Help › Choosing an AI Model: Local vs Cloud for Lightroom

Help: Choosing an AI Model

The exact model lists exposed by the plugin come from the backend at runtime. The names below reflect the curated lists shipped with the current backend (server-rs/crates/lrg-providers/ for cloud providers, server-rs/crates/lrg-api/src/{llm_models,mlx_models}.rs for the built-in local ones). Pricing and availability change over time — verify with each provider before relying on production cost estimates.

Decision factors

Choose based on:

  • privacy requirements (cloud vs. local)
  • quality expectations (description detail, keyword accuracy, edit recipe sanity)
  • runtime per image and batch throughput
  • per-image cost (cloud) or hardware cost (local)
  • available local hardware (VRAM/RAM, Apple Silicon vs. discrete GPU)

Cloud models

Google Gemini

Configure in Plug-in Manager → Optional AI providers → Google Gemini key. Models exposed today:

  • gemini-2.5-flash-lite — cheapest and fastest; good for bulk keywording.
  • gemini-2.5-flash — balanced default for analyze-and-index runs.
  • gemini-2.5-pro — highest 2.5-tier quality; use for tricky scenes or when description quality matters more than throughput.
  • gemini-3-flash-preview, gemini-3.1-flash-lite-preview, gemini-3.1-pro-preview — latest preview tier. Expect higher quality and better instruction following, but preview pricing/quotas can change.

How much a Gemini model thinks follows Analysis depth in the Analyze & Index dialog: every Gemini 3 model (gemini-3-*, gemini-3.x-*) gets the matching thinking level (Fast = low, Balanced = medium, Thorough = high), and gemini-2.5-* a thinking budget (the smallest the model takes at Fast, 2048 tokens at Balanced, the model’s own choice at Thorough). Thinking tokens count against the token limit, so the backend adds room for them on top of Max Tokens (4096 / 8192 / 16384 tokens by depth) whenever the model thinks; they are billed as output and included in the token totals the server logs. A Gemini 3 model that does not offer Balanced is asked again at Thorough, and the run says so.

Gemini 3 is sent no temperature: Google asks for its default there, and lower values can make it repeat itself. The Temperature slider still applies to gemini-2.5-*.

If a Gemini model reports that it stopped because the token limit was reached, raise Max Tokens to 4096 or higher. If the same photo keeps failing at a higher limit, the model is looping rather than running out of room; switch to a different Gemini model instead of raising the limit further.

OpenAI / ChatGPT

Configure in Plug-in Manager → Optional AI providers → OpenAI key. Models exposed:

  • gpt-4.1 — proven vision quality; the safe baseline.
  • gpt-5-nano, gpt-5-mini, gpt-5 — current GPT-5 tier; pick nano/mini for batch jobs and gpt-5 for higher-fidelity descriptions.
  • gpt-5.4-nano, gpt-5.4-mini, gpt-5.4, gpt-5.4-pro — newest GPT-5.4 tier; gpt-5.4-pro is the highest-quality option but the most expensive.

Note: GPT-5 and GPT-5.4 models ignore the Temperature slider. Their reasoning_effort follows Analysis depth instead (Fast = low, Balanced = medium, Thorough = high). gpt-4.1 does not reason: it uses the temperature and ignores Analysis depth.

GPT-5 models spend part of the token limit on internal reasoning before writing a single word of the answer, so the backend adds room for it on top of Max Tokens (4096 / 8192 / 16384 tokens by depth). If a GPT-5 model still reports that it stopped because the token limit was reached, lower Analysis depth or raise Max Tokens.

Anthropic Claude

Configure in Plug-in Manager → Optional AI providers → Anthropic key (create one at platform.claude.com). The list is not curated: it shows every Claude model your key can use that reads photos and supports structured outputs, as Anthropic’s own model list reports them, so new Claude models appear without a plugin update. As a starting point:

  • claude-haiku-4-5 — fastest and cheapest; good for bulk keywording.
  • claude-sonnet-5-5 — balanced default for analyze-and-index runs.
  • claude-opus-5-5 — highest description quality of the three.

