Now I have enough to compile the briefing. Here it is:
---
🎬 AI Video Intel — 2026-04-21 13:09
Model & Tool Releases
• Sora dies April 26. App and web go dark in 5 days. API stays up until Sep 24. Lifetime revenue was ~$2.1M against ~$15M/day in inference costs. The Disney deal ($1B, 200+ character IPs) collapsed 3 months after announcement. Export your data now: sora.chatgpt.com/exports/me. (OpenAI Help Center, The Decoder)
• Alibaba's anonymous model hits #1 on Artificial Analysis Video Arena — Elo 1,347 (text-to-video), 1,406 (image-to-video), beating Seedance 2.0 by 74 Elo points, the largest gap in leaderboard history. Appeared April 7 with no official announcement. (Vidau AI)
• Netflix open-sourced VOID (April 3) — video object inpainting model that removes objects AND their physical interactions (shadows, reflections, displaced objects fall naturally). Built on CogVideoX-Fun-V1.5-5b-InP. Preferred 64.8% vs 18.4% over Runway in human evals. Requires 40GB+ VRAM. Available on Hugging Face and GitHub. (MarkTechPost)
• Seedance 2.0 API now live on fal — global rollout to 100+ countries. Unified multimodal architecture: feed up to 9 reference images + 3 video clips + 3 audio clips in one pass. 15-second multi-angle output with stereo audio. Integrated into CapCut. No real-face generation allowed (safety restriction). (fal.ai, TechCrunch)
• Runway GWM-1 — General World Model family built on Gen-4.5. Real-time interactive world simulation for avatars, robotics SDK, and virtual environments. Gen-4.5 itself now does 1-minute videos with character consistency + native dialogue + multi-shot editing. (Runway Changelog)
• Kling 3.0 continues to lead on price/performance — Elo 1,243, $6.99/mo entry. Generates 3-15 sec multi-shot clips with chain-of-thought scene coherence, motion transfer from reference video, and lip-synced audio in 5 languages. Up to 2-minute clips (vs Sora's ~25 sec cap). (InVideo, Kling Pricing)
• Hailuo 2.3 and LTX 2.3 Fast — Hailuo offers near-cinematic quality with a meaningful free tier. LTX 2.3 Fast is the speed champion: 4K-capable video in a fraction of typical generation time, best for rapid prototyping. (WaveSpeed)
Platform Signals
• YouTube AI Slop Crackdown: YouTube wiped 4.7 billion views in enforcement actions against AI-generated content farms. Kapwing study found 278 pure-AI channels with 63B total views, 221M subscribers, earning ~$117M/year. 21% of Shorts served to new accounts are AI slop. 200+ child advocacy groups now demanding a ban on AI slop from YouTube Kids. (Kapwing Report, OutlierKit)
• TikTok: C2PA Content Credentials now actively scanning uploads — auto-labels synthetic media even without self-disclosure. Deepfakes of real people without labels = banned. Synthetic media of real private individuals banned even WITH labels. Text-based AI (scripts, captions) remains exempt. (Storrito)
• Instagram/Meta: C2PA metadata triggers automatic "AI Info" / "Made with AI" labels. No manual opt-in needed if the model embeds provenance data. (Influencer Marketing Hub)
• Algorithm signal (all platforms): September 2025 algorithm updates continue to favor original content and penalize engagement bait. AI content isn't penalized if properly labeled, but generic flood = demotion/demonetization. Users can now filter AIGC visibility on some platforms. (Virvid)
What's Working
• The $117M AI slop economy is real but under pressure. Top earner "Bandar Apna Dost" (India, anthropomorphic monkey adventures) has 2.4B views, est. $4.25M revenue. But YouTube's crackdown wiped 4.7B views — the window for undifferentiated AI content farms is closing. (Kapwing)
• AI influencer licensing averaged $45K-$120K per brand campaign in 2025. The virtual influencer model remains viable for creators who build consistent IP. (CommuniPass)
• Higgsfield AI — $1.3B valuation, $200M ARR, 15M users as of Jan 2026. The tooling layer is where the money is concentrating, not the content itself. (@wallstengine on X)
• Revenue model shift: Paid communities and AI agents outperform ad revenue and sponsorships for stability. Tom Kuegler's AI bot hit $5K annualized in its first week. Views-based income is increasingly volatile for AI creators. (CommuniPass)
• Consumer sentiment headwind: 32% of US/UK consumers view AI as negatively disrupting the creator economy (eMarketer). Differentiation and disclosure are now table stakes. (Logie)
Technique Watch
• Netflix VOID pipeline: CogVideoX + interaction-aware quadmask conditioning (4-value mask: primary object, overlap, affected regions, background). Tutorial already up for building custom VOID pipelines. (MarkTechPost Tutorial)
• ComfyUI Realtime LoRA training: New comfyUI-Realtime-Lora node trains LoRAs on-the-fly inside ComfyUI for Wan 2.2, FLUX, SDXL, SD 1.5 — minutes on a decent GPU, immediately usable for img2img, style transfer, or subject consistency within the same workflow.
• Kling 3.0 motion transfer: Upload a reference video, Kling extracts its motion pattern, applies it to a completely different subject. This is the most practical "motion LoRA" equivalent in a commercial tool right now. (InVideo)
• Seedance 2.0 multi-reference prompting: 9 images + 3 videos + 3 audio clips in a single generation pass opens up storyboard-to-clip workflows that were previously multi-step.
• Wan2.2 quantization: Wan2.2-T2V-A14B-GGUF on Hugging Face makes the 14B model runnable on consumer hardware via quantization. Relevant for your ComfyUI setup, Kilroy. (Hugging Face)
Worth Watching
• Alibaba's mystery model — 74 Elo points clear of the field with no branding. If this drops as open-source (Alibaba pattern), it reshuffles the entire local-inference landscape overnight.
• YouTube Kids AI ban pressure — 200+ child advocacy groups pushing Google. If YouTube restricts AI content on Kids, expect spillover policy changes to Shorts and main feed recommendations. (Breitbart)
• Runway GWM-1 robotics SDK — world models trained on video are now being applied to robotics and interactive simulation. This is the path from "video generation" to "world simulation" that Runway has been telegraphing.
• Post-Sora consolidation: With Sora dead, expect Kling, Seedance, and Veo to absorb its user base this month. The $6.99 Kling entry price vs Sora's $200/mo ChatGPT Pro requirement tells the story.
• C2PA detection expanding: Both TikTok and Meta now auto-detect provenance metadata. Creators using tools that don't embed C2PA data may face manual review friction. Check whether your pipeline embeds it.