I Have the same issue. What did you do exactly to fix it?
Never mind, it’s done.
don’t use
media: “/media”
this works:
homeassistant:
allowlist_external_dirs:
- /config/media/llmvision/snapshots
media_dirs:
llmvision: /config/media/llmvision/snapshots
Hi,
first of al, This is a very nice integration!!
i started to use it a couple of days ago. ’
2 questions:
i have 2 camera’s in my setup in one automation. google ai tells me what is happening from both camera’s streams. And the automation save’s the keyframe. strange thing is: Why does it seems to use only 1 camera for this keyframe? even though the google analysis says nothing to see on cam 1, activity on cam 2, the saved image is from cam1.
next question:
i can’t get the image shown on the custom:llmvision-preview-card and no images on the timeline functionality.
the key frames are stored correctly in my opinion, because this automation sends me the key frame image via telegram.
alias: AI achterdeur
description: ""
triggers:
- type: motion
device_id: 65eddb84809cc6c5861fe39898231642
entity_id: e754dec645b108ebf718301857f5c15b
domain: binary_sensor
trigger: device
for:
hours: 0
minutes: 0
seconds: 5
- type: motion
device_id: 0e51d87576bf1d2d5a164b01be3f6343
entity_id: f8f267a2a3253ae9a66473422c39ff30
domain: binary_sensor
trigger: device
for:
hours: 0
minutes: 0
seconds: 5
conditions: []
actions:
- action: llmvision.stream_analyzer
metadata: {}
data:
remember: true
duration: 6
max_frames: 5
include_filename: true
target_width: 1280
max_tokens: 400
generate_title: true
expose_images: true
provider: **
message: >-
Write a short description in dutch language of person(s) detected at the
driveway, garden, fence or back door, and what is happening, based on
this camera live feed. The description should highlight observable
traits like clothing, hair color, accessories, age or gender. Keep the
response under 1000 tokens.
image_entity:
- camera.voortuin_vloeiend
- camera.achtertuin_vloeiend
response_variable: llm_response
- action: telegram_bot.send_message
metadata: {}
data:
config_entry_id: **
message: "{{ llm_response.response_text }}"
- action: telegram_bot.send_photo
metadata: {}
data:
config_entry_id: **
file: /config/{{ llm_response.key_frame.replace('/config', '') }}
max_exceeded: silent
mode: single
Good afternoon. Is it possible to integrate Yandex AI Studio which partially supports compatibility with OpenAI API and the Gemma 3 27B model? Attempting integration via custom OpenAI does not work because it uses a Custom Endpoint https://llm.api.cloud.yandex.net/v1, but besides the API key, it also requires a catalog identifier. Here’s reference documentation: Yandex Cloud Documentation.
The newest release fixed the camera issue as i described.
But still no pictures in the timeline. New behavour:
When i klik on a even, the ai tekst appears. (just like before).
But in the log:
Login attempt failed
Login attempt or request with invalid authentication from 82-169-61-77.fixed.kpn.net (82.169.61.77). See the log for details.
Did you upgrade the Card as well? You might also need to force refresh the cache on your Browser/App. See this issue here: Invalid Authentication Error and No images in TimeLind Card - New 1.5.2 release · Issue #65 · valentinfrlch/llmvision-card · GitHub
This is likely a rate limit issue: See these issues:
Also, try increasing max_tokens.
Thx for your quick reply. Clear cache did the job. Thx
When I ditched Ring for ReoLink cameras I really missed the notifications for packages on my porch. Thanks to this thread and community, I was able to automate my own!
