OpenShot 4.0's New AI Masking Runs Local. Resolve Charges for That.

TL;DR

OpenShot 4.0 landed on August 30, and its headline feature is Object Mask: click a few points on a subject and the editor segments it and tracks it through the whole clip. The models doing the work are YOLO, EfficientSAM, and Cutie, converted to ONNX and executed on your own hardware. In OpenShot's words: no cloud processing requirement, no AI account, and no AI subscription. Around the AI, 4.0 is the biggest release the project has shipped in years: a native Qt timeline, a multi-source recording view, a color grading view with live scopes, ten new effects, and a Blur effect that finishes its benchmark workload 61.8% faster than in 3.5.1.

Rotoscoping without the rented brain

Isolating a moving subject has always been the miserable part of video editing: you either draw a mask frame by frame or you pay someone (or some cloud) to do it. Object Mask replaces that with a point-prompt workflow. You drop positive points on the thing you want and negative points on the things you do not, preview the mask, then process it through the clip.

The interesting part is the division of labor between three open models. YOLO handles object detection, EfficientSAM (a lightweight take on Meta's Segment Anything) turns your clicks into a precise single-frame mask, and Cutie, a video object segmentation model, propagates that mask across frames. Think of it as a three-person crew: a spotter who finds the actor, a sketch artist who traces one perfect outline, and a spotlight operator who then follows the actor through the scene without blinking. All three ship as ONNX conversions so libopenshot and OpenCV can run them locally, downloaded on demand rather than bundled.

The finished mask is not a dead end, either. It can be displayed directly or reused as the mask source for Blur, Pixelate, Color Grade, and other effects, and you can invert it to grade the background instead of the subject. That composability is what turns a segmentation demo into an actual editing tool.

The feature the paid editors gate

Context makes this release sharper. DaVinci Resolve has offered this kind of AI tracking mask for years as Magic Mask, but only in the paid Studio edition, and Adobe's equivalents ride a subscription. OpenShot just shipped the same category of feature in a GPLv3 editor where the download-a-model button is the entire onboarding flow. Nobody asks you to talk to sales.

OpenShot is careful about expectations: mask quality and speed depend on your hardware and which model you pick. But access does not depend on a paid cloud service, and for a lot of one-person creator operations that is the difference between using the feature and skipping it.

The rest of 4.0 pulls its weight

The AI angle is the hook, but the release notes read like three versions rolled into one:

• Recording View. Capture screen, webcam, microphone, and system audio directly into a project, with each source kept as a separate, editable track. On Linux that includes Wayland capture through PipeWire. Trim the mistake, duck the voice track, move the picture-in-picture window, all after the fact.
• Color View. Color wheels, curves, LUT controls, presets, and live scopes: vectorscope, histogram, and luma waveform. Paired with a new keyframable Color Grade effect.
• A fully native timeline. The old timeline was, until this release, literally a web page pretending to be a desktop widget. It is now native Qt, with smoother zooming, editable timecode, and better keyframe controls.
• Ten new effects. Audio Visualization, Beat Sync, Film Grain, Denoise, Shadow, Glow, Displacement Map, Timer, Color Grade, and Object Mask, bringing the total to 39.

Performance got real attention too, measured against 3.5.1 in the project's own benchmark workloads:

Why builders should care

Under the application, the engine crossed a milestone: libopenshot reaches 1.0.0, and so does libopenshot-audio. If you script video pipelines rather than click through them, the library that now bundles live capture readers, color analysis, and local object masking just declared its API stable, under GPLv3.

The bigger pattern is the delivery mechanism. Shipping segmentation and tracking as on-demand ONNX downloads, executed by the app's own engine, is a template any desktop tool can copy: no API key to rotate, no per-seat pricing, no footage upload for anyone with confidentiality rules to worry about. Local models keep eating workflows that used to justify a cloud bill.

Caveats

The performance numbers are OpenShot's own benchmarks, not independent tests, and the project itself says results vary with hardware, media, and project complexity. We have not benchmarked mask quality either: point-prompt segmentation is famously footage-dependent, and a busy background will still make Cutie's life hard. Models download on demand, so budget disk space and a first-run wait. Android support is planned, not shipped.

Key Takeaways

• OpenShot 4.0 shipped August 30 with Object Mask: point-prompt AI rotoscoping that runs YOLO, EfficientSAM, and Cutie locally via ONNX.
• There is no cloud processing, no account, and no subscription; the models are free downloads executed by libopenshot and OpenCV.
• Masks are composable: feed them to Blur, Pixelate, or Color Grade, or invert them to work on the background.
• The release also adds multi-source recording, a color grading view with live scopes, a native Qt timeline, and ten new effects, for 39 total.
• Per the project's own benchmarks, Blur is 61.8% faster and Sharpen 12.8% faster than in 3.5.1; libopenshot and libopenshot-audio both hit 1.0.0.
• A free GPLv3 editor now ships a feature class that DaVinci Resolve gates behind its paid Studio edition.

Sources: OpenShot 4.0 announcement, openshot-qt v4.0.0 release notes, libopenshot v1.0.0 release notes, UbuntuHandbook