Install

Grab the latest build from GitHub Releases:

• Windows installer — qlipq-setup-x64.exe runs a short setup wizard and adds a Start-menu shortcut and an uninstaller. Recommended.
• Windows portable — qlipq-windows-x64.zip: no install, just unzip and run qlipq.exe.
• Linux — qlipq-linux-x64.tar.gz: extract and run qlipq (loads its bundled FFmpeg libraries from ./lib).

Every build bundles the FFmpeg libraries QlipQ needs — there’s no separate ffmpeg to install.

1. Add watched folders

In Settings → Watched folders, add the folder(s) where your recordings land (for OBS this is your recording or replay-buffer output path). QlipQ can auto-detect the OBS and NVIDIA Share output folders and offer them as one-click presets. It scans these folders — including subfolders — on launch and watches for new files while it runs.

Tick Hide auto-captured highlights to keep NVIDIA App’s automatic Highlights out of the queue (they’re tagged NVIDIA APP (Highlights) in the file, versus NVIDIA APP for a manual recording; a clip’s encoder is shown in the editor’s debug panel). Highlights that aren’t hidden are marked with a Highlight badge in the queue. This sets the default; the queue’s filter bar can override it per session (show all clips, only highlights, or hide them).

2. Choose an output folder and naming template

Set an Output folder for exports. The naming template controls how exported (and renamed) files are named. Available tokens: {date}, {time}, {datetime}, {source}, {name}, {index}.

3. Pick your output quality

Settings → Output defaults controls export quality and is applied to every export:

• Quality — a named preset, a custom CRF, VBR (CRF capped by a max bitrate), or a target bitrate.
• Frame rate, resolution (down to 720p / up to 4K), codec (H.264 / H.265), container (mp4 / mkv), and audio bitrate.

The editor shows an estimated file size for the current clip, and you can override the quality per clip.

4. Edit and export

Use Sort above the clip list to order recordings by newest/oldest, name A–Z/Z–A, shortest/longest, or largest/smallest. Newest first is the default and uses the recording date when available, otherwise the file’s modified date. Sorting works with the current filters and does not change the selected clip or its edits. The sort choice lasts for the current session; clips whose duration or size is still loading appear after clips with known values.

Each clip has a thumbnail that loads when its card comes into view. Hover over a clip card briefly to play a muted preview in its thumbnail; leaving the card restores the still image. Only one hover preview plays at a time, and it loops through the original recording without moving the editor’s playhead or changing the selected clip. Clicking the thumbnail opens the clip in the editor as usual.

1. Pick a clip from the Queue (each shows its date, length, and size). A filter bar is pinned to the top of the queue: search by filename and narrow the list by status, game (the {source} label), tag, or highlights (all clips / only highlights / hide them). When any filter is active it shows how many clips match and offers Clear filters to reset. By default a clip starts playing as soon as it’s selected — turn this off with Settings → Playback → Play clips automatically when selected to open clips paused.
2. Set the in/out points on the timeline — the scrubber highlights the in/out window and marks each endpoint, and playback loops within it (pressing play jumps to the in-point and loops back when it reaches the out-point). Each point has an editable timestamp with ±0.5 / ±1 / ±5 s nudge buttons, or press Set (or the I/O keys) to capture the current playhead. Type a timestamp in the playhead field to jump, drag the scrubber (playback keeps going if it was already playing), or use the −60/−5/−1 / +1/+5/+60 second jump buttons. Timestamps read as frame-accurate timecode — h:mm:ss.ff, where ff is the frame within the second — so a single ←/→ step changes the last digits by one.
3. Keyboard shortcuts default to Adobe Premiere Pro — Space play/pause, I/O set in/out, ←/→ step a frame, Shift+←/→ jump 5 s, Home/End go to start/end, Ctrl+M export — and are rebindable in Settings → Editor shortcuts. With a clip selected, Delete removes its file from disk after a confirmation (Enter confirms the prompt) and Shift+Delete removes it immediately without asking.
4. Optionally enable crop and adjust the rectangle.
5. Toggle audio tracks and set their levels (your selection carries to the next clip); changes are reflected in the preview as you make them. On export, the enabled tracks are mixed together into one track at the levels you set.
6. Click Export clip. If a file with the same name already exists you can overwrite it or append a timestamp to keep both, and the After export setting decides what happens to the original (keep, delete, move, rename, or prompt). Use Show file to reveal the exported clip.

