ComfyUI 0.43 came out on October 27 and, among roughly forty changes, it includes a node that changes the way you build a workflow: Branch, which picks by name which branch of the graph to follow. You'll find it in the official release notes on GitHub, together with a node for adding borders to images, ACEScg color for EXR files and new memory work.
Here's how we see it. We work with ComfyUI on a perfectly ordinary 12 GB RTX 3060 with 32 GB of RAM, and for us an update is interesting if it removes a manual step or an out-of-memory error, not if it adds a name to the model list. This version has plenty of entries dedicated to new models, but only a few change everyday work: we've lined them up here. We haven't tried 0.43 yet: we'll tell you what the official notes say and where we expect it to be useful.
Quick facts
| What it is | A ComfyUI update (version 0.43.0) with the new Branch node, the Pad Image node, ACEScg color in EXR files and new memory optimizations |
| Who made it | Comfy Org, with community contributions |
| Video memory required | Unchanged: it depends on the model you use. The memory changes are optimizations |
| Does it run on a 12 GB RTX 3060? | ComfyUI does, and the new nodes need no extra hardware; Dynamic VRAM is for Nvidia cards only |
| License | ComfyUI is open source; partner nodes (Kling, FLUX 3 Max and others) use paid services |
| Commercial use from Italy | Yes for ComfyUI; always check the license of every model you load |
| Where to download/try it | GitHub, version 0.43.0 or update from ComfyUI Manager |
| Free or paid | Free; only the API partner nodes are pay-per-use |
What does the Branch node do in ComfyUI 0.43?
The Branch node connects several inputs and passes through to its output only the one you choose by name. According to the pull request that introduces it, inputs are added as you need them and each one takes the name of its label; a selector decides which one to use. It lets you switch paths without rewiring: from one prompt to another, one model to another or one image to another.
The most interesting detail is "lazy" evaluation: only the chosen input is requested, so unselected branches don't run. In a workflow with two different paths, say a light upscale and a heavy one, you don't pay the time for the one you're not using. If you type a name that doesn't exist, the node returns an error. The PR discussion shows that it needs a recent interface and Nodes 2.0 mode: 0.43 already includes frontend package 1.57.4.
What is the Pad Image node for?
The Pad Image node adds a border to an image until it reaches the size the model requires. An example from the pull request: LTX models want dimensions that are multiples of 32, so you can enlarge the image, generate and then crop back to the original resolution.
- Three modes: explicit size, a multiplier, or "up to the multiple" you choose.
- Two styles: solid color (with a feathered edge) or mirror, which reflects the edge pixels and avoids hard cuts.
- No value clamping: values outside the normal range aren't cut off, so it works with HDR (high dynamic range) images too.
If you make videos with LTX or do outpainting (extending an image beyond its edges), it's a node that saves you from installing extra packs.
What changes for those working with EXR and color?
The Convert Image Color Space and Save EXR (Advanced) nodes now support ACEScg, a working color space used in film pipelines and compositing. If you bring your generated images into After Effects, Nuke or Blender, exporting in the right color space saves you manual conversions. The support is in the release notes (entry #17023): we don't know yet how it behaves in our projects.
Does anything change for video memory and speed?
Yes, in small ways and for some models. 0.43 brings Comfy-Aimdo 0.6.0 with the first changes to put KV caches (the memory where text models keep the context they've already read) under "Dynamic VRAM" too, plus improvements for Qwen 3.x models and a lower memory peak for MOSS-Image. Dynamic VRAM is the system ComfyUI uses to load model weights only when they're needed; according to the ComfyUI blog it's for now limited to Nvidia cards, on Windows and Linux.
For our work with 12 GB of VRAM, these are the kind of optimizations you notice when loading several models in a row. We don't have numbers on an RTX 3060: we'll update the article when we do. On AMD, meanwhile, 0.43 enables PyTorch attention in the VAE.
What's new for models?
MOSS-Image gets support for editing LoRAs, color depth control and image strength control in editing; we covered it in our article on MOSS-Image Edit in ComfyUI. Support for Fibo 2.0 also arrives, along with applying LoRAs to a single head for MiniMax H3. For an overview of the models, see our guide to AI image models in ComfyUI and the one on AI video models.
Try it now
- Update ComfyUI from the Manager or download version 0.43.0 and restart (with the portable version, make a copy of the folder first).
- Open a workflow with two variants, for example two prompts or two upscale models, and insert a Branch node instead of manual connections.
- Rename the input labels with clear names ("light", "heavy") and change the selector to choose the path.
- For an LTX workflow, try Pad Image in "up to the multiple" mode with 32 and mirror style, generate and crop.
Other news
- Partner nodes: the same notes list nodes for Kling 4.0, FLUX 3 Max, Bria FIBO 2.0, PixVerse V6-Max and Ideogram Vector (SVG). These are paid services and don't run locally. Source: release notes.
- Voices and sounds: HeyGen Voice (voice cloning and speech synthesis) and Mirelo SFX sound effects have been added, again as partner nodes. Source: release notes.
- Interface: the frontend moves to 1.57.4 and an optional local agent loaded via comfy-agent arrives. Source: release notes.
Frequently asked questions
Does ComfyUI 0.43 work on a 12 GB RTX 3060?
ComfyUI does, just like previous versions. The Branch and Pad Image nodes don't need any extra hardware. What you get from the memory optimizations depends on the model: we don't have measurements on our card yet.
Is the Branch node like a Switch?
It's similar, but it chooses by name and, according to the PR, doesn't run unselected branches. Other node packs have similar functions; here it's in the ComfyUI core, so you don't need to install anything.
Should I update right away?
If you use LTX, EXR or MOSS-Image, it's worth it. Otherwise you can wait a few days to see whether fixes come out, like 0.42.1 after 0.42. Always make a copy of the folder before updating.
Where can I find the full list of changes?
In the 0.43.0 release notes on GitHub, where each entry links to its pull request.
If you're new to nodes, start with our ComfyUI beginner's guide; if you're choosing a computer, take a look at our graphics cards for AI. You'll find everything else on the ComfyUI page, and on the YouTube channel @tutorial3d we test the free workflows.
