NO FAKES puts a price on cloned faces ⚖️, Adobe automates Premiere 🎬, plus a one-person AI universe
xAI's Grok Imagine Video 1.5 is now general on the Imagine API, and the real upgrade is audio: sound effects, ambience and dialogue generate in the same pass and land on the action, not layered in later. Motion also holds together longer, with fewer warps across a clip. The Fast tier makes a 6-second 720p clip in about 25 seconds, down from 40-plus, though you are still in short 720p blocks.
Adobe has put its Firefly AI assistant inside Premiere, Illustrator, InDesign and Frame.io. In Premiere it takes on the unglamorous chores: sorting assets into bins, batch-renaming clips, identifying interview questions and dropping markers. A new Elements feature saves AI-generated characters, objects and locations to reuse across projects, Adobe's move on the consistency problem, alongside a Projects beta that holds a campaign's assets and context together.
A Guardian investigation found brands quietly running AI-generated influencers in social ads, made to look like real customers with no clear sign the people are not real. Some creators producing this work are asked to sign NDAs so they cannot discuss it. No UK rules require disclosing AI ad content, though the EU AI Act's labelling requirement starts in August, setting up a fight creators will be caught in.
A Kapwing study spun up a brand-new TikTok account and found 59% of its For You feed was AI slop, roughly three times the rate on YouTube. The concentration was highest in kids' content and lowest in categories that need a real person on camera. For creators, it is a read on where undifferentiated AI volume blends into the noise, and where showing your face still buys distance.
The US Senate Judiciary Committee has advanced the NO FAKES Act, which would give every American, not just celebrities, a federal right to authorise or block AI replicas of their voice and likeness. For creators cloning a real face or voice, the penalties set the stakes: $5,000 per work for an individual, $25,000 for a company that makes or distributes one. It still has to pass the full Senate and House.
Nvidia's Ziv Ilan walks through how his team got a video diffusion model running close to real time on a single Blackwell B200, with no new architecture. The trick is cutting the default fifty-step denoising process to as few as one, by stacking quantization, latent caching and step distillation, where a student model learns to match a teacher's output. Nvidia's open-source FastGen repo packages it, reaching the speedups real time needs.
The fast, cheap, or good, pick two rule in AI video cracked this week. ByteDance's Seedance 2.0 Mini is the cheap, fast tier that finally stopped being embarrassing: resolution is lower, 1470x630 against the flagship's 2206x946, but for inserts, drafts and doubling fight coverage at half the price it holds. Kuaishou's Kling 3 Turbo brings 3.0-level quality cheaper, with no lip-sync drift in a dialogue test.
Gerard Bisbal is building NUVAGAIA, a sci-fi universe spanning podcasts, stories, artwork, video, a wiki and interactive fiction, work that once needed a studio team and millions of dollars. He uses ChatGPT for worldbuilding, Gemini for scripts and voice, and DALL-E 3 for a consistent style, but discards anything off-tone. Without AI, he says, it would still be a folder of notes, and the coherence comes from his judgement.
This guide tests the free tiers of nine AI video generators on what matters: daily credits, resolution, clip length and watermarks. Seedance 2.0 is the only strong free tier that is both 1080p and watermark-free, with 100 refreshing daily credits, while Hailuo and Kling give sustainable credits rather than one-time trials. Most others cap output at 720 to 768p and reserve clean exports for paid plans.
LTX Director 2.0 is a free, open-source timeline editor that runs the LTX-2 video model inside ComfyUI. The v2.0 overhaul adds in-node editing: extend clips with prompts, keyframes and audio, drag-and-drop IC-LoRA, audio inpainting, and a Retake mode that re-generates only the segments you do not like. Because it runs on Lightricks' open LTX-2, generation is local with no per-clip credits, and the repo sits at 1.4k stars with example workflows.
One Node wraps the entire FLUX.2 klein image workflow into a single self-contained ComfyUI widget, replacing the usual tangle of wired nodes. It runs five modes from one panel: text-to-image, image-to-image, reference edit, a combined sketch, inpaint and outpaint mode, and faceswap. It works with both the 4B and 9B klein variants and supports GGUF models. At 210 stars and updated through 22 June, it is small but actively built.