Here's your briefing, Kilroy:
---
🎬 AI Video Intel — 2026-04-28 06:45
Model & Tool Releases
BREAKING: Sora is dead. OpenAI shut down the Sora app on April 26 — two days ago. API stays live until Sept 24. Was losing $1M/day per WSJ. Active users had cratered from 1M to under 500K. Sora creator Bill Peebles and CPO Kevin Weil both left OpenAI the same day. Disney's rumored $1B partnership never had a signed contract; they learned of the shutdown less than an hour before the public announcement. (OpenAI, Tech-Insider, TechCrunch)
HappyHorse-1.0 (Alibaba) launches as new #1. API access went live April 26 via fal, same day Sora died. Built by Alibaba's Taotian Future Life Lab, it debuted on Artificial Analysis Video Arena on April 7 and immediately took the top spot. Current Elo: 1367 T2V (no audio), 1230 T2V (with audio), 1392 I2V. Enterprise testing opens on Alibaba Cloud Bailian, commercial launch scheduled for May. Produces 1080p with synced audio, strong lighting, realistic detail. (PR Newswire, CNBC)
ComfyUI raises $30M at $500M valuation (announced April 24). Led by Craft Ventures. Total funding now $48M. 4M+ users, 60K+ community nodes, 150K daily downloads. Core stays MIT-licensed; monetization via Comfy Cloud and enterprise. Notable client: Silverside AI used ComfyUI to produce SVEDKA's 2026 Super Bowl ad — first primarily AI-generated Super Bowl commercial. NVFP4 support on RTX 50 series gives 2.5x speed + 60% VRAM reduction. (TechCrunch, FinSMEs)
Seedance 2.0 (ByteDance): Officially launched April 9. First unified audio-video joint generation architecture — audio synthesized with frame-level awareness, not post-synced. Up to 15s multi-shot output with dual-channel stereo (music + SFX + dialogue). Elo 1270 T2V (no audio), #2 behind HappyHorse. Rolling out via CapCut in Brazil, Indonesia, Malaysia, Mexico, Philippines, Thailand, Vietnam. (ByteDance, TechCrunch)
Veo 3.1 (Google): Now free for all personal Google accounts via Google Vids and Flow (announced April 2). Three tiers: full (max fidelity), Fast, and Lite (high-volume/cost-effective). New upscaler handles any source up to 4K. Native audio across all tiers. (gHacks, Google Cloud)
Netflix VOID (open-sourced April 3): Video object removal + physics-aware inpainting. Built on CogVideoX-Fun with quad-mask conditioning (primary object, overlap, affected regions, background). Apache 2.0 license. Preferred over Runway 64.8% vs 18.4% in human evals. Requires 40GB+ VRAM. (HuggingFace, MarkTechPost)
LTX-2.3 (Lightricks): Redesigned VAE for sharper details, gated attention text connector for better prompt adherence, native portrait (1080×1920) support. Weights on HuggingFace — base, fp8 quantized, and 8-step distilled variants. (LTX)
Current Artificial Analysis leaderboard (T2V, no audio): HappyHorse-1.0 (1367) → Seedance 2.0 720p (1270) → Kling 3.0 1080p (1247) → SkyReels V4 (1239) → Kling 3.0 Omni (1232). (Artificial Analysis)
Platform Signals
TikTok: Actively deprioritizing AI-generated content in the FYP algorithm, favoring "authentic human creators." AI content is explicitly excluded from the Creator Rewards Program — the program evaluates whether content is "unique to the creator and showcasing their point of view." C2PA Content Credentials used for automatic detection even without self-disclosure. Labeling required for realistic AI images/audio/video. (OpusClip, Storrito)
Instagram/Facebook: Meta unified AI content policies across both platforms (Feb 2026). "Made with AI" label required via advanced settings before posting — appears as a tag above content, similar to paid partnership labels. Notably, AI content with unique visuals + original voiceover + custom captions still scores high on originality signals. (Influencer Marketing Hub)
YouTube: AI content remains fully eligible for monetization through YPP with disclosure. Standard thresholds apply (1K subs, 4K watch hours or 10M Shorts views). Disclosure required for "realistic altered or synthetic content." Most permissive major platform for AI creators. (Vexub)
