack, orange and banana, set the colours of the film. Swapping one flavour for the other inside a box, and nothing else, is exactly the job. The limit is price. 4K costs six times 2K, so I would proof at 2K and only pay for 4K on the frame that ships.
Tavus Griffin: 26 of 54 people thought it was human
Tavus announced Griffin on 1 October, a model that sees, listens and answers as a real-time video face. In its test, people were told they would video-call another participant about their plans for the year. After one minute, 26 of 54 believed Griffin was a real person; the earlier Phoenix-4.5 fooled 1 of 41. Griffin-Lite is a research preview for select testers. It is not on the Tavus platform yet, and Tavus says it is “working on safe disclosure features” (Tavus, 1 Oct).
My read: This is a disclosure question before it is a production one. In the Tiny Challenge I borrowed the grammar of user-generated video on purpose — a phone on a tripod, a classroom, daylight from the windows. That is a style inside a film. A face that answers you live, and that half the people in a test take for human, is a different contract with the audience. Note the size of the test: 54 people, one minute, Tavus’s own study. A brand that puts a face like this in customer service should say so in the first second. My own characters stay characters, which is one reason I build mascots.
Higgsfield Ads Studio: a URL in, 20 static ads per product out
On 30 September Higgsfield released Ads Studio. You give it your website link; it takes your logo, colours, tone of voice, audience and selling points once, lets you pick products from the site, and generates static ads in batches of up to 20 per product, with a chat to ask for more. The entry gives no price or plan (Higgsfield changelog, 30 Sep).
My read: A website can give a tool your logo, your colours and your selling points. It cannot give it the idea. In cici — Every Big Dream Starts Small the cake is not the hero; it is the constant in a memory that ends in orbit. No URL contains that. So I see this as the volume end of the job: once the idea and the key visual are approved, 20 static variants per product for resizes and A/B tests. Check every pack in every one of the 20. Batch work is where a wrong flavour or an old pack slips through.
What I’d do today
Take one approved packshot. In FLUX 3 Image at 2K, box the pack, then change the background, the light and one prop around it. Do it while the API is half price, until 8 October. Lay the result over the original and check the pack at 100%. If the logo and the colour hold, it goes into your resize workflow. If not, keep compositing the real pack by hand. If you want the rest of the pipeline set up, here is how I work.
Sources
- Black Forest Labs — release notes: FLUX 3 Image, 1 Oct 2026 — docs.bfl.ai
- Black Forest Labs — FLUX 3 Image model page, accessed 3 Oct 2026 — bfl.ai
- The Decoder — Black Forest Labs launches Flux 3 Image, 2 Oct 2026 — the-decoder.com
- Tavus — Griffin, 1 Oct 2026 — tavus.io
- Higgsfield — changelog: Ads Studio, 30 Sep 2026 — higgsfield.ai