Issue #9 · Monday, July 13, 2026 · 7 stories

🚫 Meta's Instagram AI photo grab lasted 3 days 🎬 Consistent characters, local voiceovers & $18 Veo clips

Google now makes advertisers disclose AI, the labels want streaming tags, and Photoroom opens up its data pipeline.

Meta grabbed public Instagram photos for AI remixes on an opt-out-by-default basis, then pulled the feature three days later after SAG-AFTRA and privacy advocates pushed back. It landed the same week Google started requiring AI disclosure on ads and the record labels proposed streaming tags for AI music, so the whole week reads as one story: consent and disclosure are no longer optional. On the make-something side we get into Photoroom's data pipeline, a Claude prompt-stacking system for consistent characters in Seedance, local voiceovers that skip the API bill, and whether Canva's new Veo 3 clips earn the $18.

NEWS

AI Feature That Grabbed Public Instagram Photos Pulled After 3 Days · 4 min read

Meta launched a feature in its Muse Image tool that let anyone generate AI images from public Instagram photos by @-mentioning the account, with no notification or consent required. The feature was opt-out by default for public accounts over 18. SAG-AFTRA, CAA, and privacy advocates called it non-consensual exploitation of likeness. Meta acknowledged it 'missed the mark' and removed the tagging feature, though the broader Muse Image tool remains. The incident shows opt-out consent remains the default across Meta's AI systems.

Google to Label AI-Generated Ads on Search, YouTube · 1 min read

Google will label AI-generated ads across Search, YouTube, and Discover through a new 'How this ad was made' section in My Ad Center. The disclosure is automatic for ads created with Google's AI tools. For third party AI tools, brands can voluntarily disclose but aren't required unless local law says otherwise. This adds a transparency layer for viewers and a compliance factor for advertisers using AI to produce video ads.

Record Labels Propose Two AI Tags for Streaming Platforms · 4 min read

The RIAA and IFPI, backed by the Recording Academy and SAG-AFTRA, are pushing Spotify and Apple Music to adopt standardized labels: 'AI-generated' for fully synthetic tracks and 'AI-assisted' for human work with AI tools. A Deezer/Ipsos study found 97% of listeners couldn't distinguish AI from human music, yet 80% want clear labeling. Deezer and TIDAL already detect AI music at the platform level, while Apple Music relies on distributor declarations. The streaming industry's trade group has not committed to the tags.

TECHNICAL

VLM Re-Captioning at 3-4% Throughput Cost: Photoroom's Data Pipeline · 8 min read

Photoroom published the data pipeline behind their 7B image generation model. They assemble pre-training data from public and internal sources, re-caption images with Qwen3-VL at only 3-4% throughput cost, and store images as JPEG quality 92 after confirming no measurable quality loss versus PNG. The post also covers hard-won lessons on Lance fragmentation and the importance of compaction for query performance.

Prompt Stacking in Claude for Consistent AI Video Characters · 2 min read

A tutorial for Seedance 2.0 inside CapCut Video Studio demonstrates a prompt stacking system built in Claude (replicable in a spreadsheet) to chain locations, style lockups, and shot lists for consistent characters and cinematic aesthetics. Two practical hacks: prototype at 480p before 4K to save credits, and use Omni for spatial reference to eliminate flickering. The sponsored workflow is demonstrated on the short film 'Folklore' and offers actionable techniques for solo creators.

TOOLS

Local Voiceovers Without the API Bill: audio.cpp Runs VibeVoice at 4x Real-Time · 1 min read

audio.cpp is a pure C++/ggml runtime for audio models with no Python dependency, and it now runs Microsoft's VibeVoice 1.5B for text-to-speech on your own machine. The receipt from the release: a 90-minute podcast renders in about 23 minutes, roughly 4x real-time and 2.8x faster than the Python reference, before any quantization. For anyone paying per minute for cloud voiceover, that moves the whole cost line onto hardware you already own. The catch to weigh: it is a young, largely single-maintainer project and the model coverage is still filling in, so treat it as a fast local option to test, not a drop-in replacement for a production voice pipeline yet.

8 Second AI Video Clips Arrive in Canva with Veo 3 · 9 min read

Canva now offers two AI video tools: the new Veo 3 powered 'Create a Video Clip' for Pro subscribers at $18 per month and the older Magic Media for free. The Veo 3 tool generates 8 second clips and allows about 5 generations per month, with credits shared across AI features. The free Magic Media is lower quality. This is best for B-roll and social snippets, not long form production.