ComfyUI shipped v0.26.0 on June 23, 2026, and it is one of the busiest model-node drops the project has had this year. In a single release, the team added native nodes for Kling V3-Turbo, Luma Rays 3.2, Krea2, Qwen3-VL, Boogu-Image, SCAIL-2, HappyHorse 1.1, and a new advanced 3D loader. Two stability patches, v0.26.1 and v0.26.2, followed within 48 hours. If you build image, video, 3D, or audio pipelines in ComfyUI, this is the release that decides which models you can reach without hunting for third-party custom nodes.
What Shipped in ComfyUI v0.26
ComfyUI is the open-source node graph that most serious local and hybrid creative-AI pipelines run on. The core repository bundles first-party and partner nodes so a workflow can call a cloud model and a local checkpoint side by side. v0.26.0 leans hard into that model: most of the headline additions are new model integrations rather than engine changes.
The release adds video nodes for Kling V3-Turbo and Luma Rays 3.2, image support for Krea2 and Boogu-Image, multi-reference conditioning through SCAIL-2, Alibaba's audio-native HappyHorse 1.1, and text generation through Qwen3-VL, which can also be wired in as a Flux2 Klein text encoder. On the 3D side, the new Load3DAdvanced node expands the loader options for spatial workflows, and the SoniloTextToMusic partner node had its price cut in half. That is more than ten distinct model touchpoints in one version, which is why this is a stack-shifting release and not a routine bump from v0.25.
The New Partner-Model Nodes, One by One
The fastest way to read this release is by what each node lets you generate. The table below maps the headline additions to the kind of work they unlock.
| Node | Type | What it adds | Runs |
|---|---|---|---|
| Kling V3-Turbo | Video | Faster Kling image-to-video generation | Cloud API |
| Luma Rays 3.2 | Video | Updated Luma Ray cinematic model | Cloud API |
| Krea2 | Image | Memory-optimized still generation | Local |
| Qwen3-VL | Text and vision | Text generation, plus Flux2 Klein text encoder | Local |
| Boogu-Image | Image edit | Editing with negative prompt support | Local |
| SCAIL-2 | Image | Multi-reference conditioning for consistency | Local |
| HappyHorse 1.1 | Video | Alibaba audio-native video | Cloud API |
| Load3DAdvanced | 3D | Advanced 3D model loader | Local |
The video additions are the headline. The Kling V3-Turbo node targets faster image-to-video drafts, which matters when you are iterating on shot composition and do not want to wait on a full-quality render each time. If you have been comparing Kling against the field, our breakdown of Kling 3 versus Runway and Sora covers where it lands on quality and cost. The Luma Rays 3.2 node updates the Luma partner integration for cinematic motion.
On the image side, Krea2 support landed with a memory-usage factor tuned specifically for the model, so it loads more predictably on consumer GPUs. Krea2 weights have been public since their open-weights release on Hugging Face, and now they have a first-party loader. Qwen3-VL is the sleeper feature: beyond standalone text generation, it can act as a Flux2 Klein text encoder, which gives Flux2 workflows a stronger language model for prompt understanding.
How to Update and Use the New Nodes
Getting v0.26 running is a short job, but the partner nodes have a setup step that local checkpoints do not. Here is the path that works cleanly.
1. Update to v0.26.2. In ComfyUI Manager, choose Update All, or run a git pull in your ComfyUI directory, or let Comfy Desktop auto-update. Confirm the version reads v0.26.2 so you get the stability patches on top of the v0.26.0 feature drop.
2. Restart and refresh. Restart the server and reload the browser tab so the new node definitions register. The new nodes appear under their model categories after the node list refreshes.
3. Add API keys for cloud nodes. Kling V3-Turbo, Luma Rays 3.2, and HappyHorse 1.1 are partner API nodes, so they call a hosted endpoint and bill per generation. Add your provider key in the node settings before running them. Local nodes like Krea2, Boogu-Image, and SCAIL-2 need the model weights downloaded instead.
4. Wire Qwen3-VL where it helps. Use it as a text-generation node for captioning and prompt expansion, or swap it in as the Flux2 Klein text encoder to sharpen prompt adherence in Flux2 graphs. The official ComfyUI documentation has the current node reference if a connection is not obvious.
5. Bring in 3D with Load3DAdvanced. The Load3DAdvanced node gives spatial and product-visualization pipelines a more capable loader, which pairs with the TripoSplat node category cleanups in the same release.
Why It Matters for Creators
The pattern that makes v0.26 important is consolidation. A year ago, reaching Kling, Luma, Krea, and a vision-language model from one graph meant installing a handful of community custom nodes, each with its own update cadence and breakage risk. Bundling these as first-party or partner nodes means they ship tested against the core, update with the release, and behave consistently. For anyone running production work, that is fewer broken mornings after an update.
It also keeps ComfyUI as the neutral hub. The same graph can draft a shot on a cloud video model, refine a frame on a local Krea2 pass, and run multi-reference conditioning through SCAIL-2 for character consistency, without leaving the canvas. HappyHorse 1.1 brings audio-native video into that loop, which extends the audio-native workflow we covered when the model first reached ComfyUI.
Frequently Asked Questions
Is ComfyUI v0.26 free?
Yes. ComfyUI is open source and free to install. The partner API nodes, such as Kling V3-Turbo and Luma Rays 3.2, call hosted models that bill per generation through the provider, so those specific nodes carry usage costs even though the software is free.
Do I need to reinstall ComfyUI to get v0.26?
No. Update in place through ComfyUI Manager with Update All, a git pull, or the Comfy Desktop auto-updater. Your existing workflows and models stay in place.
What is the difference between v0.26 and v0.25?
v0.25 added its own batch of bundled nodes earlier in June. v0.26.0 is a separate feature release that adds Kling V3-Turbo, Luma Rays 3.2, Krea2, Qwen3-VL, Boogu-Image, SCAIL-2, HappyHorse 1.1, and Load3DAdvanced. The two are different versions with different node sets, so treat v0.26 as new ground rather than a patch.
Are the new model nodes local or cloud?
Mixed. Krea2, Boogu-Image, SCAIL-2, and Qwen3-VL run locally with downloaded weights. Kling V3-Turbo, Luma Rays 3.2, and HappyHorse 1.1 are partner API nodes that run in the cloud and require a provider key.
Can Qwen3-VL really replace a Flux2 text encoder?
It can act as the Flux2 Klein text encoder, which routes prompt understanding through a stronger vision-language model. That can improve prompt adherence in Flux2 graphs, though results depend on your prompt style and the rest of the workflow.
Do the v0.26.1 and v0.26.2 patches change features?
No. v0.26.0 carried the new model nodes. The two follow-up patches are stability and fix releases on top, so updating straight to v0.26.2 gives you the full feature set with the latest fixes.
