ComfyUI shipped v0.25.0 on June 16, adding native support for the Bernini-R Wan video model, the SCAIL-2 character-replacement model for video editing, and Depth Anything 3 for monocular depth. The same drop added core 10-bit video nodes plus 3D preview nodes (PreviewGaussianSplat, PreviewPointCloud), and a same-day v0.25.1 patch wired in Kling V3-Turbo as a partner node. This is the most directly actionable update on the list — it's free, local, and live in your graph today.
Lightricks released the LTX Trainer on June 17, exposing 13 training modes for its open-weight LTX-2 stack — LoRA, IC-LoRA, and full fine-tune — through a single YAML config. LTX-2.3 Fast currently sits at #1 among open-weight models on the Artificial Analysis Text-to-Video Arena at 1121 Elo, roughly 10 points ahead of Wan 2.2, and the 22B model runs on consumer GPUs with FP8. For creators who want a consistent character or house style, this is the cleanest open path to a custom audio-video model.
On June 18, Runway rolled out Studio Trim — trim, stitch, reorder, and export a final cut in one place, across all tiers — cutting the round-trip out to a separate editor. That follows June 5's addition of Seedance 2.0 Fast to the Runway API (text, image, or video in, with keyframe control, reference media, and generated audio, 4–15 second clips). The combo means you can now generate and assemble inside one platform.
ByteDance's Seedance 2.0 remains the model to beat on the Artificial Analysis with-audio board at 1213 Elo, accepting up to 9 images, 3 clips, and 3 audio inputs per generation with phoneme-perfect lip-sync in 8+ languages. It's now broadly accessible — live on fal since April, on PixVerse with native audio and two speed tiers, and discounted up to 70% through June 25. If your route needs synced dialogue, this is the current top pick.
Kuaishou has pushed Kling 3.0 Turbo plus an upgraded Omni as of June 17, building on the 3.0 base that brought native 4K output, built-in multilingual audio, and a physics-aware engine. Multi-shot storyboard work runs around $0.10/second, keeping Kling at the top of the value-to-quality bracket. The Turbo tier is the practical win — faster iteration at the price point creators actually use.
YouTube's renamed inauthentic-content policy is in active enforcement, targeting mass-produced, templated, low-effort uploads regardless of how they're made — in January the platform terminated 16 channels carrying a combined 35 million subscribers and 4.7 billion lifetime views. Critically, properly disclosed AI content is not penalized: correctly labeled videos still get normal distribution. The takeaway for AI creators is craft and disclosure beat volume.
TikTok's automated detector — analyzing voice patterns, visual artifacts, frame consistency, and tool metadata signatures — flags AI clips for mandatory review, and three unlabeled AI videos now cut account reach by about 60% for 30 days and pause Creator Fund earnings. Meta unified its Instagram and Facebook "Made with AI" disclosure in February. With the EU AI Act's Article 50 labeling enforcement starting August 2, labeling discipline is now a survival skill, not a nicety.
Fresh monetization data underscores that niche choice dwarfs view count: 100,000 views in finance earns about $1,500 versus roughly $185 in gaming — an 8x swing — while Shorts revenue-share runs a thin $0.03–$0.07 per thousand. YouTube's 2026 on-ramp now lets you apply at 500 subscribers with 3 million Shorts views in 90 days, but full ad revenue still needs 1,000 subs and 10 million valid Shorts views. Faceless educational video remains the highest-yield AI format.
A reality check for the local crowd: Alibaba's Wan 2.6 (1080p/24fps, up to 15s, single-pass synced audio and lip movement) shipped in December but, like Wan 2.5-Preview, never published open weights — Wan 2.2 stays the latest self-deployable version. With the new Bernini-R Wan model now in ComfyUI and Wan 2.7 rumored, the open-weight gap is the storyline to track for anyone building a local pipeline.
