OpenShot 4.0: Local AI Masking, Color Grading, and Screen Recording

For years, the honest answer to "what free editor should I use on Linux?" split three ways: Kdenlive if you wanted a serious NLE, DaVinci Resolve if you could tolerate Blackmagic's license terms and heavier install, or LosslessCut if you only needed fast trims without re-encoding. OpenShot was the friendly beginner option — easy to start, but missing pro color tools, built-in capture, and anything resembling local AI.

That changed on August 30, 2026. Creator Jonathan Thomas shipped OpenShot 4.0 with a headline trio: record screen and camera inside the app, grade footage with scopes and LUTs, and isolate subjects with ONNX models that run on your machine — YOLO, EfficientSAM, and Cutie — with no cloud upload and no subscription gate.

The release hit 319 points on Hacker News the same weekend, with threads comparing it to Resolve's free tier, Kdenlive's MLT stack, and whether a GPLv3 editor with local masking finally closes the gap for FOSS-first builders. This post is explainx.ai's read on what actually matters if you edit video locally, care about local AI without API keys, or want one desktop app instead of OBS plus an NLE plus a cloud mask service.

TL;DR — what people are asking

Why 4.0 is a different kind of OpenShot release

OpenShot 3.5 shipped in March 2026 with preview optimizations. Five months later, 4.0 is a major-version jump across libopenshot, the Qt UI, and creator workflows — not a polish pass.

Thomas framed it on the official blog as "some of the biggest creative workflow upgrades in our history" — record, grade, mask, and export in one free editor. The engineering subtext: the old web-based timeline is gone, replaced by a native Qt 6 timeline, with libopenshot 1.0.0 and libopenshot-audio 1.0.0 underneath.

For builders tracking the local ML stack, the Object Mask story is the headline: positive/negative click prompts, tracked masks across frames, masks reusable as sources for Blur, Pixelate, and Color Grade — all via downloadable ONNX weights, documented in the openshot-onnx repo, executed locally.

Color View: wheels, curves, LUTs, and scopes

Professional color work used to mean leaving OpenShot for Resolve or grading in ffmpeg scripts. 4.0 adds a dedicated Color View: large preview, clip properties, color wheels, and live scopes in one layout.

The Color Grade effect is the hub — exposure, contrast, temperature, tint, highlights, shadows, saturation, vibrance, plus separate wheels for global/shadows/midtones/highlights. Four editable RGB curves support Bézier handles. Industry-standard .cube LUTs import with blend intensity so you can dial a look instead of baking it at 100%.

Scopes ship in-editor:

• Luma waveform
• Histogram
• RGB parade
• Vectorscope (with skin-tone reference line)
• Audio Levels scope

Each scope has a Region tool — draw a box on a face or product and analyze only that area. Quick Color presets (Reset Color, Auto Contrast, Lift Shadows, Warm Up, Boost Color) and Look presets (Analog Tape, Sharpen, Blur, Shadow, Glow) all expand into the same keyframable Color Grade effect, so presets are starting points, not dead-end filters.

Why this matters for builders: if you produce tutorial or demo footage locally — the same workflow as video-use or product demo videos with Claude — consistent grade across clips no longer requires a second app.

Recording View: screen, webcam, mic, system audio — separate tracks

The Recording View is the other pillar. Capture microphone, desktop, webcam, and system audio inside OpenShot, each source saved as its own media file and timeline clip.

That separation is the design win:

• Trim a bad take on the mic without touching screen capture
• Reposition a picture-in-picture webcam (rounded PiP when screen + camera record together)
• Adjust voice-over volume independently of system audio
• Right-click a clip → Audio > Record for fast voice-over (playhead jumps to clip start, useful track selected below)

Live feedback includes mic meters, webcam preview, peak and RMS waveform rendering for recorded audio, and temporary timeline clips that swap for final media when recording stops. Audio > Record and multi-source capture support Linux Wayland via desktop portals (alongside X11 and PipeWire where available), Windows, and macOS — with platform-specific caveats (macOS system audio capture has documented limitations in release notes).

