Issue #48 · Monday, October 19, 2026 · 7 stories

🎭 AI Avatar Video Drops To $4.20 A Minute, 🗳️ And Only 27% Of AI Political Ads Are Labelled

AI Comedians Copy Real Stand-Up Word For Word, And A Seven-Prompt Character Tests US Copyright

The useful tools this week give you more control, not more magic. Curious Refuge blocks out shots in Blender with Claude before Seedance 2.5 renders them, MOSS-Image places every word and object from a JSON layout, and Fish Audio's Drama 3 takes acting notes in square brackets. Control stops at the rights line, though: AI accounts are copying comedians' sets word for word, and a seven-prompt character is testing whether anyone owns it. The more of the result you direct, the stronger your claim to it.

NEWS

ByteDance's OmniHuman 2.0 cuts avatar video to $0.07 a second, in single takes up to five minutes · 6 min read

BytePlus launched OmniHuman 2.0 on 15 October. One image plus a voice track returns a lip-synced 720p performance, with text prompts steering gestures and camera moves, and more than one person allowed in frame. The price fell from $0.12 to $0.07 per second: $4.20 a minute, or $21 for a five-minute take. It is sold through BytePlus, ByteDance's enterprise platform, so budget for how many takes a usable minute needs.

Midterm campaigns have spent $137 million on AI-made ads, and only 27% carry a disclaimer · 12 min read

The Wesleyan Media Project now counts 220 US political ads made at least partly with AI, across 34 states, with more than $137 million behind them. In the 2024 cycle it found under $50,000. Only 27% carry a disclaimer, and the researchers call their count an undercount because AI is hard to detect. For anyone producing synthetic spots, disclosure is the step most of this work skips.

TECHNICAL

Curious Refuge blocks out AI shots in Blender with Claude, then hands the animatic to Seedance 2.5 · 8 min read

Claude drives Blender through an MCP add-on to build and animate a rough 3D scene, and the render becomes the video reference for Seedance 2.5. They test at 480p first because Seedance 2.5 is expensive. The catch: Seedance copied the animatic almost too closely, unnatural movements included, so they now prompt it to take only camera and blocking from the reference and treat the figures as placeholders.

ANALYSIS

AI 'comedians' are copying real stand-up sets word for word, and Meta says it breaks no rules · 8 min read

WIRED found eight comics whose viral clips were remade by AI-generated women, with better lighting and sound, to funnel followers to OnlyFans and Fanvue. One account, Aylen Skye, has 250,000 Instagram followers. Meta told Julia Jasunas that copying her joke was not impersonation, and Instagram allows AI content if it is labelled. Jokes have very little copyright protection, so the comics have little they can do.

Tung Tung Tung Sahur took seven prompts and 15 minutes, and a US court case now tests who owns it · 5 min read

An Indonesian teenager, Noxa, made the brainrot character in OpenAI's image generator from seven prompts. After Noxa's agency sent a cease and desist over the Roblox game Steal a Brainrot, the studio sued in a California federal court, arguing purely AI-generated work cannot be copyrighted in the US. The case is pending. If seven prompts are not enough human input, characters you generate for clients may be free to copy.

TOOLS

Fish Audio's Drama 3 voice model is $10 per million characters until 12 November, with open weights · 12 min read

Drama 3 takes acting notes in square brackets, like [shaky breath] or [whispering], speaks 83 languages and targets dialogue, audiobooks and short video. The API lists at $50 per million characters and is $10 until 12 November. The weights are on Hugging Face, free for commercial use under US$500,000 in annual revenue, and the licence bans cloning a real person's voice without consent. Fish says takes vary, so render several.

MOSS-Image is a 5B open image model that places text and objects from a JSON layout, under Apache 2.0 · 10 min read

OpenMOSS released MOSS-Image on 16 October with base and fine-tuned weights, 4K-token prompts and day-one ComfyUI support. The standout is layout control: you place each object and line of text with a bounding box in a JSON prompt, which suits posters and title cards. The weights are Apache 2.0, with no revenue cap on commercial use. The guide gives no VRAM figure, only a slower low-memory mode.