Here's your briefing:
---
🎬 AI Video Intel — 2026-04-30 06:45
Model & Tool Releases
Sora is dead. OpenAI shut down sora.com and mobile apps on April 26. API stays live until Sept 24, then full sunset. Reported burn rate was ~$1M/day in compute. All user data deletes post-shutdown — export now. (The Decoder, ThePlanetTools)
Kling 3.0 launched April 24. Native 4K generation (no upscaling), unified audio-video pipeline, multi-shot sequencing. Claims #1 ELO benchmark score (1243). 30 credits/sec for 4K. Promo pricing through May 7, entry tier at $6.99/mo. OpenArt reports signups doubled within 10 hours. (Kombitz, VO3 AI, CineD)
Veo 3.1 now free for all Google accounts (10 gens/month, 720p, 8s clips). Veo 3.1 Lite launched April 3 on Vertex AI as a cost-optimized tier. Three-tier family: full / Fast / Lite, all with native audio. New standalone upscaling capability on Vertex. (Google Cloud Blog, Chrome Unboxed)
LTX 2.3 (Lightricks, March 5): 22B DiT, generates synchronized audio+video in one pass, up to 4K@50fps, 20s clips. Open-source with permissive license. Two variants: ltx-2-3-fast (iteration) and ltx-2-3-pro (production). Native vertical 1080×1920 composition. Note: LTX-2 LoRAs do NOT transfer to 2.3 — retrain required. (LTX official, WaveSpeed)
Seedance 2.0 (ByteDance, Feb 9): 60-90s video generation, native audio, multi-shot cuts. Available on CapCut/Dreamina with C2PA watermarking. API on fal.ai as of April 9. Explicitly excludes the US from availability. (ByteDance, fal.ai, AIVidPipeline)
PikaStream 1.0 launched April 2 — real-time visual engine for AI agents. Pika 2.5 remains competitive for short-form iteration with improved physics effects (Pikaffects), auto-generated synced SFX, and upgraded lipsync. (Progressive Robot, Genra)
Alibaba's new video model reportedly topped global rankings in late April for compute efficiency and visual quality (complex lighting, fabric physics, water reflections). Details sparse — watch for formal announcement. (Crypto Integrated AI News)
Grok/xAI rolling out shareable Imagine templates with Photo-to-Video, Photo-to-Style-Edit, and unlimited free AI video with consistent characters. (Crypto Integrated)
Runway Gen-4.5: No new April release. Current state: I2V + T2V, 2-10s clips, all paid plans. GWM-1 world model + robotics SDK announced but not an April update.
---
Platform Signals
TikTok is the most aggressive. As of March 2026: mandatory AI-generated label using C2PA Content Credentials for auto-detection. Posting unlabeled AI content triggers an immediate strike + 73% reach suppression within 48 hours. AI content is explicitly excluded from the Creator Rewards Program. However, properly labeled high-quality AI content (65%+ completion rate, 0.8%+ comment-to-view ratio) performs within 5-8% of organic reach. (Storrito, AuditSocials, Virvid)
Instagram/Meta unified AI disclosure across IG and Facebook in Feb 2026 — visible tag above content, similar to paid partnership labels. Originality Score system penalizes recycled/duplicate/low-effort content. No outright AI content ban but reduced reach for templated output. (Influencer Marketing Hub)
YouTube remains the most creator-friendly. AI content is fully eligible for monetization (standard YPP requirements: 1K subs, 4K watch hours). Must disclose realistic synthetic content. Key distinction: AI-as-tool = fine; AI-as-replacement-for-creativity = demonetization risk under "inauthentic content" policy (renamed from "repetitious content" July 2025). Mass-produced template content with minimal variation gets flagged regardless of views. (Virvid, BossWallah)
---
What's Working
YouTube AI history/explainer channels: $5-15 RPM. A channel at 500K monthly views earns $2,500-$7,500/mo from ads alone. Top performers hitting $10K+/mo. The niche rewards depth and narration quality — exactly where AI video supplements rather than replaces. (AI Monks / Medium)
Volume iteration strategy: Creators identifying viral formats, generating 50+ variations with different music/characters/styles. Most get 10K-50K views, breakouts hit 500K-1M. At $3-5 CPM, a single 1M-view hit = $3,000-$5,000. (AI Monks)
AI influencer sponsorships: Virtual influencers with 100K+ followers commanding $2,000-$15,000 per sponsored post. IP licensing deals averaging $45K-$120K for major brand campaigns. (CommuniPass)
Agency play: One agency scaled video revenue 340% (€180K → €792K) in year one by cutting per-video pricing 30% and increasing output 10x with AI production. (MindStudio)
Market signal: 79% of marketers increased ad spend on AI creator content; 77% say they'll shift budgets away from traditional creator marketing toward AI-generated content. (FinancialContent)
---
Technique Watch
LTX 2.3 ID-LoRA in ComfyUI: Identity-controlled video generation — train a face LoRA, combine with reference audio for consistent talking-head video. Native support merged into upstream ComfyUI (PR #13111). Adds LTXVReferenceAudio node for speaker identity transfer. (RunComfy, GitHub)
LTX 2.3 IC-LoRA: In-context LoRA for precise motion/geometry control — V2V motion tracking workflow. Structured guidance injection during generation. (RunComfy)
Consumer GPU LoRA training is practical: @ostrisai trained an LTX 2.3 character LoRA on a 5090 in ~1 day, generated 30s video in 6 minutes with ComfyUI. AI Toolkit is the workflow. (X/Ostris)
ModelScope free tier: 2,000 free API calls/day across 900+ models, plus free LoRA training on Wan2.2, Flux, and others. Free GPU app hosting. Useful for experimentation without local hardware. (X/Rohan Paul)
Distilled LoRA for LTX 2.3: Official adapter that stabilizes and speeds sampling without quality loss. Good for production pipelines. (ComfyUI Docs)
Migration warning: LTX-2 LoRAs break on 2.3 — embedding space, attention blocks, and VAE changes mean learned deltas don't map. Full retrain required. (WaveSpeed)
---
Worth Watching
Sora's death reshuffles the market. The $1M/day compute burn and shutdown validates the thesis that closed-source video generation at scale is economically fragile. Open-source (LTX, Wan) and efficient commercial (Kling, Seedance) models inherit the demand. API users have until Sept 24 to migrate.
Alibaba's unnamed model topping global benchmarks in late April could be the next open-weight release to watch. Their track record (I2VGen-XL, Qwen) suggests it may land on HuggingFace.
Seedance US exclusion creates an arbitrage — if you can access it via CapCut/Dreamina in non-US regions or through fal.ai's API, it's reportedly the strongest for 60-90s social content with native cuts.
TikTok's C2PA auto-detection means metadata stripping is becoming a cat-and-mouse game. Creators who label properly and optimize for engagement metrics (completion, comments) aren't getting crushed — the penalty is for deception, not generation.
YouTube remains the money platform for AI video. Full monetization eligibility + highest RPMs ($5-15 for educational/history niches) vs. TikTok's explicit exclusion from Creator Rewards. If you're building a revenue-generating AI content operation, YouTube is the primary channel, TikTok is for distribution/awareness only.