Seedance's 30-second 4K clips, China-only for now. DeepMind bets $75M on A24 to build film tools.
Google DeepMind is investing $75 million in indie studio A24 to co-develop AI filmmaking tools, with Demis Hassabis framing it as building features alongside working artists. No tool, release date, or capability has been named yet, so for now it is a funding signal rather than something you can use. It follows Netflix buying AI-tools shop InterPositive and Amazon MGM standing up its own AI unit.
ByteDance revealed Seedance 2.5, which generates a 30-second 4K clip from one prompt and accepts up to 50 reference images, videos, or audio files, up from 12 in Seedance 2.0. The catch for most creators is access: it arrives in China next month with no stated rollout elsewhere, and Seedance 2.0's US rollout was already delayed under Hollywood pressure over training data. Plan around the wait.
Budget retailer Rainbow Shops told its models in 2025 that AI would mean fewer jobs, then began styling products and generating avatars with AI. Model Francheska Pujols is now suing, alleging the company produced lookalike images resembling her without consent around the time her work dried up. For anyone generating brand avatars, it is a clear line: cloning an identifiable person's likeness invites a lawsuit.
The EU AI Act's core transparency rules apply from 2 August 2026, and the one that touches creators most is labelling: AI-generated content, including deepfakes, has to be disclosed as synthetic. Fines reach 35 million euros or 7% of global turnover, though the 2026 Digital Omnibus pushed the toughest high-risk deadlines to 2027 and 2028. If you publish into the EU, build labelling into your workflow now.
ElevenLabs has produced an audiobook of Homer's Odyssey narrated by an officially licensed AI replica of Michael Caine's voice, made with the 93-year-old actor's consent through a deal he struck directly. It lands as the licensed, consented counterexample to the week's other voice and likeness stories. If you want a recognisable voice in your work, this is the route that does not end in court.
Spotify is pushing generative AI into its ad business with AI voice-overs, AI-written scripts, and automated media buying, and says 7,000 brands have already made more than 20,000 creatives with the tools. The pitch is set-and-forget ad production for marketers who could not previously afford audio ads. The flip side, which Spotify acknowledges, is fewer paid jobs for voice-over artists and sound editors.
A creator built a six-episode Spanish micro-drama for vertical platforms inside InVideo Agent One, and reports the hard part was not image or video quality but memory: keeping characters, locations, and visual rules consistent across episodes. The fix was a context system feeding a character bible and a visual-language bible, plus separate director, DP, and showrunner agents. Seedance 2.0 did the generation, with Kling as a fallback.
Krea published both halves of Krea 2 as open weights on Hugging Face: RAW, the undistilled 12B base built for fine-tuning and LoRA training, and Turbo, an 8-step distilled checkpoint for fast inference. The intended workflow is to train a LoRA on RAW, then run it on Turbo. ComfyUI v0.26.0 adds a native template, and community FP8 quantisations already run it on a 12GB card.