OpenShot 4.0 Runs Subject Masking on Your Laptop With No AI Account. The Model Licenses Are the Part Nobody Read.
The free GPLv3 editor now cuts a moving subject out of a shot using models that download on demand and run on your own machine. Two of the three are Apache and MIT. The third is Ultralytics AGPL-3.0, and OpenShot's own export repo is the only place that says so.
The release post for OpenShot 4.0 contains a sentence that most video software would never print: "There is no cloud processing requirement, no AI account, and no AI subscription."
That is a real claim about a real feature. Click a few points on a person walking through a shot, and OpenShot generates a mask that follows them for the length of the clip. The models download when you ask for them. They run through OpenCV inside libopenshot. Your footage does not leave the machine. For a free GPLv3 editor that has spent eighteen years being the thing people recommend when someone cannot afford Premiere, this is the largest capability jump in the project's history, and it landed on August 30.
Here is what the post does not tell you, and what almost every write-up of the release has skipped. The masking feature is not one model. It is three, from three different research groups, under three different licenses, and one of those licenses is the kind that makes lawyers sit up. OpenShot knows this. They wrote it down. They wrote it down in a separate repository that nobody clicked.
What the feature actually does
Add the Object Mask effect to a clip. Click positive points on the subject and negative points on the background around it. OpenShot builds a selection preview so you can check the boundary before committing. Then it processes that mask forward through the clip. When the subject deforms, turns, or disappears behind something, you add more prompts on those frames and it recovers.
The mask is not baked into a render. It becomes a reusable source that Blur, Pixelate, Color Grade and the rest of OpenShot's effects can point at, and it can be inverted so you work on the background instead of the subject. One roto pass, several downstream uses. That composability is the part that makes it feel like a real tool rather than a demo, because roto is expensive and you only want to pay for it once.
The rest of OpenShot 4.0 is not AI at all and deserves a mention because it changes who this editor is for. There is a full Color view with four wheels covering global adjustments, shadows, midtones and highlights, four editable curves with Bézier handles, .cube LUT support with an intensity slider so you can blend a LUT in at partial strength, and Luma Waveform, Histogram, RGB Parade and Vectorscope scopes with a region tool that lets you draw a box over a face and read only that part of the frame. There is screen, webcam, microphone and system audio recording, with each source landing as its own separate clip so you can fix the voice without touching the screen capture. Ten new effects, including audio-reactive visualizations and a displacement map.
OpenShot's own benchmark says the Blur effect completed its test workload 61.8% faster than in 3.5.1, Sharpen improved 12.8%, and timeline rendering improved 3.4%, rising to 5.1% with transforms. Those are the project's own numbers on the project's own benchmark suite, which is the honest way to read any vendor performance chart.
The three models, and which one has teeth
OpenShot maintains a second repository, OpenShot/openshot-onnx, whose entire job is to convert upstream research checkpoints into ONNX files that OpenCV can execute without Python or PyTorch installed. That repository is where the actual answer lives.
Three model families, three jobs:
EfficientSAM turns your clicked points into a seed mask on a single frame. It is the "what did you mean by that click" model. Licensed Apache 2.0 by its authors at yformer/EfficientSAM. Clean.
Cutie takes that seed mask and propagates it forward through the video, which is the hard part and the reason the feature works at all on moving subjects. Licensed MIT, copyright Ho Kei Cheng, 2024. Also clean. OpenShot's exporter produces four ONNX slices per Cutie release package, which is what libopenshot's Cutie Object Mask backend expects.
YOLO handles the separate Object Detection effect, which also got a substantial update in 4.0 with downloadable models, model validation, segmentation masks and better control over detected objects. YOLO is where it gets complicated, because YOLO here means Ultralytics.
Ultralytics ships its models under AGPL-3.0 by default, with a paid Enterprise License as the alternative. Their own licensing FAQ is unusually direct about it: "All Ultralytics YOLO trained models fall under the AGPL-3.0 License by default. The AGPL-3.0 License covers the training code and the models produced by that training code." Their page lists the cases that require the Enterprise License, and the list includes internal business tools, any commercial product or service, proprietary software, SaaS platforms, and R&D projects that are not fully open-sourced. They state that this applies even if you train from scratch and use no pretrained weights.
OpenShot did not hide this. MODELS.md in the export repo says the ONNX files are not committed to Git in part because "pretrained YOLO checkpoints and derived ONNX exports are licensed by Ultralytics," and instructs anyone republishing them to review Ultralytics licensing first. NOTICE.md spells out the split explicitly: the repository code and documentation are MIT, and "Ultralytics YOLO source code and pretrained model weights are provided and licensed by Ultralytics." That is more disclosure than most projects in this space bother with.
Where I land on this, and where I do not
I want to be careful here, because the loud version of this story would be wrong and would scare people off a good free tool for no reason.
Say you are a video editor cutting a wedding film, a YouTube video, or a client's commercial, and you use Object Mask to roto a subject and blur the background. My reading is that you are not distributing Ultralytics weights, not serving them over a network, and not shipping software. You are producing pixels. The framing I am leaning on is OpenShot's own NOTICE, which tells you to review upstream licensing "before redistributing weights, derived ONNX exports, or using them in commercial products," and rendering a video is not redistributing an ONNX file.
Be clear about what that paragraph is, though. Ultralytics has published nothing about rendered-output-only use one way or the other. Their licensing page addresses code, models, training pipelines and products, not the pixels a downstream application emits. So the reading above is mine, not theirs, and I am not a lawyer. If your work depends on the answer, get one.
