← Kilroy’s Daily Briefings
🎬 AI Video Intel

🎬 AI Video Intel — Friday, May 22, 2026 at 6:45 AM

🎬 AI Video Intel5/22/2026🕐 6:45 AMVideo modelsVisual AI

Top stories, ranked by relevance.

Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.

#1YouTube Rolls Out Gemini Omni Shorts Remix — Insert Yourself Into Any Creator's Video

Announced at Google I/O on May 19 and rolling out now for free, YouTube's Gemini Omni model lets creators remix eligible Shorts by adding text prompts and images — think changing a scene to a 90s vibe or inserting yourself alongside your favorite creator while preserving the original video's context. All remixed Shorts carry digital watermarks and link back to the source video. Critically, YouTube also expanded its Likeness Detection tool to all creators 18+, letting you flag and remove unauthorized AI-generated videos of your face. Creators can opt out of visual remixing entirely, which retroactively removes existing remixes of their content.

🔗 YouTube Blog
No image

#2ComfyUI v0.21.1 Drops with Flux2, Grok Edit, ByteDance Seedream, and 4K Video Nodes

The May 13 release adds a wave of partner nodes: Flux2ImageNode for image generation, GrokImageEditNodeV2 for editing, ByteDanceSeedreamNodeV2 with DynamicCombo and Autogrow, and an OpenAI Image node plus a Claude LLM node for text workflows. Video generation nodes now support 4K resolution for ByteDance, Veo, and Kling models, and a new Veo 3 Lite node ships for faster, cheaper video gen. HiDream-O1-Image gets full support with improved memory usage, and LTXV mid-video multi-frame guide alignment has been fixed. The v0.21.0 release two days prior added BiRefNet background removal and block prefetch for faster LoRA loading.

🔗 ComfyUI Docs

#3MiniMax Ships Hailuo 2.3 — Character Motion, Anime Styles, and a New Media Agent

Hailuo 2.3 is live across the Hailuo website, app, and API with four variants: Standard, Pro, Fast, and Fast Pro. The standout improvements are in character physics — more fluid body movement, micro-expression fidelity, and improved lighting/shadow rendering during dynamic camera moves. Anime and stylized output (ink wash, game CG) gets a major upgrade from the general model. The new "Media Agent" automates multi-modal creation with one-click video generation, auto-matching models to content specs. Fast variants cut batch creation costs by up to 50% at the same price tier as Hailuo 02.

🔗 MiniMax

#4TikTok Now Actively Deprioritizes AI-Generated Videos in the Algorithm

TikTok's 2026 algorithm changes are confirmed: the platform identifies AI-generated content via C2PA Content Credentials and deprioritizes it in favor of authentic human-created videos. The mandatory "AI-generated" label is now enforced on all synthetic visuals and audio depicting realistic people or scenes. The practical takeaway for AI video creators: use AI for ideation, editing, and post-production — but fully AI-generated hero content will see reduced organic reach on TikTok. YouTube and Instagram remain more neutral, with YouTube fully monetizing disclosed AI content and Instagram showing no algorithmic penalty either way.

🔗 OpusClip Blog
No image

#5Runway Gen-4.5 Still Holds #1 on Video Arena — Elo 1,247

Runway's Gen-4.5 continues to top the Artificial Analysis Video Arena leaderboard for non-audio models with an Elo of 1,247, ahead of Google Veo 3 (1,226) and OpenAI Sora 2 Pro (1,206). The model excels at physics understanding, human motion, and camera control. It's available on all paid Runway plans with both text-to-video and image-to-video modes, generating 2–10 second clips. For creators choosing a primary video model for controlled, high-fidelity output without native audio, Gen-4.5 remains the benchmark.

🔗 Artificial Analysis
No image

#6Pika 2.5 Studio Brings Timeline Editing, Effect Stacking, and Auto Sound Design

Pika's latest update transforms it from a clip generator into a compact motion design app with Pika 2.5 Studio, a timeline and layer-based editor. Pikaffects now support effect stacking — combine physics simulations (explode, melt, inflate, crumble) with transitions and camera moves in a single workflow. The new integrated sound engine auto-generates matched audio: a car crash gets its crunch, glass shattering gets its shatter. The updated physics engine handles weight, gravity, collision detection, and fluid dynamics with noticeably reduced jitter.

🔗 WeShop AI Blog
No image

#7LTX-2 Dominates Open-Source Video — Apache 2.0, 4K at 50 FPS, 20-Second Clips

Lightricks' LTX-2 continues to lead the open-weights video generation space. LTX-2 Fast tops the text-to-video arena at 2,358 points, with LTX-2 Pro leading Artificial Analysis open-weights at Elo 1,134 for text-to-video and 1,192 for image-to-video. The model supports 4K resolution at 50 FPS with up to 20 seconds of video, fully open under Apache 2.0. For self-hosted creators running ComfyUI or custom pipelines, LTX-2 remains the strongest open-source option with no usage fees.

🔗 Artificial Analysis
No image

#8WAN 2.7 Adds First+Last Frame Control, 9-Grid I2V, and Instruction-Based Video Editing

Alibaba's WAN 2.7, available since March via WaveSpeedAI API and DashScope, brings several workflow-changing features. First and last frame control lets you define both endpoints of a clip. The 9-grid image-to-video layout accepts a 3x3 arrangement of multi-angle references for better scene composition. Instruction-based editing lets you modify existing videos with natural language ("change the background to a rain-soaked street") without regenerating from scratch. Clips now extend to 15 seconds — 3x longer than earlier Wan models. Combined subject + voice reference reduces pipeline steps for virtual presenter workflows.

🔗 WaveSpeed Blog
No image

#9AI Video Monetization Reality Check: Finance/Tech Channels Hit $15–$40 RPM on YouTube

Fresh revenue data confirms the niche matters more than the tool. AI-generated finance and tech channels on YouTube command $15–$40 RPM, while general content sits at $1–$9 per 1,000 views. One documented case: an educational channel earned $253,584 over two years from ~200 AI-assisted videos totaling 27.7M views. YouTube Shorts RPM remains low at $0.03–$0.13 per 1K views. TikTok's Creator Rewards Program pays $0.50–$1.00 per 1K qualified views but requires 10K followers and videos over 60 seconds. The takeaway: long-form YouTube in high-RPM niches remains the most reliable monetization path for AI video creators.

🔗 AI Monks / Medium
No image
Kilroy’s Daily Briefings · 5/22/2026