0.18.0 Beta 2

[blakeblackshear]

Images

• ghcr.io/blakeblackshear/frigate:0.18.0-beta2
• ghcr.io/blakeblackshear/frigate:0.18.0-beta2-standard-arm64
• ghcr.io/blakeblackshear/frigate:0.18.0-beta2-tensorrt
• ghcr.io/blakeblackshear/frigate:0.18.0-beta2-rk
• ghcr.io/blakeblackshear/frigate:0.18.0-beta2-rocm
• ghcr.io/blakeblackshear/frigate:0.18.0-beta2-tensorrt-jp6
• ghcr.io/blakeblackshear/frigate:0.18.0-beta2-synaptics

Changes since Beta 1

Security

• GHSA-q8jx-q884-jcq9: bound REGEXP evaluation with a per-eval timeout to prevent ReDoS on the database thread
• Fix JWT exp claim (use max_age per JWT_SESSION_LENGTH) and prevent a new token being issued on logout

Backend

• GenAI: fix Gemini tool calling, catch an OpenAI error, expose whether a provider supports embeddings, clean up llama.cpp and use the API key when configured, and restrict the semantic search tool to English on JinaV1
• Zones: resolve friendly names against the correct camera and render zones with friendly name + ID in the chat prompt
• Face recognition: convert crops to RGB before embedding and catch faces that become empty after cropping
• Runtime camera toggles: persist across config saves, re-apply overrides on save/delete without re-broadcasting, and stop dropping batched add/remove config updates in object processing
• Embeddings and event cleanup: always clean up embeddings so they don't outlive their events, and prune object filters for unsupported labels when cleaning tracked objects
• Various autotracking and ONVIF fixes
• Add -vaapi_device to the birdseye restream vaapi encode preset so hwupload can init on ffmpeg 8
• Resolve export hwaccel args global value against the correct config path
• Don't report auto-populated object/audio filters as camera overrides
• Derive stale replay cameras from bounded directory listings to avoid scanning all clips at startup
• Improve error messages for Intel GPU stats
• Fix notification test button being blocked by websocket auth
• Replace deprecated openvino-dev with a modern package for OpenVINO model generation
• Ensure preview and review directories are valid after deleting a camera from the UI
• Handle indefinite events when a segment needs to forcibly be ended for a camera
• Bump sqlite-vec to 0.1.9
• Documentation updates

Frontend

• Zone editing: disable the ID field when a zone is referenced in any required_zones, and delete all required_zones references when the zone is deleted
• Exports: center the multi-camera export range on the current playback position and fix a crash when camera group icons are invalid
• Motion search/preview: fix motion-preview playback rate and don't stay on the page when the flow is cancelled
• Improve the classification model edit dialog (edit enabled/save_attempts, plus state/motion/interval configs)
• Show a numeric keyboard on mobile for number-specific fields in Settings
• Fix the birdseye reordering widget in Settings
• Update several object icons for Frigate+ 2026.2
• Pin react-zoom-pan-pinch to fix History zooming reset
• Fix minor export/review activity calendar timezone handling bug
• Various frontend tweaks and fixes (logger i18n, scrollbar, overflowing text, etc.)

Full changelog: v0.18.0-beta1...v0.18.0-beta2

Major Changes for 0.18.0

Breaking Changes

There are several breaking changes in this release, Frigate will attempt to update the configuration automatically. In some cases manual changes may be required. It is always recommended to back up your current config and database before upgrading:

