Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Released May 13, ComfyUI v0.21.1 adds first-party partner nodes for HiDream-O1-Image, Google Veo 3 Lite, ByteDance Seedream V2, Grok Image Edit V2, and a Claude LLM text-generation node. Under the hood: LoRA async loading, block prefetch, and dynamic VRAM tuning deliver noticeably faster iteration. If you haven't updated in a week, do it today — this is the biggest partner-node batch drop in months.
Open-sourced May 8 under MIT license, HiDream-O1-Image is an 8B-parameter pixel-native unified transformer that handles text-to-image, editing, and subject-driven personalization up to 2048x2048 — no external VAE required. It debuted at #8 on the Artificial Analysis Text-to-Image Arena, making it the top-ranked open-weights image model. A distilled Dev variant ships at 28 steps for faster inference, and ComfyUI v0.21.1 already has native node support.
Kuaishou's Kling 3.0 4K mode, which went live April 24, renders native 3840x2160 at up to 60fps — pixel-level detail during diffusion, not upscaled post-generation. Multi-shot storyboards let you define up to 6 camera cuts in a single prompt with Subject Binding 3.0 holding character variation under 10% across the full sequence. The V3 4K API is available via invite, with broader platform integrations rolling out.
TikTok's automated detection now hits 94.7% accuracy on synthetic faces and 87.3% on AI backgrounds. The penalty structure has shifted to immediate strikes: first offense is content removal plus a strike, second is a 7-day posting ban, third is 30 days, fourth is permanent monetization ban, fifth is account termination. Proactively labeled AI content only sacrifices about 5-8% organic reach. Label everything or risk your channel.
Launched May 13, Higgsfield Supercomputer is a cloud-native AI agent with 40+ built-in tools and three layers of persistent memory. Tell it "build a full week of Instagram ads plus competitor analysis" and it selects models (Kling 3.0, Seedance, Claude Opus, GPT-5.5 Pro), generates all assets, and drops them where you need them. It self-learns brand guidelines over time. Accessible via browser or Telegram — no local setup.
OpenAI's Sora web and app shut down April 26 as scheduled. The API remains until September 24, but OpenAI has confirmed all user data will be permanently deleted after the deadlines. The shutdown was driven by unsustainable compute costs and limited monetization. Kling, Veo, and Seedance are absorbing displaced users, with the Disney $1B investment deal also unwinding following the discontinuation.
Google's Veo 3.1 Lite, available via the Gemini API since March 31, offers 720p video at $0.05/second and 1080p at $0.08/second — less than half the cost of Veo 3.1 Fast with the same generation speed. Supports text-to-video and image-to-video, landscape and portrait ratios, and 4/6/8-second durations. For volume creators building automated pipelines, this is the most cost-effective on-ramp to Google's video stack.
Following a Disney cease-and-desist in February and a Senate letter in March, ByteDance has deployed IP guardrails inside Seedance 2.0 that block generation of recognizable celebrity likenesses, copyrighted franchise characters, and branded visual identities. The system is model-level filtering, not post-generation moderation. If your workflow relies on Seedance for character-driven content, test your prompts — some previously viable outputs may now be blocked.
Lightricks' LTX-2.3 (released March 5, Apache 2.0 license) continues gaining traction as the go-to open-source video pipeline. It outputs 4K at 50fps in clips up to 20 seconds with synchronized audio, and uniquely supports native portrait video at 1080x1920 — trained on portrait data, not cropped from landscape. For creators shipping 9:16 content to Reels, Shorts, or TikTok, this is the cleanest free pipeline with full commercial rights.