ArtStation flips NoAI on by default, and TikTok wires Claude straight into Ads Manager.
World Labs debuted Atlas, a world model that turns a single image into up to a minute of 1440p video and explorable 3D scenes. The key differentiator: camera trajectories are native inputs, giving pixel-perfect control over the virtual camera, unlike text-prompt video generators like Sora. Atlas also outputs point clouds and 3D Gaussian splats, but early access is limited to select enterprises.
TikTok launched Agentic Hub, an AI marketplace that connects tools like Claude and ChatGPT to Ads Manager via MCP, removing the need for API credentials or developers. The platform also introduced official AI skills like Viral Video Creator and upgraded Symphony Creative Studio for generating video ads from briefs. A furniture company produces 100 videos per week with this setup, but human oversight remains essential since AI recommendations should be challenged.
KitBash bought ArtStation from Epic in August and has now switched the NoAI tag on by default, applied it retroactively across existing uploads, and brought in Cloudflare to block known scrapers. It is not banning AI art, only filtering it out of search. Artists' reaction was mostly that it is too late, since three years of scraping already happened under the old owner. Many had already left for Cara.
Freelancer.com listings for AI cleanup jumped 87% to 10,760, Upwork remediation gigs rose 70% year over year, and Fiverr searches for 'AI cleanup' grew 20x since 2023. The numbers reflect a reality for AI-generated content: cheap first passes still require significant human fixing, with one illustrator offered $500 to repair 13-15 AI illustrations and a tarot deck needing 2-3 hours of Photoshop per image.
A hands-on comparison shows Alibaba's Wan 3.0 generates a 15-second 720p clip in 58 seconds using 7.5 credits, versus ByteDance's Seedance 2.5 taking 3m15s and 22.5 credits. Wan also outputs 1080p/30fps and accepts document/webpage inputs, at $0.10 per second versus Seedance's $0.462. But Seedance delivers more cinematic results with better camera work, lighting, and character consistency, making it the pick for reference-heavy projects.
Alexander Diel's team at Duisburg-Essen found viewers rate slightly imperfect AI food images as more uncanny than either realistic or stylised ones, and want to eat them less. That travels past food: the almost-real zone is the worst place to land. 42 percent of restaurant operators lost money in 2025 and 26 percent now use AI, mostly for marketing. One owner who went viral was review-bombed, threatened, and quit.
HR Endless Sampler renders MiniMax H3 video of any length by chunking one long latent. It runs Gemma 4 12B to plan shot timing, then hands each chunk the frames before it and writes a fresh prompt for what comes next. That is a tool conceding one over-stuffed prompt cannot carry a long shot. The demo is 625 frames at 1080p on a 16GB card, H3 only, 76 stars.
🎬 Cool Video of the Week
This week's pick is mine. I ran Seedance 2.5 on a black and white silent-era hotel scene, and it held up better than I expected: the character stays consistent across the shot, the physics behave, and it delivered the full thirty seconds in one generation instead of stitched fragments. The part worth passing on is what went wrong before it went right. Over-engineering the prompt hurt the output as much as under-writing it did, and both versions produced slop. Somewhere between too little direction and too much is where the model does its best work, and finding that line is the actual skill.
Clip by Apoorva. Tap the image to watch on Instagram.