โ† Back to archive
Issue #51 ยท Monday, October 26, 2026 ยท 6 stories

๐Ÿ“บ Neighbours Cast Pushes Back On An AI Recut ๐ŸŽ๏ธ F1's Camera Moves That Never Happened

Pennsylvania's AI ad label starts 21 Dec, buyers want labels too, and two fixes for broken text.

The most interesting AI video this week is the kind nobody was told about. F1 appears to be generating the camera move between two real shots, and the giveaway is the one you will see in your own work: the text on the car turns to mush mid-move. Fremantle's Neighbours recut uses AI only to reframe, and cast members say they found out when journalists called. Pennsylvania's new ad law and Storyblocks' buyer surveys point the same way: hidden AI is the problem, labelled AI mostly is not. Canva Craft and PixText round it out with two attempts to fix the text that breaks first.

NEWS

Neighbours is being recut into an AI-assisted vertical microdrama, and the cast says nobody asked ยท 5 min read

RoseBerry Media and Fremantle are turning two of Margot Robbie's Neighbours storylines into about 150 minutes of vertical micro-episodes for the Epis app, using an AI-assisted tool called RedSnapper. Fremantle says RedSnapper only does catalogue analysis, editing and vertical reframing with a human editor signing off every stage, and does not generate scenes, alter faces or use synthetic voices. The cast is not reassured. Stefan Dennis called it "morally and legally wrong"; Matthew Werkmeister, who signed his first contract at 13, says he never consented to his image going through AI onto a platform that did not exist then. The MEAA is helping cast members and the Australian Writers' Guild is negotiating. The lesson for anyone recutting footage they did not shoot: no new pixels does not mean no consent problem, and old contracts never imagined a 9:16 reframe.

Pennsylvania makes AI ads carry a label from 21 December, and nobody knows what "false" means yet ยท 5 min read

Governor Josh Shapiro signed SB 806, the Artificial Intelligence Deceptive Advertising Act, on 22 October. From 21 December any ad that uses images, video or audio "significantly modified or generated by artificial intelligence" in a way that creates a false impression of the product must say it "contains content that was modified or generated using artificial intelligence", the first time the content appears, in the same medium, and where feasible in a way that is hard to strip out. There is no knowledge test: the "knowingly, recklessly or willfully" wording was cut on the Senate floor. Fines run to US$3,000 plus costs, enforced by the attorney general and district attorneys only. Parody, satire, obvious fiction and text-only ads are exempt. The catch, per Loeb & Loeb: the law never defines "false" against what, so a product shot generated rather than photographed may or may not need the label. If you make AI ads for US brands, plan the disclosure into the edit now rather than arguing the threshold later.

TECHNICAL

F1 broadcasts appear to use AI to invent the camera move between two real shots ยท 2 min read

Fans spotted pit-lane-to-cockpit transitions at the Singapore Grand Prix that no physical camera could fly, and F1 creator Matt Amys says the broadcast has used them since the Dutch Grand Prix in August. His read of the method: feed the model the last frame of one real camera and the first frame of the next, and let it generate the move in between. It is first-and-last-frame generation, the same trick most video tools now offer, used as a transition rather than as footage. The tells are the ones you will see in your own: decals and text on the car go illegible mid-move, and the speed dips then jumps. F1 did not answer PetaPixel, and a similar George Russell clip was taken down, so treat the method as informed guesswork, not confirmed.

ANALYSIS

Stock buyers say human-made is better, but 87% used AI content anyway ยท 2 min read

Storyblocks ran four surveys this summer of more than 1,400 people who license stock for work. About two-thirds said human-made content is better, yet 87% had used AI-generated visuals or audio in the past year, and half had rejected a stock asset specifically because it was AI. Between 74% and 94% want AI content labelled, 63% work somewhere with an AI-content policy, and the availability of AI content ranked last among what buyers want from a platform. Read it with the source in mind: Storyblocks does not accept AI uploads, so it has a stake in this answer. The usable signal still holds for anyone selling AI footage to brands: buyers use it, distrust it, and expect a label, so disclose before the client asks.

TOOLS

Canva Craft: an image model trained on templates, not photographs, with text in 55 languages ยท 1 min read

Canva released Craft, its first in-house image model, to every Canva user on 23 October through Dream Lab and the public API. It renders at 2K in more than 30 aspect ratios and writes on-image text in 55 languages, because it was trained on millions of human-made Canva templates rather than photos, so layout and type are the point. Canva says it debuted fourth overall on DesignArena and first for portraits. What Canva does not say: what it costs in credits, or which plans get it. Cheaper Craft-fast and Craft-ultrafast variants and an MCP connector are listed as coming, with no dates. Worth a test for thumbnails and covers where text usually breaks, but keep your type in the edit for anything that ships.

PixText fixes garbled text inside an image you already generated, Apache 2.0 ยท 5 min read

PixText, from ATH-MaaS, repaints only the text regions you mark in an existing image, using the surrounding pixels as context, so a good generation with a misspelt sign no longer has to be rerolled. Weights are Apache 2.0, with the inference code on GitHub and a public Hugging Face Space to try it. The fast MeanFlow variant runs four steps, the Teacher variant 25. Limits: it needs a CUDA GPU with FlexAttention, Python 3.11 or newer, there is no ComfyUI node on the model page, and no VRAM or timing figures are published. Useful as a repair pass; it does not replace setting final type in the edit.