1. Simply copy your current config file to a new location
2. Stop Frigate and make a copy of the frigate.db file

• GenAI now supports multiple providers. As a result, the global genai config is now a mapping of keys to GenAI provider setups, and a new roles field defines what tasks each provider will be used for (object descriptions, review summaries, embeddings, chat, etc.). Existing configs will be automatically migrated. Documentation
• Frigate no longer saves annotated JPEG snapshots (.jpg) to disk. Only a clean, unannotated WebP snapshot (<camera>-<id>-clean.webp) is now stored in /media/frigate/clips/. The clean_copy configuration option has been removed — clean snapshots are always saved when snapshots are enabled. The snapshots.quality default has changed from 70 to 60, reflecting the switch from JPEG to WebP encoding. Users who previously relied on the annotated .jpg files on disk should instead use the /api/events/<id>/snapshot.jpg endpoint, which now honors timestamp, bounding_box, crop, height, and quality query parameters for both in-progress and completed tracked objects. When query parameters are omitted, the camera's snapshots config values are used as defaults. Existing configs containing clean_copy will be automatically migrated. Documentation
• The config format for zones and masks has changed to add enabled and friendly_name fields. A config migrator will make the necessary changes, but if migration fails you may need to edit your masks and zones manually.
• Intel GPU stats reporting has changed to no longer require intel_gpu_top and now reads directly from the kernel's per-client DRM usage counters. This removes the need for CAP_PERFMON, privileged mode, or perf_event_paranoid tuning, and adds support for newer Intel GPUs (including Battlemage). Reported usage values may differ from previous versions because the measurement method has changed; the underlying GPU workload is unchanged for the same configuration.
  • Note that this change will require a Linux kernel version of 6.5 or newer on your host. Kernels older than 6.5 do not expose per-engine counters for GPUs using GuC submission.
• FFmpeg has been updated to version 8.
  • FFmpeg 5 is deprecated and will be removed from Frigate in 0.19.
  • FFmpeg 8 introduces a known issue with go2rtc streams that transcode video using hardware acceleration. A small change to the go2rtc config is required. Documentation
• The /api/export DELETE HTTP API endpoint has been removed and refactored to support bulk deletion of exports. A POST to /api/exports should be used instead. Documentation
• The sync_recordings config option has been removed and replaced by the new Media Sync pane in Settings.
• The timelapse_args export config option has been removed and replaced by custom exporting ffmpeg arguments. Documentation
• The ui.date_format and ui.time_format config options have been removed as these fields have been unused since date/time format and handling has been determined by i18n.

New Features

Frigate 0.18 introduces several major new features.

Full UI Configuration

Frigate 0.18 introduces complete UI-based configuration management. The new Settings experience provides:

• A categorized sidebar covering every section of the config (cameras, detectors, motion, recording, review, GenAI, semantic search, face recognition, LPR, classification models, networking, authentication, and more).
• Per-field validation, inline descriptions, and links back to the documentation.
• A clear distinction between global and camera-level configuration, with the ability to override global values per camera and reset back to defaults at any time.
• Highlighting of unsaved changes, a "Save All" button that summarizes pending edits across sections, and a Reset to Default/Global button on each section.
• Indicators showing which fields require a restart to take effect, with a one-click restart prompt when a restart-required field is changed.
• Most of Frigate's configuration options can now be saved and applied dynamically without restarting, including most enrichment, motion, ONVIF, GenAI, and go2rtc stream settings.
• Manual YAML file configuration is still fully supported. Documentation pages now include "Frigate UI" tabs alongside YAML examples to show exactly where each option lives in the new Settings UI.

See the documentation.

Profiles

Profiles let you define named overrides on top of your base camera configuration and switch between them on the fly — without restarting Frigate. Profiles can override a camera's detection, motion, recording, snapshots, review, zones, notifications, objects (including masks and filters), enrichments (face recognition, LPR), audio, and birdseye settings, as well as the camera's enabled state.

• Profiles are defined at the top level of the config with optional friendly names, and per-camera overrides are defined under each camera.
• The active profile is persisted across restarts and can be changed from the UI or via the frigate/profile/set MQTT topic.
• The current profile state and the list of available profiles are published over MQTT.
• The Settings UI shows which fields are overridden by the active profile, supports editing the base config independently from any profile, and provides per-profile camera enable/disable from the camera management view.
• A dedicated Profiles page in the main menu shows a card-based summary of all configured profiles and their overrides.

See the documentation.

GenAI

Frigate 0.18 includes a major expansion of GenAI capabilities. In addition to multi-provider support (see Breaking Changes), there is a new dedicated llama.cpp provider with automatic model probing and context sizing, and GenAI configuration can now be updated dynamically without restarting Frigate.

See the documentation.