Where it does matter, and where I would actually stop and read: if you are building something on top of OpenShot or libopenshot, bundling those ONNX files into a product, running them as part of a service you sell, or shipping them inside an appliance. Then Ultralytics' position is that you need an Enterprise License, and Ultralytics' position on AGPL scope is aggressive by the standards of how most people read AGPL. That aggressiveness is not universally accepted as a matter of copyright law, but you do not want to discover the disagreement from the receiving end.
The practical upshot is a nice one. The masking pipeline you will actually use every day, EfficientSAM plus Cutie, is Apache plus MIT. That half is as unencumbered as open source gets. The Ultralytics dependency sits in the detection effect next door, and if you never touch Object Detection you never touch AGPL-licensed weights at all.
One documentation mismatch worth knowing before you go looking
The OpenShot 4.0 post says: "OpenShot now supports downloadable YOLOv5 ONNX models."
The supported outputs table in openshot-onnx's own MODELS.md lists four families, and YOLOv5 is not among them. What it lists is YOLO26, YOLO11, YOLOv8 and YOLOE-26, sourced from ultralytics/assets releases v8.4.0 and v8.3.0, each in n/s/m size tiers as -seg segmentation exports.
I cannot tell you which document is stale. It is possible the editor accepts a v5 model that the export utility no longer builds, or that the blog copy simply says v5 as shorthand for YOLO. Either way, if you go hunting for yolov5n-seg.onnx because the release post told you it exists, check the export repo's catalog first. yolo/models.json is described as the consumer-facing release catalog with the friendly dropdown names and zip filenames, so that is the file that reflects what actually ships.
Put this into practice
The smallest real thing you can do in the next ten minutes is roto one subject out of one shot and blur what is behind them.
- Download OpenShot 4.0 from openshot.org/download. Windows, macOS and Linux. Free, GPLv3, no account.
- Drop a clip on the timeline where a person or object moves against a background that is not a matching color. Give the first attempt an easy shot. A subject who crosses in front of something the same brightness will make you think the tool is worse than it is.
- Right-click the clip and add the Object Mask effect. Click positive points on the subject. Two or three usually gets you started. Then click negative points on the background immediately around them, especially anywhere the boundary is ambiguous.
- Look at the selection preview before you process. This is the step people skip and then blame the model. The preview is showing you EfficientSAM's answer to your clicks on one frame. If it is already wrong there, adding frames will not fix it. Move the points.
- Process the mask through the clip. Cutie carries it forward. Scrub through and find the frames where it slips, which will be wherever the subject changes silhouette or something passes in front of them.
- Add more prompts on the bad frames. This is the actual loop and it is where the time goes. Budget for it.
- Point an effect at the finished mask. Add Blur, set the mask as its source, and invert if you want the background blurred rather than the subject. Then try Color Grade the same way so you can push the subject warmer than the room.
The first-run download happens on demand. OpenShot pulls the models it needs from the openshot-onnx releases rather than shipping them inside the installer, which is why the installer is not enormous and also why your first mask will pause while something downloads.
If you want to know exactly what you are pulling before you pull it, the file layout is published: YOLO release packages are zips containing model.onnx and a classes.names label file, Cutie packages contain four ONNX slices, and EfficientSAM packages carry their own checksums and sizes in efficient-sam/models.json.
Honest limitations
OpenShot publishes no performance figures for masking. The blog says plainly that "performance depends on your computer and the selected model," and there is no VRAM number, no timing table, no minimum spec. The audio visualization effects got a full benchmark chart with frames-per-second figures across ten modes. Masking got a sentence. Read that gap for what it is: they measured the thing that was easy to measure.
The prompting loop is manual and it does not scale to long takes. Every frame where the subject deforms is a frame where you might be clicking. On a two-second insert this is delightful. On a three-minute walk-and-talk with occlusions, you are going to find out why roto artists get paid.
The models are not in the repository. MODELS.md states that generated model binaries are ignored by Git and must be built locally or downloaded from GitHub Releases. That is a reasonable engineering decision and it also means the exact artifacts you receive depend on what has been attached to a release, not on what is in the source tree.
XMem, which appears in the repo, is not supported. The README says it "remains in experiments/ as historical scratch work and is not promoted as a supported release family." If you find it and try to use it, that is on you.
Object Detection and Object Mask are different features with different licensing profiles, and the release post presents them side by side as one AI story. They are not one story. Know which one you are using.
What I would actually watch next
The interesting question this release raises is not whether OpenShot can compete with Resolve. It cannot, and it is not trying to.
The interesting question is what happens to the "AI feature means subscription" assumption when a volunteer-funded editor ships prompted segmentation that runs on a laptop with no account, and the only thing standing between that and total freedom is one upstream model family's licensing posture. EfficientSAM is Apache. Cutie is MIT. If someone exports a permissively licensed segmentation model into the same ONNX slot that Ultralytics currently occupies, the AGPL question disappears entirely and the whole pipeline becomes something you can build a business on without reading a single license page.
That swap is a weekend's work for whoever wants to do it. The export scripts are MIT and published. python yolo/scripts/export_yolo_seg_onnx.py is right there in the README, and nothing about the architecture requires the model behind it to be Ultralytics'.
If you try the masking on real footage, I want to know where it broke. Specifically: how many prompt frames did a ten-second shot cost you? That number is the whole story of whether this replaces a paid tool or just makes a free one more fun.
Medium metadata
Suggested kicker image: the OpenShot 4.0 Object Mask screenshot showing positive and negative point prompts over a subject with the generated mask preview beside it.
Sources: