ByteDance just announced Seedance 2.5 on stage at Volcano Engine 2026 this morning — a single native 30-second clip in one pass, accepting up to 50 multimodal references (images, video, and audio). It's in enterprise beta now with public launch slated for early July, so the headline spec is real but you can't queue jobs yet. Heads up: any "Seedance 2.5" Elo score floating around is fabricated — the only legit leaderboard numbers belong to 2.0, which still tops the image-to-video boards at roughly Elo 1,195 with audio.
Kuaishou shipped Kling 3.0 Turbo last Wednesday at 18:00 CST — a fast-preview mode that spits out 1-to-15-second drafts from text or multi-image references at 480p/720p, priced from ¥0.8/sec for 720p and ¥1/sec for 1080p with audio baked in (roughly $0.11 and $0.14 per second). Alongside it, Kling 3.0 Omni got an editing-pipeline bump to handle 3-to-15-second clips and 4K input/output. Translation: rough out concepts on Turbo cheap, then commit your winner to full 3.0 for production-grade audio and lip-sync.
Z.ai's SCAIL-2 dropped open-weights and got merged into ComfyUI (kijai's PR landed June 9), bringing end-to-end motion transfer and cross-identity character replacement with no intermediate pose rig. Set replacement_mode to true and it swaps the person in a driving video for your reference character, with animal-driving and zero-shot SAM3D mesh control as emergent tricks. A community node, scail-auto-extend, already handles arbitrary-length generation via automatic chunking, anchoring, color matching, and stitching.
Aleph 2.0 — Runway's flagship video-to-video editor — came to the Figma Weave canvas on June 18, after going live via the Runway API on June 2. It edits existing footage from text prompts plus up to 5 keyframe images placed at specific timestamps, handles 2-to-30-second inputs, and carries an edit through every frame a subject appears in. That's the post-production play: product swaps, packaging changes, relighting, and background cleanup without re-rolling the whole clip.
As of May 2026, YouTube rolled out automatic detection that identifies photorealistic synthetic media and slaps on the "Altered or Synthetic Content" label even when creators forget. Pair that with the January wave that terminated 16 channels holding 4.7 billion lifetime views, 35 million subscribers, and an estimated $10 million in annual ad revenue under the inauthentic-content policy. Disclosed, human-shaped AI work still monetizes at full rate — templated copy-paste slop is what's getting torched.
TikTok now runs an automated scanner on uploads — checking voice patterns, visual artifacts, frame consistency, and tool metadata signatures — and flagged videos get mandatory human review before they publish. Meta unified its AI rules across Instagram and Facebook in February, requiring an in-post "Made with AI" disclosure separate from the caption, shown as a tag above the content. The upside: properly labeled AI content is fully ad-eligible and, per the platforms, does not reduce algorithmic reach.
The numbers people are passing around: Noah Morris runs roughly 20 faceless channels totaling 2.5 million-plus subscribers, with one court-case video costing $250 to make and pulling over $20,000 off 5 million views. Production economics have collapsed to under $3 per video, with competitive monthly tool stacks running $47 to $180. High-RPM lanes still rule — finance at $15-$40 per thousand views, true crime $8-$13 RPM — but those same templated channels are exactly what the platform crackdowns are hunting.
ComfyUI 0.25.0 dropped around June 16, bundling the new LatentCut node for slicing latents at precise points — genuinely handy for temporal video work like dropping specific frames or chunked batch processing. The same release wave folded in SCAIL-2 model support and a Wan2.5 image-to-image API node. If you're building local pipelines, this is the update that turns the past two weeks of model drops into actual node-graph workflows.
For the self-hosting crowd, the open-weight tier keeps maturing: Wan 2.6 ships 1080p up to 15 seconds with native audio-visual sync under Apache 2.0 — meaning commercial and client work with zero licensing fees. LTX-Video still generates 30fps at 1216x704 faster than real-time, and Tencent's HunyuanVideo holds quality-leader status. The real unlock is GGUF quantization dragging VRAM needs from 60GB-plus down under 10GB, putting these on consumer cards instead of just A6000s.
