Alibaba's Wan 2.7 is now live in ComfyUI via Partner Nodes, with upgrades across image quality, audio, motion dynamics, stylization, and consistency. The headline for production work: support for up to five real-person image inputs, vocal-timbre references, and a full menu of task types including first-frame, first-plus-last-frame, audio-driven image-to-video, video continuation, reference-to-video, and video edit. The one caveat is the open-weight question — for now this is API-based access through the nodes, so local self-hosters are still waiting on downloadable checkpoints.
Tencent's lightweight 8.3B-parameter HunyuanVideo 1.5 now has native ComfyUI support with engineering optimizations for fast inference, and it's fully open-weight. With offloading and VAE tiling on, the pipeline runs 720p 121-frame text-to-video and image-to-video at a peak 13.6GB VRAM — comfortably inside an RTX 4090. A step-distilled 480p I2V model cuts end-to-end generation time by roughly 75 percent, landing clips in about 75 seconds on a single 4090. For local creators, this is the most capable open model you can actually run today.
Wan 2.6's Reference-to-Video mode is available in ComfyUI, letting you feed a 2-to-30-second reference clip and have the model extract a character's appearance, motion patterns, and voice to generate new footage with the same identity. It delivers precise lip-sync and motion transfer, and supports up to three simultaneous references for multi-character scenes with preserved identities. Clips run up to 15 seconds. For anyone building recurring AI characters, this is the consistency tool to test this week.
OpenAI discontinued the Sora web and app experiences on April 26, 2026, and the Sora API will follow on September 24, 2026, after which all associated account data is permanently deleted. Sam Altman reportedly told staff OpenAI is winding down all products built on its video models. If any part of your pipeline still calls Sora 2, you have a hard migration deadline — and you should export any stored content before the cutoff.
As of March 2026, all AI-generated videos on TikTok must carry the built-in 'AI-generated' label, and the platform deployed automated detection that scans uploads for voice patterns, visual artifacts, frame consistency, and metadata signatures from known AI tools. Content that realistically depicts people, places, or events without disclosure risks removal. The practical takeaway: label proactively in the posting interface, and strip or expect the detection of tool metadata in your exports.
Meta unified its AI-content policy across Instagram and Facebook in February 2026, requiring in-post disclosure separate from the caption — a tag above the content, styled like a paid-partnership label. The cross-platform gotcha is that disclosure rules don't line up: a clip that needs a label on TikTok may need none on Instagram or YouTube, and vice versa. If you cross-post the same AI video everywhere, you now need a per-platform labeling checklist instead of one blanket setting.
After thousands of faceless AI channels saw monetization suspended under the inauthentic-content policy earlier in 2026, YouTube clarified that the rule targets low-effort, mass-produced, templated output — not AI-assisted creators who add their own insight. The platform frames it as enforcement of a years-old policy now sharpened by better AI detection, not a new restriction. If your channel ships identical templated videos at scale, you're exposed; if AI is speeding up genuinely differentiated work, you're fine.
ByteDance rolled out Seedance 2.0 Mini in June 2026, a faster, lower-cost tier in the Seedance 2.0 family, launching first on Dreamina before wider availability. The broader Seedance 2.0 line currently offers 100 free daily credits — roughly 10 to 20 watermark-free 1080p clips a day, no card required — with per-10-second cost ranging from about $0.22 on fal.ai's Fast endpoint up to $3.20 in Dreamina credits. For high-volume short-form creators chasing cost-per-clip, the Mini tier is worth a benchmark pass.