Recordings land in the project assets folder, keeping source files with the project — closer to how OpenMontage treats a repo as the unit of work, but for a traditional NLE project file. Compare to X's built-in video editor for social-only capture; OpenShot targets full project ownership on disk.

Ten new effects — and where local AI fits

OpenShot 4.0 adds 10 effects to libopenshot:

Benchmarks in the release post claim audio visualization modes exceed common 24/30/60 fps project rates — Bars and VU modes above 1,000 FPS in effect-only tests (not full export).

Object Mask is the builder-relevant piece. Add positive points on the subject, negative points on background, preview, then track through the clip. Masks feed other effects or invert for background-only edits — the same problem class as AI background blur guides and object removal tooling, but offline and GPLv3.

Object Detection also got a refresh: downloadable YOLOv5 ONNX models, segmentation masks, per-object visibility controls.

Audio-side improvements in 4.0 also include a denoiser, speech enhancement, and optional ComfyUI templates for heavier restoration chains (Reduce Noise, Enhance Voice, Improve Audio Resolution, Repair Audio).

Local AI masks via ONNX — no cloud, no subscription

Cloud video tools gate masking behind uploads and meters. OpenShot 4.0's pitch is the opposite: download models when needed, run on your hardware. From the official announcement:

This workflow runs locally. OpenShot uses free machine learning models that can be downloaded when you need them and run on your own computer… There is no cloud processing requirement, no AI account, and no AI subscription.

The openshot-onnx repository exports OpenCV-friendly ONNX packages:

Models download as zip release assets; runtime execution goes through libopenshot + OpenCV DNN — no PyTorch required after export. GPU backends (CUDA, OpenCL, auto-detection) depend on your OpenCV build.

That aligns with the broader 2026 pattern explainx.ai tracks — run inference where your data already lives (personal local AI systems, ONNX in the browser, closed vs open alternatives) — applied to pixels instead of tokens.

For privacy-sensitive footage — interviews, unreleased product shots, internal training video — local ONNX masks mean rushes stay on disk, unlike cloud background removal tools that re-encode uploads and often strip C2PA provenance manifests.

Honest limits:

• Performance is hardware-bound. A heavy Cutie pass on 4K footage will hurt on an ultrabook; desktop GPUs help.
• ComfyUI integration is separate and heavier. OpenShot 4.0 also exposes optional ComfyUI workflow templates (denoise, enhance voice, style transfer). User docs describe these as experimental and hardware-hungry — think discrete NVIDIA GPU and large RAM, not "runs everywhere ONNX masks run."
• Mask quality ≠ Resolve Magic Mask. Local click-to-track is a huge speedup over manual rotoscoping; it is not Hollywood compositing.

Native Qt timeline and the 61.8% Blur speedup

The timeline rewrite matters if you abandoned OpenShot over janky zoom or web-view glitches. Clips, transitions, keyframes, rulers, and menus are native Qt now — editable timecode in the ruler, smoother zoom, keyframe menus with interpolation controls, nested keyframes on Color Grade wheels/curves scale when retiming clips. Underneath, libopenshot 1.0.0 adds live capture readers, color analysis, local object masking, Qt 6 support, and stability fixes — the C++ engine OpenShot-Qt wraps.

Performance vs 3.5.1 (OpenShot's own benchmarks):

Timeline percentages look modest next to Blur, but the release notes argue timeline paths touch almost every interaction — preview, thumbnails, scroll, keyframes — so small gains compound across a session.

What Hacker News argued about (319 points)

The HN thread treated 4.0 as a legitimate FOSS competitor conversation, not a niche release. Recurring comparisons:

DaVinci Resolve (free tier) — Still the color and Fusion reference. Resolve's free version is not GPLv3 and ties you into Blackmagic's ecosystem, but it is the bar for scopes, nodes, and pro delivery. OpenShot 4.0 gives you scopes and wheels without an account; Resolve gives you depth.