This uses Anthropic’s own API, not its OpenAI-compatible endpoint. Entering https://api.anthropic.com/v1 as the Other AI server does not work: that endpoint cannot list models (it answers “anthropic-version: header is required”) and ignores the answer format. Use the Anthropic key field instead.

What the backend does for you:

  • Answer format. Every answer is held to the requested fields by Anthropic’s structured outputs, so keywords, caption and title arrive in the expected shape.
  • Thinking. Newer Claude models think before answering, and some cannot turn it off. The backend sends Analysis depth as the model’s effort (Fast = low, Balanced = medium, Thorough = high) where the model lists that level, and otherwise leaves the model at its own default. It adds room for thinking on top of Max Tokens (4096 / 8192 / 16384 tokens by depth), so a photo is not cut off by reasoning you never see. Thinking tokens are billed as output.
  • Temperature. Current Claude models reject a temperature setting, so the plugin’s temperature slider does not apply to them.
  • Prompt caching. The part of the prompt that is the same for every photo of a run (instructions, keyword vocabulary, taxonomy) is cached on Anthropic’s side, so from the second photo on it costs a fraction of the normal input price. This only kicks in once that part is long enough (roughly 500–4,000 tokens, depending on the model).
  • Safety filter. If a model declines a photo, the task reports it with Anthropic’s reason. For the newest models (Opus 5.5, Opus 5, Sonnet 5.5, Fable 5/5.1) the request lets Anthropic retry a declined photo on the model it recommends; the photo is then indexed and the task tells you which model answered.

If Claude reports that it stopped because the token limit was reached, raise Max Tokens to 4096 or higher. If the same photo keeps failing, the model is repeating itself; pick a different Claude model.

Local models

Local providers run on your own machine, so privacy is the strongest argument for using them. Quality of small open-weights vision models has improved significantly, but cloud frontier models still lead on tricky scenes.

There are two kinds of local option: the engines built into the backend (nothing else to install), and external servers you run yourself (Ollama, LM Studio, or any OpenAI-compatible server as the Other AI server).

Built-in: llama.cpp and MLX (no external app)

The backend runs vision models itself. Open Plug-in Manager → LrGeniusAI, find the Local AI Model sections, pick a model, and click Download — it then appears in the model dropdown as On this Mac · <model> or On this PC · <model>.

  • llamacpp — llama.cpp compiled into the backend, using GGUF models. Shipped in the Windows release (Vulkan, any GPU vendor). Recommended entries: Gemma 4 E4B as the default, Gemma 4 12B (QAT) if you have 24 GB of RAM, Ministral 3 8B or Qwen3.5 9B as alternatives, Qwen2.5-VL 3B on modest hardware.
  • mlx — Apple’s MLX stack via a small Metal helper process, Apple silicon only. Recommended entries: Gemma 4 E4B as the default, Gemma 4 E2B for speed, Ministral 3 8B or Qwen3-VL 4B as alternatives.

The macOS release ships MLX only. A backend built from source with the llamacpp feature offers both on a Mac, which is worth a side-by-side run on the same 10–20 photos: MLX is Apple’s native inference stack, while llama.cpp reuses the shared prompt prefix across the photos in a batch (MLX re-processes it per photo), which matters more the larger the batch.

Either engine can also download a model that is not in the list — pick Other model from Hugging Face…; it is checked before anything is downloaded. See Local AI Models.

Both reuse models you already have: llama.cpp picks up GGUFs under ~/.lmstudio/models, and MLX picks up LM Studio’s MLX models and the huggingface-cli cache. Full guide: Local AI Models.

One caveat that applies to both: do not enable keyword aliases or bilingual keywords with a local model. They turn every keyword into a structured object, which small models handle badly — the Analyze & Index dialog warns about it.

