Immich remote machine learning
Separate Immich Machine Learning Server: When It Helps
Immich can move its machine-learning container to a more powerful computer, which is useful when the main photo server is memory-constrained or optimized for low power. This architecture is most valuable when Smart Search and face-detection queues are slow but the main server is otherwise adequate. It should not be treated as a generic way to split every Immich task across two machines.
Quick answer
What this Immich workload needs
Use remote machine learning when the main server is the bottleneck for Smart Search or Face Detection and you already have, or can justify, a stronger second host. Immich notes that image previews are sent to the remote ML container and that the service has no built-in security measures, so keep it on a trusted private network.
Live Amazon hardware
Current products that fit this decision
These listings come from this sprint's dedicated Amazon catalogue. Barebones systems, laptops, enclosures, adapters, unrelated accessories, ambiguous product classes and obvious multipacks are excluded by the normalizer.
Buying decision
Match the hardware to the bottleneck
Decide whether the separate host is always-on, shared or on-demand. The planner weighs library size, ML features, current queue pressure and available accelerator hardware before recommending whether separation is worth the extra system.
Interactive decision tool
Remote Machine Learning Planner
Use the workload inputs to identify a practical hardware tier or likely bottleneck before comparing live products. Results are planning guidance; verify current Immich support and the exact hardware configuration before deployment.
Compatibility checklist
Four checks before you purchase
Use a supported deployment
Immich currently favors Linux and Docker Compose for production; full VMs are supported while Docker-in-LXC is not recommended.
Keep PostgreSQL local
Do not treat a faster network as permission to move the standard database data path onto a network share.
Verify hardware access
Quick Sync, NVENC, VAAPI and ML acceleration need the right host, drivers and container device access.
Measure before upgrading
CPU, RAM, SSD, GPU and network bottlenecks create different symptoms. Change the measured bottleneck first.
Remote ML solves a specific bottleneck
Immich’s remote machine-learning feature exists to move expensive ML processing away from a constrained server. It is particularly relevant to small-memory devices or efficient application hosts that are otherwise good enough. If browsing, database and video jobs are the real problems, adding a separate ML box may leave the main bottleneck untouched.
Smart Search and Face Detection use the remote service
Immich documents that Smart Search and Face Detection can use remote machine learning. Facial Recognition itself uses the saved outputs in the database, so the workflow is more nuanced than saying every face-related task runs remotely. Understanding that split helps diagnose which jobs should become faster after the remote service is enabled.
Image previews cross the network
The server sends image previews to the remote ML container for processing. Immich warns that the remote service is an internal service with no security measures of its own. Keep it on a trusted LAN, avoid exposing the port directly to the internet and be cautious about paid or public processing hosts where preview data would leave your controlled environment.
A second Intel host can be enough
Remote ML does not automatically require an expensive NVIDIA card. Immich supports accelerated ML backends including OpenVINO for Intel graphics, CUDA for NVIDIA, ROCm for AMD and several ARM/Rockchip paths. A spare Intel mini PC may provide a meaningful improvement over a very weak main server while remaining efficient and easy to cool.
Discrete GPUs help larger or more frequent queues
A GPU becomes more compelling when the library is large, models are heavier or new assets arrive continuously. Immich notes that larger models benefit more from hardware acceleration when enough VRAM is available. Buy based on actual queue pressure and model requirements rather than assuming every home library needs a gaming-class card.
Version matching matters across both hosts
Immich warns that version mismatches between the main server and remote machine-learning container can cause bugs and instability. A second host therefore adds an update responsibility. Document the compose file and upgrade both sides together rather than treating the remote box as a forgotten appliance.
Fallback behavior can improve resilience
Immich recommends adding the remote URL rather than simply replacing the local ML URL when you want local fallback. That way jobs can continue if the remote machine is unavailable. If the main server is intentionally too weak to process ML, you may choose remote-only behavior, but then failed remote availability means ML jobs will not complete until the service returns.
Multiple URLs are not automatic load balancing
Immich states that multiple ML URLs are tried sequentially rather than automatically load balanced. Running several remote containers only helps as a true pool when an external load balancer coordinates them. This matters for buyers: two GPUs on two machines do not automatically double throughput just because both URLs are entered.
Network speed is usually not the primary limit
Remote ML sends previews rather than the full original library for every operation, so basic wired networking can be adequate. Latency and reliability still matter, especially during large batch jobs. Upgrade the network for the whole storage and backup architecture, not solely because machine learning is on another host.
A shared workstation can be an economical accelerator
Some households already own a powerful desktop with a supported GPU. Running remote ML there can avoid buying a second dedicated high-end server, especially for one-time indexing of a large library. The trade-off is availability: jobs pause or fall back when the workstation is off. That may be perfectly acceptable if new asset volume is modest.
The main server still needs adequate core resources
Moving machine learning does not eliminate Immich’s CPU, RAM, database and storage requirements. The main server continues handling uploads, application logic, database work and other background jobs. A remote ML host is an accelerator tier, not a replacement for a stable primary server.
Separate ML is an architecture, not an upgrade reflex
If the main server has an underused Intel iGPU or existing discrete GPU, local acceleration may be simpler. Separate ML is most valuable when it preserves a low-power primary host, reuses stronger hardware or isolates bursty AI work. Compare the cost, energy and maintenance of two machines against upgrading the primary server once.
Questions people ask
Separate Immich ML Server questions
Can Immich run machine learning on another computer?
Yes. Immich supports a remote machine-learning container that the main server can call over the network.
Which Immich jobs use remote machine learning?
Immich says Smart Search and Face Detection use the remote service. Facial Recognition uses saved model outputs in the database rather than directly calling remote ML.
Is the remote ML service secure to expose online?
No. Immich warns that the machine-learning container is an internal service without security measures. Keep it on a trusted network.
Can I use an Intel mini PC for remote ML?
Yes when the hardware and supported backend fit the models. Immich supports OpenVINO for Intel GPUs.
Do I need NVIDIA CUDA for remote ML?
No. CUDA is one supported backend among several. Intel, AMD and certain ARM/Rockchip acceleration paths are also documented.
What happens if the remote ML server is off?
If the local ML URL remains configured, Immich can fall back. If you configured remote-only processing, affected ML jobs will fail until the remote service returns.
Can I use multiple remote ML servers?
You can enter multiple URLs, but Immich says they are tried sequentially and are not automatically load balanced.
Does the remote ML version need to match?
Yes. Immich warns that version mismatches between hosts may cause bugs and instability.
Will remote ML speed up video transcoding?
No. Video transcoding and machine learning are separate services and acceleration paths.
When is a separate ML server worth it?
When Smart Search/Face Detection queues are a real bottleneck and moving them lets you keep the primary Immich server simpler or lower power.
Official references and methodology
Verify current Immich support before deployment
Cloudzat follows Immich’s current remote-ML behavior and security warnings, then layers workload sizing on top. Live mini PCs and GPUs are candidates from a dedicated catalogue; the page does not claim a specific model is supported without checking the selected backend, drivers, VRAM and deployment.
- Immich - Requirements
- Immich - Hardware Transcoding
- Immich - Hardware-Accelerated Machine Learning
- Immich - Remote Machine Learning
- Immich - Docker Compose Installation
- Immich - System Settings
As an Amazon Associate, Cloudzat may earn from qualifying purchases. Prices, seller terms, exact configurations, firmware and software support can change.