Kdenlive — The Linux power-user NLE with years of MLT maturity, proxy editing, and effect chains. Kdenlive users on HN noted OpenShot finally has features Kdenlive had for years (scopes, recording integration), while OpenShot backers pointed at simpler onboarding and the new ONNX mask path without hunting plugins.

LosslessCut — Praised for instant lossless splits. OpenShot is not trying to replace that — different tool. Use LosslessCut to chop, OpenShot to finish with grade, PiP, and masks.

Local AI vs cloud editors — Commenters flagged ONNX masking as the anti-Runway/CapCut move: your footage stays on disk. That rhymes with how explainx.ai covered Diffusion Studio's edit-as-code bet and OpenCut's MCP surface — three different answers to "how should automation touch video," all local-first in different ways.

Agent-shaped editing — OpenShot remains a mouse-and-timeline NLE. It does not replace video-use (LLM + ffmpeg EDL) or agent pipelines like OpenMontage. It complements them: capture and grade in OpenShot, hand off rough cuts to an agent, or polish agent output with scopes and masks.

OpenShot 4.0 does not compete with MiniMax-H3 video editing benchmarks or Fable 5 agentic launch-video edits. It targets editors who want Resolve-like scopes and masks without leaving a free, offline app:

Who should actually install 4.0?

Getting started (August 2026)

OpenShot 4.0 is available from openshot.org/download for Linux (AppImage and distro packages), Windows, and macOS. Minimum libopenshot version is 1.0.0. Release artifacts: OpenShot 4.0.0 · libopenshot 1.0.0 · libopenshot-audio 1.0.0. ONNX models: OpenShot/openshot-onnx — build or download zip packages; review upstream licenses in MODELS.md.

Practical first steps after install:

1. Recording View — record screen + mic as separate clips; test PiP layout before a long take.
2. Look > Adjust Colors on a clip — open Color View, enable waveform + vectorscope, try a .cube LUT at partial intensity.
3. Object Mask — download the smallest ONNX model first; test on a short 1080p clip before committing to 4K.
4. Export presets — 4.0 refreshes social presets (TikTok, Instagram, LinkedIn, etc.) if you ship short-form from the same project.

What this means for what you build

If your stack is local-first AI, OpenShot 4.0 is a datapoint — not just for LLMs on laptop RAM, but for vision models on timeline footage without SaaS lock-in. Pair it with agent editors when you need conversational cuts; pair it with OpenCut when you need MCP automation into an NLE; pair it with cloud generators when you need footage that does not exist yet (ViMax, Runway comparisons).

The 4.0 release does not make Resolve obsolete. It does give FOSS users a credible single-app path from capture → grade → local mask → export — the workflow indie builders and educators asked for in that 319-point thread.

Related reading

• Diffusion Studio: video editing as code — agent-first timeline as JSX/TS
• video-use: edit video with Claude Code — conversational ffmpeg EDL editing
• OpenCut: plugins, rewrite, headless MCP — open NLE with automation surface
• OpenMontage: agentic video production — git-owned multi-stage pipeline
• Closed-source AI vs local open alternatives — the local-first decision frame
• How to blur video backgrounds with AI — cloud/browser blur vs local mask workflows
• Remove objects from video with AI tools — same isolation problem, different tools
• BGRemover.video — cloud vs local background removal — upload-based blur vs on-device masks
• AI video editing tools — complete guide — landscape map including local and cloud options
• Build a personal local AI system — hardware and software stack for offline inference

Sources: OpenShot 4.0 release announcement (Jonathan Thomas, August 30, 2026); OpenShot/openshot-qt release v4.0.0; libopenshot 1.0.0 release notes; OpenShot/openshot-onnx; 9to5Linux — OpenShot 4.0 overview; Linuxiac — Recording View feature summary; Hacker News discussion (~319 points, August 2026).

Version numbers, benchmark figures, and model availability reflect OpenShot 4.0 as of August 31, 2026. ONNX model performance varies by hardware; optional ComfyUI workflows have separate GPU requirements documented in the OpenShot 4.0 user guide.