Released today, September 9, v0.35.0 adds LTXV generated-keyframe nodes, native video trimming and cropping, color-space conversion, AVIF output, and TRELLIS memory optimizations on top of the Comfy Compiler. It also brings open-source Pixal3D mesh reconstruction and SenseNova U1.5 pixel-space generation into core. For anyone doing multi-shot work, keyframe control plus in-graph trimming means fewer round trips out to an NLE.
OpenAI kills the Sora API on September 24, 2026, five months after the consumer app died on April 26 following an 84-day run. OpenAI's own migration table leaves the "recommended replacement" column blank — there is no in-house successor, so every path leads to another vendor. If any part of your pipeline still calls Sora endpoints, this is a hard cutoff with no grace period and no data recovery.
Shipped September 6, ComfyUI's new core sparse attention node ships with three backends tuned for MiniMax H3, paired with a memory compiler that reclaims wasted VRAM and speeds up long generations. H3 is a 33B open-weights multimodal model — 42.5 GB minimum download, 2K output with native stereo audio up to 15 seconds. This is the difference between H3 being a cloud-only toy and something you can actually run on a workstation.
As of the September 1 rollout, Instagram is requiring accounts built around an AI persona to carry an "AI-generated profile" label — skip it and the algorithm marks your content non-recommendable, which torches both reach and monetization. Creators who use AI tools but still appear on camera themselves are exempt. Note the Reels Play Bonus is now invite-only, so affiliate links in Reels and direct brand deals are the real revenue paths.
Dropped September 5, RefMods are roughly 1 MB conditioning adapters for MiniMax H3 — encode a character once, then reuse it across generations without re-referencing source images every time. A companion 11 MB "H3 Semantic Bridge" adapter ships with its own nodes and workflows to tighten prompt adherence. Also out that day: H3 Singularity, a fine-tune targeting HDR clarity, distant-face restoration, and stronger action motion.
Posted September 7, AI Movie Studio 2 is an open-source workstation driving local ComfyUI with a 3D storyboard, a multi-reference continuity system, Long Take mode, and XML timeline export into your editor. It's model-agnostic — local ComfyUI plus Fal and Replicate cloud drivers — with built-in LoRA support on every local driver. Positioned squarely as the free alternative to LTX Studio and Storyboarder.
Muse Image is now callable directly on Meta's Model API at $0.01 per image, after months of powering Meta AI internally. It's an agentic model — it decomposes multi-part prompts and refines in a chain of thought before rendering, and it'll hit web search for factual prompts. ComfyUI added official Muse Image nodes on September 4 in v0.34.4 with text-to-image, editing, reference-image support, and 14 aspect ratios.
The September 7 release added GPT-6 Astra and reasoning-effort levels to OpenAI Chat nodes, but the bigger deal for production is what it removed: DALL·E versions, Kling Video Extend, the LTX-2 partner nodes, and several ByteDance Image options are gone. If you have archived workflows leaning on any of those, they'll break on update. Pin your version or audit before you pull.
The September 5 v0.34.5 release added Comfy Cloud beta nodes that run curated Flux 2, Z-Image Turbo, Mage Flow, MiniMax H3, and Music 3 workflows on Comfy Cloud GPUs from inside your local graph. That matters for the H3 crowd — the open weights are royalty-free under the Community License, but that license excludes open-weight use in the US, EU, UK, and South Korea without separate authorization. Hosted access sidesteps the local-weights question entirely.
Alibaba's Wan 2.7 suite — four Apache 2.0 models on a shared 27B MoE backbone — debuted at 1762 Elo on Artificial Analysis text-to-video, well clear of LTX-2.3 Pro at 1484. It does first/last frame control, up to five reference video inputs, voice cloning, and instruction-based editing at 15 seconds. Alongside LTX-2.3 and HunyuanVideo 1.5, that's the trio with licenses clean enough to build a commercial pipeline on without a lawyer.