Chat

Frigate 0.18 introduces a Chat interface for talking to your configured GenAI provider directly from the Frigate UI. Chat is implemented as a tool-calling LLM agent with access to Frigate-specific tools, so you can ask natural language questions about your cameras, tracked objects, and review activity.

• Streaming responses with markdown support (including tables), per-message copy buttons, message editing, and the ability to stop a streaming response.
• Tool calls are grouped and displayed inline as the agent works, with quick links to relevant tracked objects and review items.
• Tools include searching tracked objects, finding similar objects via semantic search, inspecting live camera context, toggling camera features, starting VLM monitoring on a camera, and a recap tool for summarizing recent activity.
• Tracked object thumbnails and similarity results render as cards in the chat, and any event can be attached to a chat message via the paperclip button or by clicking a thumbnail.
• A persistent quick-replies row makes it easy to start common queries from a fresh chat.

Remote Embeddings

Semantic Search embeddings can now be generated by an external GenAI provider instead of running locally. This is especially useful when you would prefer to offload embedding work to another machine or to a hosted provider.

• Any configured GenAI provider that supports an embeddings endpoint can be selected as the embedding provider for Semantic Search.
• The semantic search settings UI includes a dropdown for choosing which configured GenAI provider should produce embeddings.

See the documentation.

Motion Review and Motion Search

Frigate 0.18 introduces a redesigned Motion Review experience along with a new Motion Search feature.

• Motion Search. A new motion search workflow lets you draw or select a region on a camera image and find time periods where motion occurred in that area, similar to a motion-based version of object review.
• Motion previews. Motion preview clips are generated and shown directly in the review pane, quickly highlighting areas of motion in the frame where Frigate did not detect and track any objects.
• Heatmap filter. A heatmap-grid filter on the motion review page lets you constrain motion previews to specific regions of the camera image.

See the documentation.

Debug Replay

Debug Replay is a new tool that lets you replay recorded video through Frigate's detection and motion pipelines as if it were a live camera. This makes it dramatically easier to tune motion settings and object filters against real footage rather than waiting for live activity to reproduce an issue.

• Configure the debug replay camera directly in the UI without manually editing your configuration.
• Adjust motion and object detection configuration on the replay camera and see the effect on detections in real time.
• The Debug view's overlays (regions, motion, paths, zones) are fully supported during replay.

See the documentation.

Exports

Frigate 0.18 brings substantial improvements to exports, including fast multi-camera exporting from History, single-click exporting of multiple review items in Review, and more.

See the exports documentation.

Case Management

Exports can now be organized into named cases, making it much easier to manage exports related to a single incident or investigation.

• Cases are displayed separately from individual exports in the UI and can be filtered, edited, and deleted.
• Exports can be assigned to a case at creation time (including from the single-camera and mobile review export dialogs) or moved between cases later.
• Bulk selection (similar to the Review page) supports reassigning or deleting multiple exports across cases at once, with admin-only actions properly gated.
• New cases can be created directly from any export dialog, and exports can be removed from a case without deleting them.
• Stale or interrupted in-progress exports are reaped automatically on startup.

Custom Exports

Frigate 0.18 provides a dedicated exports API that accepts arbitrary ffmpeg arguments, allowing for fully custom export pipelines (custom filters, encoders, output formats, and so on). A CPU fallback path is automatically used if hardware-accelerated encoding fails. For security, custom ffmpeg arguments are restricted for non-admin users and a number of dangerous flags (such as -filter_complex, -vf, and -attach) are blocklisted to prevent arbitrary file reads via filter sources.

In addition, exports now use Frigate's Job infrastructure end-to-end, providing more reliable progress reporting and better cleanup of interrupted jobs. Per-camera record.export.hwaccel_args settings cascade up through camera and global configuration, allowing per-camera tuning of export hardware acceleration. See the recording documentation for export configuration details.

Other Improvements

There are many other improvements to highlight in this release.

