ComfyUI news, October 2026: Qwen-Image 2.1, Comfy Agent and more

Early October was busy for anyone who makes a character's photos and reels in ComfyUI. A new open image model reached day-one support, Comfy's own assistant arrived, an open video model learned to make sound, and Civitai kept reshaping how models are listed. This briefing sorts the releases by what they change for a creator who posts the same person every day, and links every claim to where it was published.

The two weeks at a glance

Qwen-Image 2.1: ten references and a Turbo

Alibaba's Qwen team released Qwen-Image 2.1 on September 20 as a single model that both generates and edits. The visual generator has 7 billion parameters, small for a model at this level, paired with a Qwen3-VL 8B text encoder, according to the project's GitHub page. It natively works at 2K, with recommended portrait sizes of 1792 by 2400 and 1536 by 2752 for vertical posts. For a side-by-side with Krea 2 on portraits, references and licenses, see Krea 2 vs Qwen-Image 2.1.

Two features matter most for a recurring character. Editing accepts up to ten reference images, and the page lists identity preservation for people among its improvements, which is the job a face-reference workflow does today. And it outputs native transparency, so a cut-out of her for a thumbnail or a story overlay no longer needs a separate background-removal pass. The Turbo variant released on October 9 does the same work in eight steps, which turns a slow edit loop into something you can iterate on.

Musetta studioDemo data

In ComfyUI it needs no extra node pack: core support landed in version 0.37.0, and Comfy-Org publishes repackaged weights with example workflows. If an older Qwen workflow from Civitai still opens with red nodes, the guide to missing ComfyUI nodes walks through the fixes.

Comfy Agent: workflows from a chat

On October 1, Comfy Org introduced Comfy Agent. You describe the result you want, and it plans a workflow, places and connects the nodes on the canvas you already have open, runs the graph and iterates when something breaks. Coverage of the launch describes it as a beta that is available to everyone on Comfy Cloud, with a Comfy Desktop version promised within weeks, reusable skills and up to five parallel chats (Kombitz).

For creators this is the biggest change in how ComfyUI is used since the node graph itself: the graph stays visible and editable, but you no longer have to know which loader feeds which sampler. Until the desktop version ships, it works on Comfy's cloud machines rather than your own GPU. Disclosure: Musetta's own assistant, the Brain, does a similar job for setup on your local ComfyUI, so weigh our view accordingly; the comparison with plain ComfyUI lays out where each fits.

Kandinsky 6.0: open video that makes sound

Kandinsky Lab released Kandinsky 6.0 on October 6 under the MIT license, according to ComfyUI Wiki's release note. It comes in two sizes, a 29-billion-parameter Pro and a 3-billion-parameter Lite, each with a distilled 10-step checkpoint, and generates 5-second clips at 24 frames per second with synchronized audio: dialogue, ambience and music. A separate super-resolution model upscales to Full HD.

Sound has been the missing half of open reels. Until now a creator made the clip with an open model such as Wan 2.2 and added audio afterwards; see the Wan 2.2 reels guide for that workflow. Kandinsky 6.0 does both in one pass. Five seconds is short for a full reel, but it is the length of a hook, and the 3B Lite is the version most home GPUs can try first. Note that one write-up describes the variants differently, so check the model cards before you download 29B of weights. For the strongest open model with sound and the license catch that comes with it, see our MiniMax H3 review.

ComfyUI 0.39 and the smaller changes

ComfyUI 0.39.0, released October 5, is mostly plumbing, and two items affect daily work. A new --offline launch flag and a --disable-partner-nodes flag replace the old --disable-api-nodes, useful if you want a machine that never calls a paid API by accident. And the Luma Ray 2 partner nodes are deprecated, so a workflow that depends on them needs a replacement soon. The same release adds partner nodes for HeyGen Video 1.0 and Ideogram 4.5.

Civitai keeps reorganising

Civitai's late-September notes reorganised model types into fine-tunes, adapters, component replacements and workflow additives, capped hand-added resources in its generator at ten with at most three checkpoints, and opened audio LoRA training on ACE-Step in its Training Studio. The July change to its Creator Program, which lets eligible creators charge a license fee per generated image, is still being debated by smaller creators, as one creator's analysis shows. None of this changes how you run a downloaded model locally, but it changes where the trigger words and versions you rely on are listed.

What to do this week

If you would rather not track which version each workflow needs, that is the job Musetta was built for: it runs on your own ComfyUI, keeps her consistent across models and installs what a new workflow is missing. If you prefer a different front end altogether, the ComfyUI alternatives guide compares the options.