← Back to archive
Issue #50 · Friday, October 23, 2026 · 7 stories

🎸 A Band Found AI In Its Own Video Via The Credits 🥓 Just Eat's AI Puts Bacon In Soda

YouTube asked 694 uploaders about AI: 72% already use it, and most want themselves in the shot.

The cheapest generation is the one you planned. CozyClay lets you block a shot in the browser, with no GPU, before you pay for a single take, and a creator named PJ's fix for stiff 3D-guided shots keeps your composition through image-to-video. A side-by-side on the same effects shot shows Seedance 2.5 holding together where Kling 4.0 slops, so pick the model by the test, not the launch post. Even Just Eat's bacon-in-soda menu photos were a planning failure: nobody checked the output against the dish.

NEWS

YouTube's own survey: 72% of young uploaders already use AI, and most put themselves in the shot · 1 min read

YouTube's Culture & Trends team and NRG surveyed 694 US video posters aged 14 to 44 in April. 72% say they already use AI on their videos or images, more than two-thirds of those posting AI video have featured themselves, and 61% want to insert themselves into content. The demand is for the creator as the character, not for faceless clips, which is the case for likeness tools and consistent self-casting. It is a commissioned survey of stated interest, not usage data, and the sample is small.

TECHNICAL

Iron Goddess, a mystery model on the Arena, smudges its fastest frames, and the fix for stiff 3D-guided shots · 4 min read

Theoretically Media tested Iron Goddess, an unclaimed video model in Arena battle mode: big, fast motion, but frame by frame the fastest moments go smudgy, while Seedance 2.5 holds together with less movement. Nobody has said who made it; the Alibaba guess is a guess. The useful part is a fix credited to a creator named PJ for 3D-guided generations that come out stiff: screenshot the previs, lay it out as a storyboard and go image-to-video, which keeps the composition, lens and aspect ratio. The video is sponsored by Recraft, disclosed in the description.

ANALYSIS

Fontaines D.C. found out from the credits that their video used AI, and now ban it in contracts · 3 min read

The 'Marianne' video credited "AI artists", and frontman Grian Chatten told the LA Times that neither the band, their label nor management were told; someone on the VFX team used AI in the workflow, out of more than 100 people on the shoot. The band has since written into its contracts that AI cannot be used in its music videos. For anyone selling AI work into music or brand jobs, the lesson is disclosure up front: a client who learns about it from the credits can blow up the relationship and close the door for the next vendor.

Just Eat pauses its AI menu photos after one showed bacon and eggs floating in soda · 3 min read

Just Eat trialled AI-generated images for UK restaurants with no photos of their own, tagged as AI-generated, and paused it after complaints. A Northern Irish soda bread sandwich came out as sausages, bacon and eggs in pints of fizzy drink, salt and pepper chips became plain fries under a pile of salt, and a vegetarian cheddar listing showed an orange fondue sauce. The tool misread the menu descriptions. Labelling did not save it: when a picture is a promise of what arrives, a wrong one costs refunds and trust, which is why Zomato pulled AI dish images in 2024.

TOOLS

CozyClay is free previs that blocks your shot before you pay for a take · 1 min read

CozyClay is an open-source previs studio that runs in the browser with npx cozyclay: no GPU, no Python, no model download, and a one-click Pinokio install. You block the set, pose characters and plan camera moves, then export a keyframe pack (first and last frame, greybox clip, camera and prompt) plus depth passes for your video model. The point is to settle the composition before you burn paid generations. It is AGPL-3.0 with 961 GitHub stars, there are no tagged releases yet, and the agent panel bills to your own API key.

Kitten TTS 2 clones a voice on your CPU, and its free licence stops at US$1 million · 1 min read

Kitten TTS 2 is a 1.7B speech model whose default weights are packed to 1.58 bits, about 954 MB, and run on a CPU with no GPU. It clones a voice from 5 to 30 seconds of one speaker and ships 115 voices, with named voices in 16 languages, though the text normaliser is tuned for English. The catch is the licence: the weights are free for commercial use only until you or your affiliates pass US$1 million in annual revenue or total funding. The model card gives no speed benchmarks, so test it on your machine first.

QUICK HITS

Cool Video of the Week: Seedance 2.5 vs Kling 4.0 on the same effects shot · 1 min read

A man in a suit, seen from behind, fires a glowing blue pistol at a restaurant brick wall that bursts into blue particles around a lit fireplace and a set tableM.E.G. Pictures ran the same 30-second shot through both models: a man in a suit fires a glowing blue pistol at a restaurant's brick wall, and the wall erupts in blue particles beside a lit fireplace. Seedance 2.5 is on the left, and on his read, and ours, it holds the room, the table and the man together where Kling 4.0 slops. Kling 4.0's pitch was control over luck, so a side-by-side like this is the test that matters. Posted by @MukiTanaka.