Backend

• Masks and zones can now be dynamically enabled and disabled via MQTT. Documentation
• Camera enabled, detect, record, snapshots, and audio toggles now persist across Frigate restarts.
• Added support for the AXERA AI accelerator. Documentation
• Nvidia 50 series GPUs are now supported following an upgrade of CUDA and ONNX Runtime dependencies. As part of this change, the CUDA version was tuned to avoid overaggressive BFC behavior on these cards.
• The Intel drivers have been updated to support Battlemage GPUs.
• There is now a dedicated llama.cpp GenAI provider with automatic model probing and context sizing. Documentation
• Custom (API-created) events can now be configured as detections or alerts based on the event label, can run indefinitely until explicitly ended, and support a configurable pre_capture value when created.
• ROCm has been updated to 7.2.0 and now supports RDNA4 AMD GPUs. JinaV1 also works correctly under the new ROCm. Documentation
• Added VLM monitoring, which runs a configured GenAI provider in a loop against a camera's live view to power chat-driven workflows.
• ONVIF support has been refactored to add profile selection, decouple relative move from autotracking, and dynamically update ONVIF configuration without a restart. Documentation
• A new generic process watchdog auto-restarts subprocesses that crash or stall.
• The license plate detection box coordinates are now published in tracked_object_update MQTT messages. Documentation
• A new frigate/<camera>/<role>/status topic and a stopped value on the LWT topic make it easier to monitor camera health and Frigate availability. Documentation
• Configurable listening ports via the networking config section for advanced use cases. Documentation
• New /api/recordings endpoints for deleting recordings, syncing media (replacing the old scheduled sync_recordings cleanup), and pulling review clips by review item ID.
• Snapshot responses now include an X-Frame-Time header.
• Frigate+ API calls now run in a thread pool to avoid blocking the event loop.
• API performance improvements for several recording related endpoints.
• ffmpeg has been updated to 8.1.1, and go2rtc to v1.9.14.
• HTTP/2 support has been enabled in nginx. Users who access Frigate over port 8971 instead of port 5000 may see a significant performance improvement in Live and History view, especially if their installation has more than a couple of cameras.
• Face recognition optimizations to increase accuracy and reject poor candidates.
• Many internal type-safety improvements and a significant increase in mypy coverage across the codebase.
• A number of security hardening fixes around authentication, role enforcement, custom export ffmpeg arguments, and SQL parameterization.

Frontend

• There is a new Media Sync function which allows checking for orphaned files which can be deleted directly from the UI. A verbose mode writes a per-media-type report of orphaned paths to /config/media_sync. Documentation
• A new camera connection quality indicator is shown on the camera metrics page, making it easy to see when cameras are having connection issues.
• Greatly improved the efficiency of metrics loading and rendering, including hiding rather than unmounting graphs when switching tabs.
• The Classification Wizard now allows regenerating examples in cases where the images are not representative of all scenarios, and supports reclassifying both images and faces after training.
• The system metrics page now shows GPU and NPU temperatures (Nvidia, Rockchip, Hailo, and others where available).
• A new Messages tab in the Logs section and Debug Replay shows the messages exchanged between Frigate's backend and the frontend/MQTT.
• Preview generation has improved, with smoother transitions between hours and a new qmax parameter for higher-quality previews at high and very_high settings.
• Cameras can now be deleted directly from the UI, with full cleanup of associated media.
• Instant snapshots can now be downloaded directly from the controls in History view.
• The Tracking Details pane now shows additional scoring information including computed and top scores per timeline entry.
• The motion and object debug overlays now match between Debug view and Replay objects lists, and area is shown as a proper percentage.
• Camera group and streaming settings are now saved to the browser local storage separately for each user.
• The frontend has been migrated to React 19, with corresponding upgrades across the dependency tree (framer-motion, react-konva, react-grid-layout, vaul, sonner, and others).
• The WebSocket layer has been rewritten on top of useSyncExternalStore for finer-grained, per-topic subscriptions and significantly fewer unnecessary re-renders.
• Many fixes around mobile usability, including export crash fixes on iOS, viewport scaling tweaks, and improved drawer/dialog behavior.
• New translations and substantial improvements to existing translations across many languages, including a new i18n key check in CI to keep translations in sync.
• Many small UX improvements and tweaks.

---

This discussion was created from the release 0.18.0 Beta 2.

