Now I have enough to compile the briefing. Here it is:
---
🎬 AI Video Intel — 2026-04-22 06:45
Model & Tool Releases
• Happy Horse 1.0 (Alibaba/Taotian Future Life Lab) — The biggest shakeup this month. Appeared anonymously on Artificial Analysis Video Arena ~April 7, now confirmed as Alibaba's work. Elo 1333–1387 T2V (#1) and 1391–1406 I2V (#1, all-time record), beating Seedance 2.0 by 74 Elo — the largest gap in leaderboard history. 15B param, native audio-visual in a single pass. Led by Zhang Di, ex-Kuaishou VP who built Kling. Going open-source. (CNBC, WaveSpeed)
• Sora dies April 26 — Four days out. Web + app go dark Saturday. API stays until Sept 24. Downloads dropped 45%, Disney killed a $1B deal, active users cratered from 1M to under 500K. Export your content now. Kling saw +4% WAU to 2.6M in the week after the announcement. (OpenAI Help Center, TechCrunch)
• Veo 3.1 now free for all Google accounts — 10 clips/month via Google Vids, ~12/day via Google Flow (50 daily credits). 8-sec max on free tier. Veo 3.1 Lite launched March 31 for devs at <50% cost of Veo 3.1 Fast. 720p/1080p, native audio, scene extension for 60s+ videos. (Google Blog, MindWired)
• ComfyUI v0.19.3 (April 17) + April 20 partner nodes — LTX text generation updates, JsonExtractString node, OOM fix for quantized models. April 20: Sonilo (video-to-music, frame-synced soundtracks), ERNIE-Image (precise text rendering in images), ACE-Step 1.5 XL (commercial-grade music gen) all shipped as partner nodes. (ComfyUI Changelog, ComfyUI Blog)
• Netflix VOID [older context, April 3] — Open-source (Apache 2.0) video object removal + physics-aware inpainting. Built on CogVideoX. Preferred over Runway 64.8% vs 18.4% in human eval. Requires 40GB+ VRAM. (MarkTechPost, HuggingFace)
• PikaStream 1.0 [older context, April 2] — Pika pivoted hard from offline gen to real-time video chat for AI agents. 24fps, 1.5s latency, one GPU. AI Selves can join Google Meet. Different product category than traditional T2V. (Pika Blog)
• Hailuo 2.3 + Media Agent — MiniMax shipped better micro-expressions, wider art styles, Fast variant at 50% cost reduction. Media Agent (multi-modal creation) rolled out worldwide across web, mobile, and API. (MiniMax)
• fal launched enterprise API for GPT Image 2 (April 21) — OpenAI's next-gen image model now available via fal with T2I, editing, upscaling, and video gen. (EIN Presswire)
Platform Signals
• TikTok: AI content banned from Creator Rewards Program. Fully AI-generated videos (visuals + voiceover) cannot monetize. AI-assisted workflows (color correction, auto-captions, AI B-roll) remain eligible. The line: "AI as a tool" = OK, "AI as the producer" = banned. Repeated unlabeled AI violations = permanent loss, no reinstatement. 1.3B+ videos already labeled via C2PA. (Storrito, Napolify)
• TikTok: AI slop slider launched. Users can now reduce AI content on their FYP via "Manage Topics" — slider-based, less/more, not on/off. Platform is actively downranking content flagged as "synthetic slop." (Daily Dot, Techlicious)
• YouTube: "Inauthentic content" policy still in effect (renamed from "repetitious content" July 2025). AI-generated scripts + stock footage + recycled clips = demonetized. AI as creative tool with creator personality = fine. Must disclose altered/synthetic content that could be mistaken for real events. (BossWallah, Supertone)
• TikTok algorithm now demands 70%+ completion rate to go viral (up from 50% in 2024). Watch time is the top signal, followed by search relevance and originality. Instagram prioritizes "sends per reach" (DM shares). (Virvid)
What's Working
• AI avatar + TikTok Shop is still printing: one creator case study shows ~$13K commission from a single AI avatar product video. (AIImageToVideo)
• "Slightly uncanny" performs. Content that feels slightly off/unnatural is interrupting scroll and driving engagement on TikTok — the weirdness is the hook, not the bug. (AIImageToVideo)
• Virtual influencers in tech/fashion/fitness commanding $2K–$15K per sponsored post at 100K+ followers. IP licensing deals averaging $45K–$120K for large brand campaigns. (CommuniPass)
• Paid communities outperforming all other creator monetization models for recurring revenue in 2026. AI creators pivoting from platform payouts to subscriptions. (CommuniPass)
• 32% of US/UK consumers view AI as "negatively disrupting" the creator economy — a headwind to track. (Logie)
Technique Watch
• Sonilo video-to-music in ComfyUI — reads visual pacing + emotional cues, syncs soundtrack frame-by-frame. Eliminates the manual audio-matching step in AI video workflows. (ComfyUI on X)
• Netflix VOID quadmask conditioning — 4-value mask system (primary object, overlap, affected regions, background) for interaction-aware removal. Removes a person AND physics-correctly handles objects they were holding. Useful for cleanup/VFX workflows. (MarkTechPost)
• Wan 2.7 subject+voice reference — 150 reference frames for character consistency across clips, preserving face/skin/hair/clothing/voice. First-frame and last-frame generation for precise narrative control. (WaveSpeed)
• Kling 3.0 AI Director — script-based scene transitions with auto-scheduled camera angles (shot/reverse shot), structured multi-shot sequences up to 15s in a single generation cycle. (Kling3.org)
Worth Watching
• Happy Horse going open-source — If the #1 leaderboard model actually drops weights publicly, it rewrites the ComfyUI/local-gen landscape overnight. Watch for the HuggingFace drop. (CTOL Digital)
• Sora API sunset Sept 24 — Any production workflows still on Sora API have ~5 months. Migration targets: Runway Gen-4.5 (quality), Veo 3.1 (free tier + native audio), Kling 3.0 (multi-shot), Seedance 2.0 (cost). (The Decoder)
• Seedance/Disney legal fight — Disney's Feb 13 cease-and-desist to ByteDance over Seedance 2.0 training data is unresolved. MPA also piling on. Could set precedent for all video model training. (Wikipedia)
• Pika's real-time pivot — PikaStream 1.0 targeting live video calls, not content gen. If this catches on with agent builders, it's a different market than the rest of the pack. (Pika Blog)
• Runway + NVIDIA Rubin partnership — Runway announced it's building on NVIDIA's next-gen Rubin platform. Likely means Gen-5 will have a significant compute advantage. (Runway)