Comfy Org launched Comfy Agent on October 1: describe what you want, and it plans the graph, drops and wires the nodes on the canvas, runs it, and debugs itself when something breaks. It supports up to 5 parallel chats with independent histories, reusable custom "skills" for repeat tasks, and simultaneous editing by both you and the agent without lockout. It's in beta, live for everyone on Comfy Cloud day one, runs on frontier Anthropic models, bills against existing Comfy Credits after free starter chats, and a Comfy Desktop build is promised in weeks.
Lightricks' LTX-2.5 sits atop the Hugging Face trending board as of today with 5,864 likes and 1.58M downloads, covering image-to-video, text-to-video, and video-to-video. It generates a 10-second 720p i2v clip in 6.8 seconds on Nvidia superchips, ships open weights, and had native ComfyUI support on day one. On September 29, Tiled Fusion landed for LTX's Native Resolution feature, pushing edits to 4K and 8K.
The September 29 core release adds native Ming-Image support, Wan ID-V2V identity-preserving video-to-video, w6a8 quantization, and HDR color space handling. The w6a8 path is the headline for anyone on a 16 or 24 gig card — it's the difference between loading a frontier backbone and watching it OOM. HDR color spaces also finally line ComfyUI up with the 10-bit output the hosted models are now producing.
Kuaishou's Kling 4.0 is scheduled for full release in October 2026, with early access rolling out to a limited group starting September 28. It does up to 30 seconds in a single pass — double Kling 3.0 — with 10 flexible keyframes and 10-bit HDR at 1080p and 4K. Kling 4.0 Flash is already in limited rollout at up to 20 seconds and 720p for cheaper, faster everyday work.
Announced September 29, LongLive-Plug distills few-step, CFG, and long-context LoRAs a single time per backbone, then reuses those adapters across 54 downstream models. For anyone maintaining a stack of fine-tunes, this collapses the retrain-per-checkpoint tax that's been the quiet cost center of running custom video pipelines. Early reports frame it as a plug-in layer rather than a new base model.
Qwen-Image 2.1 is the number two trending image model on Hugging Face right now at 2,795 likes, and the community shipped four major adapters in six days: a Consistency LoRA on October 1 that holds frames stable through edits without repainting drift, Turbo8 eight-step distillation across t2i, editing and RGBA, AnyAngle for Gaussian-splat camera rotation, and LanPaint 2.2.0 for alpha-transparency inpainting. The uncensored GGUF variant has cleared 1.3M downloads on ComfyUI compatibility alone.
H3's tooling matured fast in the last week: a Transformation LoRA on October 1 that morphs cleanly between two images, a 360 Orbit LoRA turning a single photo into a full camera orbit, Akatz Labs' Character Swap weights for replacing characters in reference video, and Prompt Builder 1.8.0 adding masked clip editing via SAM 3.1. Fizgig H3 Still also now pulls single-frame stills at 3 megapixels and up without banding.
TikTok requires visible on-video labels — not caption-only — for any AI-generated visuals or audio depicting realistic people or scenes, and it auto-detects synthetic media through C2PA Content Credentials even when you don't self-disclose. Three unlabeled AI videos triggers roughly a 60% reach reduction for 30 days plus temporary Creator Fund suspension. Meta auto-flags on Instagram and Facebook with an optional AI-creator account label added in May 2026, and the EU's August 2 requirement makes visible labeling a legal obligation, not a platform preference.
AI-generated Shorts are earning roughly $0.03 to $0.10 per thousand views, so a million-view video nets $30 to $100 in ad revenue. US Shorts RPM sits at $0.328 versus $0.193 in Australia and $0.166 in the UK, and a 274-channel study found Shorts RPM running just 3 to 14 percent of long-form in nearly every niche — you need 11,000 to 34,000 Shorts views to match 1,000 long-form views. Finance and B2B top out at $0.10 to $0.15; entertainment and gaming bottom out at $0.01 to $0.04.
Released September 28, Vlo 0.3 is an open-source video editor with direct ComfyUI integration, prebuilt MiniMax H3 and Qwen workflows, and SAM2 masking built in. It targets the gap most AI video creators still fill with a round trip to Premiere or Resolve. Worth a look if your post-production step is currently the slowest part of your pipeline.