[ofcedwards]

I know it's been available for a few versions now, but I've only just started using genAI features with this beta, using llama.cpp as the provider. I'm really impressed with the descriptions that have been generated, especially on a relatively old/underpowered GPU, but I noticed a few issues.

The first issues was on this version: 0.18.0-a573ea4. I'm not 100% sure on the second. It was a newer build than beta 1, but older than 0.18.0-a573ea4.

First: I'm using a custom multi-line prompt for object descriptions. I have genAI debug logging enabled for testing purposes, and found the logs to not handle this prompt very well. The prompt seems to get split at every line break, creating a lot of log entries. An excerpt of this is below.

2026-07-26 18:14:02.778794770  [2026-07-26 18:14:02] frigate.genai                  DEBUG   : Sending images to genai provider with prompt: Your task is to analyze a sequence of images containing a person taken in chronological order from a security camera.
2026-07-26 18:14:02.778797078  
2026-07-26 18:14:02.778798228  ## Image Context
2026-07-26 18:14:02.778799213  
2026-07-26 18:14:02.778800700  These images are from a residential security camera that covers or is affixed near this area: driveway.
2026-07-26 18:14:02.778801509  
2026-07-26 18:14:02.778802429  ## Analysis Guidelines
2026-07-26 18:14:02.778803213  
2026-07-26 18:14:02.778804824  - Focus on the behavior of the person based on its actions and movement, rather than describing its appearance or the surroundings.

My prompt is longer, but I think this demonstrates what I'm trying to say.

Second: I'm new to using llama.cpp and locally-hosted AI in general, so this might be my ignorance. But I was using a minimally-configured instance of llama.cpp that automatically set 4 slots with a shared context size of 8192 between all slots. Frigate picked this up automatically as an 8192 context size, and llama.cpp ended up crashing during a period of high activity, I believe because the 4 slots demanded more than 8192 total context. I have a minimal excerpt of Frigate logs from that period, which are below. I've since manually set the number of slots and context size per slot, instead of using the automatic shared context from llama.cpp, and I haven't had any issues. I don't think there's really much to do on the Frigate side here, since I believe Frigate sent proper requests based on the information it had, but this might be something worth adding as a warning in the docs. Again, I'm new to local AI, so this one might just be completely my mistake on the setup.

2026-07-24 17:58:45.012922428  [2026-07-24 17:58:45] frigate.genai                  DEBUG   : Sending 66 images to create review description on Driveway
2026-07-24 17:58:49.099702685  [2026-07-24 17:58:49] frigate.genai.plugins.llama_cpp DEBUG   : llama.cpp model '/models/model.gguf' uses ~50 tokens for 240x180 images
2026-07-24 17:58:49.125859388  [2026-07-24 17:58:49] frigate.genai                  DEBUG   : Sending 28 images to create review description on Front Door
2026-07-24 17:59:01.964267841  [2026-07-24 17:59:01] frigate.genai.plugins.llama_cpp WARNING : llama.cpp returned an error: ('Connection aborted.', RemoteDisconnected('Remote end closed connection without response'))
2026-07-24 17:59:01.964486426  [2026-07-24 17:59:01] frigate.genai                  DEBUG   : Invalid response received from GenAI provider for review description on Driveway. Response: None
2026-07-24 17:59:01.964763507  [2026-07-24 17:59:01] frigate.genai.plugins.llama_cpp WARNING : llama.cpp returned an error: ('Connection aborted.', RemoteDisconnected('Remote end closed connection without response'))
2026-07-24 17:59:01.965012109  [2026-07-24 17:59:01] frigate.genai                  DEBUG   : Invalid response received from GenAI provider for review description on Front Door. Response: None

[NickM-27]

If you want to see the prompt and result it's recommended to enable debug save thumbnails. The logs are meant for debugging individual items not leaving it enabled.

Yeah generally more context is needed. Every local AI has a different default implementation so it's difficult to say for sure.

[EPiSKiNG]

Does ghcr.io/blakeblackshear/frigate:0.18.0-beta2-rk have support for ffmpeg 8.0 ?

[NickM-27]

You can bring your own but one is not included