Ollama

Install and start Ollama from ollama.com, then pull at least one vision-capable model. Recommended starting points:

ollama pull qwen3-vl:4b-instruct-q4_K_M     # fast, ~6 GB VRAM
ollama pull qwen3-vl:8b-instruct-q4_K_M     # better quality, ~10 GB VRAM
ollama pull gemma3:4b-it-q4_K_M             # good general default
ollama pull gemma3:12b-it-q4_K_M            # higher quality if you have VRAM
ollama pull llava                            # legacy fallback

Browse all vision models: ollama.com/search?c=vision. Nothing to configure in the plugin: Ollama on this computer is found automatically. See Ollama Setup.

LM Studio

Worth running as an external server mainly if you already use it for other things, or want its model browser and per-model tuning; otherwise the built-in engines above cover the same ground with one fewer app running.

Download from lmstudio.ai, enable server mode, and download one or more vision models from inside the app. Recommended:

  • qwen/qwen3-vl-4b — fast baseline.
  • qwen/qwen3-vl-8b — better description quality.`
  • gemma-4-e4b / google/gemma3-12b — strong general-purpose options.

On Apple Silicon prefer the MLX variants of the same model — they run significantly faster than the GGUF builds. Like Ollama, LM Studio on this computer is found automatically. See LM Studio Setup.

Other AI server (OpenRouter, llama.cpp server, LiteLLM, vLLM)

Any server that speaks the OpenAI chat API can be entered as the Other AI server in Plug-in Manager → Optional AI providers, with an API key if it needs one. That covers OpenRouter (hundreds of cloud models behind one key, some free), llama.cpp’s llama-server, LiteLLM, vLLM, and Ollama or LM Studio running on another computer. Its models appear under the server’s name, e.g. OpenRouter · google/gemini-2.5-flash. See Other AI Server.

Quick recommendations

Workflow Suggested first try
Cheap bulk keywording (cloud) gemini-2.5-flash-lite, gpt-5-nano or claude-haiku-4-5
Balanced default (cloud) gemini-2.5-flash, gpt-5-mini or claude-sonnet-5-5
Best description quality (cloud) gemini-2.5-pro, gpt-5.4, gpt-5.4-pro or claude-opus-5-5
Privacy-first, simplest setup Built-in llamacpp with Gemma 4 E4B
Apple Silicon, local Built-in mlx with Gemma 4 E4B (E2B if 8–16 GB)
Windows with any discrete GPU, local Built-in llamacpp (Vulkan) with Gemma 4 E4B
Modest hardware / 8 GB RAM mlx Gemma 4 E2B or llamacpp Qwen2.5-VL 3B
Already running Ollama / LM Studio Ollama qwen3-vl:8b or LM Studio qwen3-vl-8b
Many cloud models, one key Other AI server: OpenRouter
A GPU server on your network Other AI server: llama-server, vLLM or LiteLLM

Practical recommendation

The dropdown in Analyze & Index always reflects what the backend currently advertises — newer models that ship with future backend updates will appear automatically. If a model you expect is missing, check that the corresponding API key or server is configured and reachable from the backend. The Plug-in Manager → Status section reports whether any provider is available, and the line under Other AI server whether that server answers. A saved model that is not offered right now stays selected, marked (not available now) — the task then tells you what to start or fix instead of quietly switching to another provider.

When evaluating, run the same batch of 10–20 representative photos through two candidates and compare:

  • keyword coverage and accuracy
  • description quality and language correctness
  • runtime per image and end-to-end batch time
  • system load (local) or token cost (cloud)

Throughput on cloud providers

Cloud runs are dominated by network round trips, not by the model. The backend overlaps up to four requests at a time for OpenAI, Gemini and Anthropic, which hides most of that latency without looking like a burst to their rate limiters.

Ollama, LM Studio and the Other AI server deliberately stay at one request at a time: the server on the other end is usually a single model that serialises the work anyway, and piling requests on it only competes with Lightroom for the same machine.

Edit this page on GitHub