Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Comfy-Org dropped v0.34.5 this morning with beta nodes that push curated workflows onto Comfy Cloud GPUs, covering Flux 2, Z-Image Turbo, Mage Flow, MiniMax H3 and Music 3. That's a real escape hatch if your 4090 is pinned rendering overnight batches — same graph, rented silicon. Z-Image Turbo remains the speed pick here at 8 steps out of Alibaba's Tongyi Lab, up to 2048x2048 native.
OpenAI kills the Videos API and every Sora 2 alias — sora-2, sora-2-pro, and all dated snapshots — on September 24, 2026, with no replacement model in the docs. The app already went dark April 26. Every migration path leads to a different vendor, so this isn't a model bump, it's a rewrite: Gemini Omni, Veo 3.1, Kling 3.0 at roughly ten cents a second, or Seedance 2.0 self-serve from fifteen bucks.
Instagram renamed its May "AI creator" tag to "AI-generated profile" and is now telling accounts built on AI personas to apply it — or get marked non-recommendable, which guts reach and downstream monetization. That's a reversal from May, when Meta explicitly said the label wouldn't affect recommendations. Critically, if you use AI tools but appear as yourself on camera, you do not need the label.
Yesterday's v0.34.4 wired Meta's Muse Image into ComfyUI for text-to-image and editing with up to 10 reference images — the multi-reference character-consistency play. Muse Image is agentic: it decomposes prompts, refines in a chain of thought, and hits web search for factual prompts. At one cent per image and the No. 2 slot on Arena for text-to-image and single- and multi-image editing, it's cheap enough for keyframe farming.
v0.34.3 on Tuesday added H3 Max to text-to-video, first-last-frame and reference nodes; v0.34.4 followed with a turbo variant. H3 Max is the fal-post-trained speed build — in head-to-head testing it turned a 15-second 768p clip in about 18 seconds, versus roughly five minutes for a 15-second 1080p Seedance 2.5 render. H3 caps at 15 seconds and 12 input slots, so it's a draft-and-iterate engine, not a finisher.
Alibaba's speed-tuned Wan 3.0 Prime landed in ComfyUI v0.34.1 and runs up to 7x faster than base Wan 3.0, doing 2-to-30-second clips at 480p through 1080p with a matching audio track and first-frame steering. Pricing runs 6.8 cents per second at 480p to 28 cents at 1080p. Same release quietly retired the legacy Veo nodes — check your old graphs before you hit queue.
Wan 3.0 entered public beta August 7 with native 30-second 1080p single-pass clips, in-pass audio, and document-to-video from PDF, PPT, XLS and URLs up to 100MB or 50 pages. But it shipped with no technical report, no model card, no parameter count, and no weights. For local pipelines, Wan 2.2 from July 2025 remains the last true open-weight flagship in the line.
Google DeepMind pushed Omni 1.1 Flash to general availability August 27, adding scene extension, first-and-last-frame interpolation, reference clips and selectable output resolution — and ComfyUI's v0.34.2 shipped support the same day with 4K output. It also ships free inside YouTube Shorts and the YouTube Create app for conversational editing. Scene extension is the quiet headline: it's the cheapest fix for the 15-second ceiling everywhere else.
Shorts RPM sits at three to five cents per thousand views in 2026 — roughly $10 to $60 per million. Channels blending AI production with a visible human presence are outperforming pure-AI channels on every monetization metric, and about 21% of new-user feeds now get flagged as AI-generated filler. Faceless documentary and exam-prep channels reported monetization pausing overnight — the trigger is scale without a human fingerprint, not AI itself.
The v0.34.0 major release added PixVerse V6 video nodes with native audio, plus HDR video saving with the AV1 codec — a genuinely useful delivery upgrade for anyone finishing in a real NLE instead of shipping straight to a feed. Same drop added Pixal3D, TRELLIS2 mesh generation, and SAM 3D Body segmentation for anyone stitching 3D into their video pipeline.
