# ComfyUI-SAM3: text-prompted segmentation inside ComfyUI

> ComfyUI-SAM3 is a ComfyUI custom node package that wraps Meta's SAM3 model for open-vocabulary image and video segmentation driven by natural language prompts, with box and point inputs, a live test gallery and a one-click installer built on the experimental comfy-env package. It suits ComfyUI users who want promptable segmentation in a node graph, and it installs through a path the project itself flags as experimental.

**PozzettiAndrea/ComfyUI-SAM3** — ComfyUI wrapper for segment anything 3

- Repository: https://github.com/PozzettiAndrea/ComfyUI-SAM3
- Stars: 579 · Forks: 84
- Language: Python
- License: NOASSERTION
- Published: 2026-10-22 · Updated: 2026-10-22 · Language: en
- Canonical page: https://hysenlabs.com/projects/pozzettiandrea-comfyui-sam3

## Open-vocabulary segmentation as ComfyUI nodes

ComfyUI-SAM3 is described as a ComfyUI integration for Meta's SAM3, the Segment Anything Model 3, providing open-vocabulary image and video segmentation using natural language text prompts.

The distinction between open-vocabulary and class-based segmentation is the whole point of the wrapper. A class-based segmenter knows a fixed list of things it can find, and anything outside that list is invisible to it. An open-vocabulary segmenter takes a phrase and finds the region that phrase describes, so the vocabulary is not fixed in advance. For a ComfyUI user, that means a node that takes text as an input alongside images and video, which is a fundamentally different shape from a fixed-label workflow.

The package metadata names it as a ComfyUI custom node with the publisher id pznodes and the display name SAM3, at version 0.1.21, and its own description repeats the warning about the experimental comfy-env package and the pixi package manager.

The credits section names the two upstream pieces. SAM3 comes from Meta AI Research, and the ComfyUI integration is the project itself. The third credit matters for anyone auditing the code: the interactive points editor is adapted from ComfyUI-KJNodes by kijai under the Apache 2.0 licence, and the SAM3PointsEditor node is based on the PointsEditor implementation from KJNodes, simplified for SAM3-specific point-based segmentation.

That last credit is the one to read carefully. A simplified fork of a points editor is where the interaction bugs live, and the licence obligations of Apache 2.0 flow into this repository.

## Three install paths, ordered by speed and reliability

The installation section gives three options and ranks them by speed against reliability, which is an unusual and useful thing to state plainly.

The first is ComfyUI Manager, marked as recommended: search for SAM3 in the Manager and click Install from the highest version displayed, and if that does not work, try nightly. The second is Manager via Git URL, using the Install via Git URL option with the repository URL. The third is the manual route, described as the most reliable.

The manual route is four commands:

```bash
   cd ComfyUI/custom_nodes
   git clone https://github.com/PozzettiAndrea/ComfyUI-SAM3.git
   cd ComfyUI-SAM3
   pip install -r requirements.txt --upgrade
   python install.py
```

So the node package goes into the custom_nodes directory of an existing ComfyUI installation, the Python requirements are installed, and install.py does whatever model setup is needed. The order of the list is worth reading as advice: the Manager path is fastest because it automates the same steps, and the manual path is most reliable because you can see what each step did.

One detail to notice in the Manager instructions. You are told to install from the highest version displayed, and to try nightly if that does not work, which implies the version list is not in a stable order or that some entries are pre-release builds, and that picking the newest is not automatically the safest choice for a node that also has to fetch model weights.

The project also asks for feedback, pointing at a Discussions page and an Issues page, and asks for problems to be reported during installation or use of the nodes.

## The experimental comfy-env and pixi path

The warning at the very top of the README is the first thing a reader sees, and it is not boilerplate. The warning says the project uses the experimental package comfy-env to attempt a one click isolated install, and that it will download and use the pixi package manager.

The words to weigh are attempt and experimental. The one click install is not a supported path, it is an attempt at one, and the mechanism is a package that the project itself labels experimental.

What the mechanism does is create an isolated environment for the node rather than installing into the ComfyUI Python environment directly. That is a sensible goal, because a segmentation model with its own pinned dependency set is exactly the kind of thing that breaks a shared environment when it is installed carelessly. The pixi package manager is what makes the isolation possible, and pixi is downloaded as part of the process.

The requirements file shows the pinned versions:

```bash
comfy-env==0.3.89
comfy-3d-viewers==0.2.44
```

Two things follow from that short list. The comfy-env package is pinned exactly, which is what you want for a tool that manages its own environment, and there is a second dependency on a 3D viewers package that a segment-anything wrapper would not obviously need.

The repository layout is consistent with a package that ships more than nodes. There is a prestartup_script.py, which is ComfyUI's mechanism for running code before the server starts, an install.py, a nodes/ directory, a web/ directory for the frontend side, workflows/ with example graphs, an assets/ folder, a docs/ folder, and two pixi-related manifests, comfy-env-root.toml and comfy-test.toml, the latter presumably for running the project's own tests in an isolated environment.

