Published September 23, the v0.2 distill runs text-to-image and reference-driven editing in 5 transformer passes instead of 40 with classifier-free guidance switched off entirely. It ships as a rank-256, alpha-256 bf16 LoRA at 1.3 GB that sits on top of the unmodified base transformer — sample with sigmas 1.0, 0.875, 0.75, 0.5, 0.25. Trained with DMD2-style distribution matching; community ComfyUI conversions are already on Hugging Face and a 4-step variant is live on Civitai.
Alibaba's 7B open-weight model shipped September 20 with a 64-channel RGBA autoencoder at 16x spatial compression, meaning alpha comes straight out of the model rather than a background-removal pass. It edits from up to 10 reference images and accepts local edits via circles, painted annotations, or masks. ComfyUI v0.37.0 added native support with three official templates plus MoGe 3 geometry estimation.
Alibaba PAI released the Union ControlNet on September 24 — eight structural controls plus inpainting collapsed into a single 7.5 GB branch with ComfyUI support out of the gate. That's one load instead of juggling separate depth, pose, canny and scribble models. Pair it with the community Fix LoRA, also out this week, which addresses the base model's washed-out colors and broken fine texture via a documented 20-step workflow.
The September leaderboard refresh puts Kuaishou's Kling v3 on top at 1,934, ahead of Alibaba's HappyHorse 1.0 at 1,816 and ByteDance's Seedance 2.0 Fast at 1,747. Scores come from roughly 1,394 blind votes across text-to-video, image-to-video and video editing. Veo 3.1 now holds at #3 on the with-audio board, and Runway Gen-4.5 — which led at launch in late 2025 with 1,247 Elo — has fallen out of the top ten entirely.
Alissonerdx's September 22 pack adds prompt-directed VFX Edit alongside documented Style Transfer and Head Swap adapters. Same day, Fun ControlNet Union 2.0 landed with Scribble, Layout and Gray control and double the injection blocks. H3 sits within a few Elo of the arena lead while outputting 2K at under a third of mainstream per-second pricing, and accepts up to 9 images, 3 video clips and 3 audio clips as references in a single request.
OpenAI shut the Sora app on April 26 and the API sunsets September 24, 2026. Reporting pegged the burn at roughly $1 million a day against about $2.1 million in lifetime subscription revenue. If you still have a Sora call in a production pipeline, it broke overnight — the practical replacements are Kling for price-to-quality, Veo 3.1 for native dialogue and ambient audio in one pass, and Runway when you need granular control.
Alibaba launched Wan 3.0 in public beta August 6 on Alibaba Cloud Model Studio and Qwen Cloud: 30-second clips in a single run, document-to-video input for the first time in the family, at 0.3 to 1.2 yuan per second depending on resolution. The Apache-2.0 "1.3B plus 14B" story circulating in the community appears in no Alibaba announcement — the open Wan line still stops at 2.2. What did ship open in July was WanSong, Wan-Dancer-14B and Wan-Streamer under Apache 2.0.
The September 15 core release introduced a generic looping mechanism built from Start Loop and End Loop nodes. Anything wired between them runs once per iteration, and you can iterate over a count, an index range, or a list. That's variant sweeps, seed grids and per-frame passes handled in the graph itself instead of via a third-party pack that breaks every update.
Meta is renaming the "AI creator" label to "AI-generated profile" to make disclosure clearer for viewers. Creators who don't properly label an AI-generated profile can see reach reduced; creators who do label won't be penalized for having a synthetic person as their subject. That follows the April 30 expansion of original-content protections from Reels to photos and carousels — accounts that mostly repost work they didn't create no longer get recommended to non-followers across Explore, Feed or Discover.
The full rewrite announced August 10 takes effect February 1, 2027, and builds on July's clarification that generic, repetitive, or template-driven mass-production content can't be monetized — nor can fake AI "experts." Properly disclosed AI content remains fully monetizable with no algorithmic penalty, and automatic AI detection with auto-applied disclosure labels has been live since May. The revenue ceiling is still real: faceless finance and investing channels report roughly $8K to $25K a month at 500K views once affiliates are layered in.