X/Twitter: Aggregator payouts slashed to 60%, with another 20% cut coming next cycle. Prioritizing original content creators. Premium+ accounts average 1,550 impressions/post vs under 100 for free accounts (15x disparity). Creator monetization payouts expanded to $100M in 2026. (ALM Corp)
Cross-platform: Content identified as recycled from another platform sees up to 40% reach reduction. (ShortSync)
What's Working
Faceless YouTube at scale: Creator Adavia Davis runs 5 channels doing documentary-style videos (history, Minecraft, animals, celebrity, anime) — revenue $40-60K/mo. Creates 6-hour background videos designed for sleeping audiences. Production cost ~$60/video using Claude for scripts + ElevenLabs for narration. 85-89% profit margin. (AI Video Bootcamp)
AI UGC agencies: Two-person shops doing ad creatives, video localization, and social packages hitting $20K/mo on $200-300/mo in tool costs (Runway Gen-4.5 + Kling + ElevenLabs + HeyGen). Single ad creatives: $500-1K. Monthly social packages: $2-5K. (AI Video Bootcamp)
AI avatar UGC: Creators filming one base video then using Argil AI / HeyGen to generate infinite script variations at $250-1K per deliverable, ~30 min production time. Revenue $5-10K/mo on $58-128/mo tool spend. (AI Video Bootcamp)
Content machine operators: Batch-processing 100+ variations via API at 2-5 cents per piece, scaling winners to 1,000+. Revenue $2-8K/mo across TikTok Creator Fund ($1-1.5K/mo) + Shorts revenue share + affiliate + brand deals. Traction typically starts week 6-8. (AI Video Bootcamp)
YouTube remains the AI creator's best friend — full monetization, no algorithmic penalty for properly disclosed AI content, and the long-form format rewards the "sleep video" / background content niche where AI production shines.
Technique Watch
Netflix VOID quad-mask pipeline: The 4-value mask system (primary object → overlap → interaction-affected regions → background) enables physics-aware removals that handle falling objects, displaced items, and secondary effects. Integrable into ComfyUI via CogVideoX nodes. (MarkTechPost)
NVFP4 quantization in ComfyUI: Native support for NVIDIA's 4-bit floating point on RTX 50 series. Real-world impact: models that required 24GB VRAM now fit in ~10GB with negligible quality loss. Combined with 8-step distilled models (LTX-2.3, Wan2.2 GGUF), local video gen is becoming viable on consumer hardware. (NVIDIA)
Seedance 2.0's unified audio-video architecture: The frame-level audio awareness means sound design doesn't need a separate pass — footsteps, ambient, dialogue, and music generated in a single forward pass. Competitive advantage for short-form content where audio sync is historically the weakest link.
Wan2.2 GGUF quantization: Community-contributed GGUF format of the 14B model now on HuggingFace, enabling efficient local inference. Pairs with your existing ComfyUI Wan setup. (HuggingFace)
Worth Watching
Sora's death reshuffles the market. ~500K users need new homes. Runway Gen-4.5, Veo 3.1, and Kling 3.0 are the most-cited migration targets. The API staying live until Sept 24 gives developers a window, but pipeline dependencies should be migrated now. (ZSky AI)
HappyHorse commercial launch in May will be the next inflection point — currently API-only via fal. If pricing undercuts Runway/Kling (Alibaba can subsidize), expect rapid adoption. (Meyka)
Veo 3.1 free tier is a strategic play by Google to build distribution before monetizing. Creators building workflows on it now may face paywalls later — but the 4K upscaler accepting any source video is a genuinely useful tool regardless.
TikTok's AI content suppression + Creator Rewards exclusion is the biggest platform risk for pure AI content plays. The arbitrage is moving to YouTube (permissive) and agency/B2B work (doesn't depend on platform algorithms). Instagram sits in the middle — AI content with enough human creative layer still gets reach.
Creator economy at $280B projected for 2026 — but the 32% of US/UK consumers who view AI as "negatively disrupting" the creator economy is a sentiment headwind for brand partnerships. (ngram)