Here's your briefing, Kilroy:
---
🎬 AI Video Intel — 2026-04-24 06:45
Model & Tool Releases
• Sora dies in 2 days. OpenAI's consumer app shuts down April 26; API follows September 24. Cited reasons: compute costs, user decline (1M → <500K MAU), copyright exposure, and pre-IPO resource consolidation. Disney's $1B character integration deal — announced December 2025 — was abandoned with less than an hour's notice to Disney. Export your data now or lose it. (OpenAI Help Center, The Decoder)
• Seedance 2.0 public launch (April 15). ByteDance opened access in 100+ countries — but explicitly excludes the US. Multimodal input (text + 9 images + 3 video clips + 3 audio tracks), native lip-sync that reviewers say beats post-production dubbing tools, 90%+ usable output rate. Now in ComfyUI with 1080p support as of April 20. (AkitaOnRails, ComfyUI Blog)
• ComfyUI batch update (April 20): ACE-Step 1.5 XL (4B-param music generation — full songs in seconds), ERNIE-Image for text rendering, Sonilo via Partner Nodes, Seedance 2.0 1080p. NVIDIA's RTX Video Super Resolution node for 4K upscaling now available. LTX-2.3 NVFP4 coming soon. (ComfyUI Changelog, NVIDIA Blog)
• Hailuo 2.3 + Media Agent. MiniMax upgraded the video model with better physics, micro-expressions, and stylization (anime, ink wash, game CG). Hailuo Video Agent → "Media Agent" for one-click multimodal creation. Also shipped Hailuo 2.3 Fast for batch generation. Same pricing as Hailuo 02. (MiniMax)
• Netflix VOID (April 3). Open-source video inpainting under Apache 2.0. Interaction-aware object removal that respects physics — erase a car from a collision and it generates the surviving car continuing down an empty road. Preferred over Runway 64.8% vs 18.4% in human eval. Needs 40GB+ VRAM. (Hugging Face, MarkTechPost)
• PixVerse V6 (March 30). 15s 1080p with native audio, 20+ cinematic lens controls (focal length, aperture, DoF, chromatic aberration), multi-shot short films from a single prompt. Starts $0.22/gen on Segmind, no subscription. (PixVerse)
• Wan 2.7 shipped late March — 4K, 30-second sequences, native lip-sync, first/last frame generation, 9-grid I2V, instruction-based editing. Open-source. (WaveSpeedAI)
• Runway Gen-4.5 still holds #1 on Artificial Analysis T2V benchmark (1,247 Elo). [older context — December 2025 release, no new update this week]
Platform Signals
• TikTok: C2PA auto-detection is live. Unlabeled AI content that gets flagged retroactively eats a 12–48 hour distribution hold — early engagement signals frozen, effectively killing FYP momentum. Users can now filter AIGC volume via "Manage Topics." The label itself isn't a negative ranking signal, but user behavior around labeled content creates a performance gap in practice. (AuditSocials, Storrito)
• YouTube: Early 2026 wave of faceless AI channel monetization suspensions under the "inauthentic content" policy. AI narration and AI-assisted editing remain allowed — what triggers suspension is mass-produced content with no original editorial direction. YouTube clarified faceless channels are not banned. Finance/productivity/health niches still monetizing at $10–$30 RPM. (MiraFlow)
• Instagram/Meta: C2PA metadata auto-detection → "Made with AI" label appears automatically. No evidence of distribution penalty beyond user perception effects.
• Cross-platform: All three platforms now auto-detect via C2PA. The operational takeaway: label proactively or get hit with holds/flags. (AuditSocials comparison)
What's Working
• Faceless educational YouTube remains the highest-margin AI content format. Channels earning $3K–$15K/month are common; multi-channel operators report $20K–$50K+ combined. Key: must show editorial direction, not mass-gen slop. (Unkoa, Subscribr)
• AI influencer brand deals: $10K+/month reported. Forrester's 2026 AI Creator Economy Report says 74% of consumers will follow/purchase from AI creators if quality and transparency are maintained. (CommuniPass)
• Revenue diversification is mandatory: Creators with 4+ revenue streams earn 5x more than single-source. Ad RPM alone is unreliable — views convert inconsistently even at scale.
• YouTube pays best: 55% ad rev share, $1.61–$29 RPM range. TikTok Creator Rewards: $0.40–$1.00/1K views. X: $0.50–$5 RPM. (Hailuo guide)
Technique Watch
• Netflix VOID for post-production: Physics-aware object removal is a genuine workflow unlock for anyone doing compositing or cleanup. Apache 2.0 means commercial use. If you have an A100/H100, this replaces paid inpainting tools.
• Seedance 2.0 native lip-sync eliminates the separate audio pipeline for most use cases. 90%+ success rate means less re-rolling. The US exclusion is the bottleneck — access requires VPN or API workarounds through third-party platforms like fal.ai.
• PixVerse V6 cinematic controls: Focal length, aperture, DoF, lens distortion as generation parameters — not post-processing. Multi-shot + native audio from a single prompt for product ads.
• ComfyUI App View: Simplified interface that hides node complexity. NVIDIA RTX Super Resolution node for 4K upscaling directly in the pipeline. Targeting non-technical users and rapid prototyping.
• Wan 2.7 first/last frame generation: Lets you bookend a video with specific compositions. Combined with 9-grid I2V and instruction-based editing, this is becoming the most flexible open-source option.
Worth Watching
• Sora content deadline: April 26. All user data permanently deleted post-shutdown. If you or clients have assets there, export now.
• Seedance 2.0 US availability: ByteDance is blocking US access directly. API access via fal.ai and other intermediaries is currently the workaround. Watch for regulatory or policy shifts.
• YouTube inauthentic content enforcement is tightening. The line between "AI-assisted" (allowed) and "AI mass-produced" (suspended) is editorial direction. Channels with scripted narration, curated sources, and deliberate editing survive; auto-generated content farms don't.
• DeepSeek V4 Flash/Pro announced today (April 24) — 1M token context, hybrid attention. Not video-specific but the architecture advances feed downstream into multimodal models. (Bloomberg)
• Modality convergence accelerating: Netflix (video inpainting), Cohere (ASR), Google (any-to-any) — the "multimodal" category is fragmenting into specialized pipelines. Expect more single-purpose open-source models displacing general-purpose paid tools. (HF State of OS Spring 2026)