To say thank you 
, and because I feel it’s my duty to give back, here’s my automation:
id: notification_package_detected
alias: Notification Package Detectedtrigger:
platform: state
entity_id: binary_sensor.doorbell_person
to: ‘on’platform: time_pattern
minutes: ‘/20’action:
sequence:
TAKE CAMERA SNAPSHOT
- delay: '00:01:00' - service: camera.snapshot data: entity_id: camera.doorbell_fluent filename: "/config/www/CCTV/check_for_packages.jpg" - delay: '00:00:05'CHECK CAMERA IMAGE FOR PACKAGES
- metadata: {} data: prompt: > There is a package on or near the steps. Answer as one word: True or False. filenames: "/config/www/CCTV/check_for_packages.jpg" response_variable: generated_content action: google_generative_ai_conversation.generate_contentSET PACKAGE STATUS TO TRUE OR FALSE
- metadata: {} action: input_text.set_value data: value: "{{ generated_content['text'] }}" target: entity_id: input_text.package_detectedTOGGLE INPUT BOOLEAN
- service_template: > {% if is_state('input_text.package_detected', 'True') %} input_boolean.turn_on {% else %} input_boolean.turn_off {% endif %} entity_id: input_boolean.package_detectedSEND IOS NOTIFICATION
- metadata: {} data: title: EZ House Alert message: Package Detected at Front Door data: image: '{{ filename }}' entity_id: camera.doorbell_fluent action: > {% if is_state('input_text.package_detected', 'True') %} notify.mobile_app_zachs_iphone {% endif %}mode: single
Thank you! 
after updated to Version 1.5.2 it stop or dont use my memory i have about 60-80 pictures in memory that it stop use is it some data base I ned to go back to?
I am trying to add openrouter to the llm vision integration. When I add the api_key and choose for the openai/gpt-4o-mini model I get an error saying "home assistant can not add open router to llm vision message “Could not connect to the server. Check your API key or IP and port”
I am sure the api key is correct. What can be wrong ?
Images missing in notifications just me? I can see them in the media directory and they appear in the timeline card with the description. I tested using the reolink example code (no llm vision) with the snapshots in the same media directory and that worked. Snapshots are in media\llmvision\snapshots I did notice in the notification configuration for llm vision it indicates those photos should be in sub directories for each camera which mine are not.
Am I alone and or any other tips to debug?
I was just trying with Claude Sonnet 4.5, but getting the error:
“Full Response: {“type”:“error”,“error”:{“type”:“invalid_request_error”,“message”:“temperature and top_p cannot both be specified for this model. Please use only one.”}”
I tried setting temperature to 0 but still getting the error. I guess there should be a way to disable one of those 2 settings? Thanks!
Hi, this project is amazing! Would be nice to be able to put URL option for the file path for people that don’t keep frigate as home assistant add on. From what I understand currently you can only add local file path. Thank you!
This might be why I keep getting 400 errors with ollama and llmvision. Thanks for the heads up.
This may be a simple referencing mistake, but do I have the correct format here? Is there something needed before referencing the variable or an escape sequence of some kind?
I’m trying to get my Alexa device to read the {{response.response_text}}. But the blueprint based automation keeps erroring out
“Stopped because of unknown reason at December 4, 2025 at 4:54:04 PM (runtime: 0.01 seconds)”
If I change what it reads to {{ label }} it works just fine and reads it aloud. I looked all over the blueprint, and it would appear both {{ label }} and {{ response.response_text }} are valid.
this works just fine,
- action: notify.send_message
metadata: {}
data:
message: "{{ label }}"
target:
entity_id: notify.joe_s_echo_speak
this does not work,
- action: notify.send_message
metadata: {}
data:
message: "{{ response.response_text }}"
target:
entity_id: notify.joe_s_echo_speak
Just curious, before I drop the cash on a 360 camera, has anyone played with these types of cameras and seen how LLMs interpret the warped footage?
On my side the same. I am on iOS and for the sake I cannot make it happen that I see a snapshot in my notification.
Is it possible to add a delay from issuing the snapshot to saving it and sending it to analysis in the blueprint? My Arlo cameras take forever to produce one so its often the snapshot from the last trigger that ends up being sent for analysis.
I have noticed the same
I timed it, from sending command to actually seeing a new snapshot is 5-7 seconds so the delay has to be after a snapshot request.