Preview vs. export. The preview decodes frames in-process and tonemaps HDR sources to SDR for display — it’s a visual guide, and exports always use the original file, untouched. Its sharpness is set by Settings → Preview quality (720p / 1080p / 1440p / Source; default 1080p) — higher is sharper but costs more decode/GPU work, so lower it if playback stutters. If an HDR clip (especially a Windows HDR desktop recording) previews too dark, Settings → HDR preview → Brightness lifts it with an adjustable gamma (higher = brighter; 1.0 = off; Reset restores the default). Both affect the preview only.

Diagnosing preview stutter. If the preview stutters or the audio drops out, enable Settings → Playback → Show debug panel in the editor. The panel adds a Debug card under the editor showing the clip’s details, whether it’s decoding on the GPU (hardware) or in software, and — while playing — live video/audio buffer levels, dropped frames, and audio underruns. Software decoding of heavy 1440p/4K AV1/HEVC is the usual cause; it never affects exports.

Suggest a highlight trim

Highlight suggestions are optional and run through a local Ollama vision model. Install Ollama, start it, and download the default model:

ollama pull qwen3-vl:4b-instruct

Qwen3-VL requires Ollama 0.12.7 or newer. You can choose another installed local vision model in Settings → Highlight suggestions. QlipQ connects to Ollama at 127.0.0.1:11434; it does not download models automatically and rejects cloud models.

Use the explicit -instruct variant. Ollama’s qwen3-vl:4b tag selects the thinking variant, which can spend the entire output budget reasoning without returning a highlight answer. If you previously configured qwen3-vl:4b, download qwen3-vl:4b-instruct and select it in Settings → Highlight suggestions. Existing model choices are preserved.

1. Open a clip and find Highlight suggestion below the In/Out controls.
2. Describe what to find, such as “consecutive headshots”, “a multi-kill”, or “a successful boss fight”, then click Suggest highlight. Playback pauses while analysis starts. Longer recordings are analyzed in overlapping sections; the section count shows progress. Cancel or selecting another clip stops analysis.
3. Review the suggested range and its description. Your current trim stays unchanged until you click Apply trim. Dismiss keeps your existing edit.
4. Applying the suggestion adds up to 3 seconds of lead-in and 2 seconds of aftermath, bounded by the original clip. A single-instant event, such as a kill notification, gets the same padding. Adjust the normal In/Out controls, preview the result, and export when ready. The applied trim is saved like any other edit.

QlipQ asks the model to compare complete plays and consider execution quality: consecutive aimed headshots should rank above routine or assisted multi-kills. It also asks the model to distinguish new eliminations from persistent kill-streak banners and repeated kill-feed entries. Suggestions aim to include the action sequence, rather than just the notification that follows it. These judgments still depend on the model; it can misread the HUD, misdescribe a play, or choose the wrong moment.

This is an experimental suggestion: analysis samples video at roughly one frame per second, at up to 1280 × 720 to retain HUD detail, in 30-second sections with a 5-second overlap. QlipQ requests a 40K-token context to fit the sampled frames and allow up to 4,096 output tokens, which needs more memory than a short chat. Some models generate thinking text even when asked not to; that uses part of the output budget and can make analysis take longer. QlipQ only uses the final answer. It does not analyze audio and may miss brief action, jokes, or events that require game-specific context. It may report no clear highlight. Accuracy and speed depend on the model, your hardware, and the recording; review every suggested trim.

If Ollama is unavailable or the model is missing, the editor shows an error with setup instructions. An output limit error means the model ran out of tokens before finishing its answer; try a shorter clip or a local non-thinking vision model. An empty answer error means Ollama finished without returning an answer; retry or choose another local vision model. These errors do not mean that no highlight was found. A failed or cancelled analysis leaves your current trim unchanged.

Next: set up the OBS replay buffer.