Own AI
Your models, your computer: Lumavra talks to them directly.
What "Own AI" is
From Pro, Lumavra can use your own AI models instead of our service – on your PC at home, on a Mac, or on a server you rent yourself.
The app talks to your computer directly. None of it goes through our servers: addresses live in your iCloud, keys only in your iCloud Keychain. Pictures on your own model cost no credits, and name searches on your model do not count against your plan's AI runs.
What your model can take on:
- Find names – name ideas for your searches
- Improve – improve and vary prompts in the AI studio, propose picture ideas
- Analyse – describe what a photo shows (needs a model that sees pictures)
- Pictures from words and pictures from a photo – with ComfyUI, filling gaps too
On your model too, Lumavra writes the request and checks the answer against the same rules as with the service. You set it up in the app under Settings › Own AI.
What you need
- For text: LM Studio (or Ollama) with a language model.
- For pictures: ComfyUI with an image model.
- A computer: an NVIDIA graphics card with 8 GB or more, or a Mac with Apple silicon and 16 GB of memory or more. More memory means larger, better models.
Your iPhone and your computer have to reach each other – on the same Wi-Fi, or away from home through your own VPN (see below).
Setting up LM Studio
- Download and install LM Studio from lmstudio.ai.
- Find and download a model under Discover (suggestions in the next section).
- Under Developer, start the server and turn on Serve on Local Network.
LM Studio shows the address, usually
http://<your computer's IP>:1234. - In Lumavra: Settings › Own AI › Add service – type "LM Studio", enter the address. The app reads the loaded models itself.
- Under Tasks, choose the model for "Find names" and "Improve" and tap Check.
Tip: many models "think" at length before answering. Lumavra rarely needs that; without thinking they usually answer in a few seconds.
Which text models?
| Memory | Suggestion |
|---|---|
| 6–8 GB | Bonsai 27B by PrismML – a 27B model, heavily compressed: 1-bit ≈ 3.9 GB, ternary ≈ 7.2 GB, Bonsai 2 27B ternary ≈ 5.9 GB. Sees pictures with its vision file. Or Qwen3.5-9B (Q4). |
| 12–16 GB | Gemma 4 12B or gpt-oss-20b; Qwen3.5-9B at higher quality (Q8) |
| 24–32 GB | Qwen3.8-27B, Qwen3.6-35B-A3B or Gemma 4 31B (each Q4) |
Analyse needs a model that sees pictures – such as Gemma 4, Qwen3.5-9B, Mistral Small 3.2, or Bonsai with its vision file. The numbers are a guide for the model files; long requests need a little more.
Setting up ComfyUI for pictures
- Download ComfyUI from comfy.org (the desktop version) and install an image model – easiest through one of the bundled templates.
- In its settings, allow listening on the network so your iPhone can reach ComfyUI
(usually port
8188). - In Lumavra: Settings › Own AI › Add service – type "ComfyUI". The app recognises which models are there and suggests the fitting ones.
- Under Tasks, assign "Pictures from words" and "Pictures from a photo" and tap Check – with a test picture if you like.
Which image models?
| Model | Memory | Strength |
|---|---|---|
| Z-Image Turbo | 16 GB and up | Photorealistic and fast (8 steps); Apache 2.0 |
| Qwen-Image Edit | 24 GB (fp8) | Pictures from your photo and filling gaps; text in the picture; Apache 2.0 |
| Qwen-Image | 24 GB (fp8) | Pictures from words with clearly readable text; Apache 2.0 |
| FLUX.2 [klein] 4B | 8–13 GB | Fast and versatile; Apache 2.0 |
| FLUX.1 / FLUX.2 [dev] | 24 GB and more | Very high quality; licensed for personal use only |
| SDXL | 8 GB and up | Runs almost anywhere; a huge choice of styles |
Pictures from a photo need Qwen-Image Edit. Mind the model's licence if you use pictures for more than personal purposes.
Away from home
At home on Wi-Fi your computer's address is enough. Away from home you reach it through your own VPN, such as Tailscale or WireGuard – not through us, and without opening ports on your router.
Rent instead of buy
Without a fitting graphics card you can rent models – for text, