Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Wan 2.7 officially landed in the ComfyUI Template Library via Partner Nodes — search "Wan2.7" and you're generating. Supports up to five real-person image inputs, vocal timbre reference, audio-driven image-to-video, and first-plus-last-frame control. Catch: local weights still haven't dropped as of this morning, so hosted via wan27.org is your fastest path for now.
The June 1 release is dense: native NVIDIA Cosmos Predict2 support in 2B and 14B text-to-image variants plus image-to-video, Veo 3 Lite integration, TripoSplat for image-to-3D Gaussian splat output, and 4K resolution unlocked for Veo and ByteDance models. A dependency-free MediaPipe Face Detection reimplementation also ships as a built-in node. Plenty to wire up this weekend.
Lightricks' LTX-2.3 is a 22-billion-parameter Diffusion Transformer running 4K at 50fps for up to 20 seconds with synchronized stereo audio in a single pass. Native portrait mode at 1080x1920, no cropping, four checkpoint variants including a distilled 8-step for speed runs. Open weights, permissive license, ComfyUI nodes included. If you have the VRAM, this is the local workhorse right now.
Announced at Computex, Cosmos 3 is a two-tower Mixture-of-Transformers omnimodel that generates text, images, video, audio, and action tokens in a unified architecture. Cosmos3-Nano sits at 16B total for workstations; Cosmos3-Super hits 64B for datacenters. Licensed under OpenMDW-1.1 for commercial use, weights on Hugging Face. Runway is a founding co-developer of the Cosmos Coalition. This is a different category of open model than what existed six months ago.
Since May, YouTube auto-applies AI disclosure labels to photorealistic content where creators didn't self-disclose. The January enforcement wave already terminated 16 channels accounting for 4.7 billion lifetime views and an estimated 10 million dollars in annual ad revenue. Low-effort automated channels are losing monetization; human-directed AI content in high-CPM niches still pulls 10 to 15 dollars RPM. Label your stuff and show your creative direction — that is now table stakes.
Effective June 1, Meta's new "About this ad" destination centralizes AI disclosure across all its platforms. An "AI info" label auto-applies to content flagged via industry-standard AI markers or self-disclosure. Political and electoral synthetic creative requires mandatory disclosure; deceptive AI media gets suppressed rather than deleted under the revised policy. If you're running paid AI creative on Meta properties, audit your disclosure workflow today.
MiniMax quietly shipped Hailuo 2.3 with significantly improved complex body physics, new anime, illustration, and ink-wash style support, and tighter facial micro-expressions. Same pricing as Hailuo 02 — they're calling it a global record for cost-effectiveness. Live on the website, app, and fal.ai for pipeline integration. If movement quality was your reason to pass on Hailuo before, this is worth revisiting.
The Sora web app went dark April 26. The API sunsets September 24, 2026. OpenAI is refunding long-term Pro subscribers. They did broker a licensing deal with Disney right before the shutdown, which is a fine parting irony, but it does nothing for your workflow. Veo 3.1, Runway Gen-4.5, and Kling 3.0 are the natural landing spots depending on what you need. The migration window is open and closing.
