ByteDance shipped Draft Mode for the Seedance 2.5 API, letting you rough out composition, motion, dialogue, and camera moves at 480p, then re-infer the winner at 1080p using the same draft ID — prompt, seed, aspect ratio, and duration all carried forward. ByteDance's own math: a 5-second 1080p shot that normally takes four attempts gets 57% cheaper and roughly twice as fast; at twenty attempts the savings hit 77%. ComfyUI nodes for it landed September 25.
Dropped September 25 — a character swap LoRA for MiniMax H3 that re-casts the subject of an existing reference clip while holding the original motion and timing. That's the piece the H3 edit stack was missing, and it slots straight into the VFX Edit, Style Transfer, and Head Swap LoRAs from September 22. For anyone running a faceless channel with recurring characters, this is the consistency lever.
September 24 was a Qwen-Image 2.1 avalanche: the official Fun Acc LoRA collapses generation to four steps in a 346 MB PDD file, and Fun ControlNet Union packs eight structural controls plus inpainting into a single 7.5 GB branch. Add the Pruna 5-step and 8-step no-CFG LoRAs from the 23rd and a community Fix LoRA for washed-out color, and the whole pipeline just got dramatically cheaper per frame. ComfyUI v0.37.0 on September 20 added native 2.1 support with three official templates.
Also September 24: a retrained VAE decoder that eliminates the checkerboard artifacting people have been fighting in Qwen-Image 2.1 output. It's a straight decoder swap — no retraining, no prompt gymnastics — which makes it the highest ratio of quality gained to effort spent this week. If you've been upscaling to hide the pattern, stop doing that.
Lightricks' 22B open-weights audio-video model is sitting at roughly 5,100 likes, leading all video models on HF trending for the week of September 25 — well ahead of the rest of the generative-media pack. It does joint video and audio through a dual-stream diffusion transformer, up to 4K and 50fps, multi-shot in one generation, and a 10-second 720p clip in 6.8 seconds on dual GB200s. The community license is free under $10M ARR with no mandatory branding.
OpenAI's Sora API shut down September 24, following the web and app experiences on April 26. Anything in your pipeline still calling Sora endpoints is now failing silently or hard. The migration targets people are actually landing on: Veo 3.1 for synced dialogue, Kling 3.0 Turbo on cost, Seedance 2.5 for long-form image-to-video, Runway Gen-4.5 for control surfaces.
From September 20, and still underrated: a LoRA that re-renders existing MiniMax H3 footage from a new camera angle using a MoGe depth warp. That's coverage generation from a single take — get your establishing shot, then manufacture the reverse and the three-quarter without burning a new generation lottery ticket. ComfyUI v0.37.0 shipped MoGe 3 support the same day, so the depth half of the pipeline is native now.
Fresh September 2026 figures put AI Shorts RPM at roughly one to seven cents per thousand views, with creators keeping 45% of their allocated pool share. Translation: a million-view AI Short is a $30 to $100 event, and channels doing 10 to 50 million monthly views land in the $200 to $2,000 range. The Partner Program bar is still 1,000 subs plus 10 million qualified Shorts views in 90 days.
TikTok is now scanning uploads for AI characteristics — voice patterns, visual artifacts, frame consistency, and tool metadata signatures — and applying labels via C2PA Content Credentials whether or not you disclose. YouTube's AI-use attribute covers videos, Shorts, and livestreams where realism could mislead; Meta unified Instagram and Facebook disclosure back in February with an above-post tag. The practical upshot: label proactively, because getting caught not labeling is the expensive path.
September 25 brought a node that pulls single-frame stills out of MiniMax H3 at up to 3 MP. It's a small thing that solves a real problem — thumbnails, key art, and reference frames that actually match your video model's look instead of being generated by a separate image model with a separate aesthetic. Pair it with obvpm Timeline from September 21 for extend, prepend, bridge, and loop editing inside ComfyUI.