## Box, point and text inputs, and the live test gallery

The package description gives the three input types the model works with: image and video segmentation with text prompts, boxes, and points.

That list maps onto the node set. The text prompt is what makes the segmentation open-vocabulary, the box is a coarse region constraint, and the point is the interactive refinement. The SAM3PointsEditor node is the interactive one, and it is the node derived from KJNodes, which tells you it is meant to be driven by a human placing points rather than by a fixed pipeline stage.

The repository also links a live test gallery hosted on GitHub Pages, at pozzettiandrea.github.io/ComfyUI-SAM3/, with a gallery preview image and a link labelled View Live Test Gallery. A live gallery of workflow examples is more useful than static screenshots for a node package, because the thing a user needs to know is how the nodes are wired together, and a page of runnable graphs answers that better than a folder of PNGs.

The workflow examples live in the repository as well, in the workflows/ directory, so the gallery is a presentation layer over content that is versioned with the code.

The project also links a short video attachment, which appears to be a demonstration of the segmentation in action, and the interaction between the video and the gallery suggests the author is trying to show both the behaviour and the wiring.

## Where ComfyUI-SAM3 is the wrong tool

The first case is a workflow that does not need ComfyUI at all. If you want to segment a video with a text prompt once, a standalone segmentation tool is a smaller thing to install and a smaller thing to debug. The wrapper earns its keep when the segmentation is one stage in a larger graph, where the model output has to be consumed by other nodes, and that is exactly the situation it is built for.

The second limitation is the install path itself. A package that downloads a package manager, creates an isolated environment and pulls model weights is a heavier operation than a custom node that pip installs a library, and the project is upfront that the mechanism is experimental. For a machine where ComfyUI runs in a locked-down container, on a platform where the extra download is not welcome, or for an organisation that reviews every dependency that enters the environment, the manual path is the only defensible one, and even that runs install.py.

The third is the dependency footprint. The requirements file is two lines, but one of them is a 3D viewers package, which suggests the node set reaches beyond segmentation into previewing volumes. That is a reasonable thing to want, and it is also a wider install than the description implies, so anyone who only needs masks should check what the 3D viewers package pulls in before committing to it.

The alternative worth comparing is Meta's own SAM3 implementation, used directly. The difference in approach is that the upstream repository is the model and the training code, while this is an integration layer that adapts a trained model to a node graph. If you want to fine-tune, to run it as a service, or to control the inference loop directly, the upstream code is the better starting point. If you want segmentation as a stage in a ComfyUI graph next to your other models, this wrapper saves you the plumbing, and the credits make the provenance of the parts clear enough to audit.

## Conclusion

ComfyUI-SAM3 fits a ComfyUI user who already has a working ComfyUI installation and wants promptable segmentation as nodes in a graph, since the manual install is four commands and the wrapper handles the model plumbing. It is the wrong choice if you want a standalone segmentation tool, if you cannot run a GPU workload, or if your environment policy rules out a package that downloads and drives the pixi package manager to create an isolated install, which the project warns about at the top of its README. Before you install it, read that warning, decide between the Manager path and the manual path, and check the requirements file, which pins comfy-env at 0.3.89 and pulls in a separate 3D viewers package, so the dependency footprint is larger than a segment-anything node might suggest.

## FAQ

### What does ComfyUI-SAM3 do?

It is a ComfyUI custom node package wrapping Meta's SAM3 model for open-vocabulary image and video segmentation using natural language text prompts, with box and point inputs alongside the text, so the vocabulary is not limited to a fixed class list.

### How do I install ComfyUI-SAM3?

The three documented options are ComfyUI Manager by searching for SAM3 and installing the highest version, or trying nightly, the Manager Install via Git URL option, or the manual route: clone into ComfyUI/custom_nodes, run pip install -r requirements.txt --upgrade, then run python install.py, which the README calls the most reliable option.

### What is the warning about comfy-env in ComfyUI-SAM3?

The README opens with a warning that the project uses the experimental package comfy-env to attempt a one click isolated install, and that it will download and use the pixi package manager. The one click path is therefore an attempt rather than a supported route.

### Where does the ComfyUI-SAM3 points editor come from?

The interactive points editor is adapted from ComfyUI-KJNodes by kijai under the Apache 2.0 licence, and the SAM3PointsEditor node is based on the PointsEditor implementation from KJNodes, simplified for SAM3-specific point-based segmentation.

### What licence is ComfyUI-SAM3 released under?

The licence identifier is recorded as a custom one GitHub cannot classify, with the terms in the LICENSE file at the repository root. The package version is 0.1.21, the last push was on 2026-08-24, and the project has no GitHub releases.

## Sources

- [Issues](https://github.com/PozzettiAndrea/ComfyUI-SAM3/issues)
- [PozzettiAndrea/ComfyUI-SAM3 on GitHub](https://github.com/PozzettiAndrea/ComfyUI-SAM3)
- [README](https://github.com/PozzettiAndrea/ComfyUI-SAM3/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/pozzettiandrea-comfyui-sam3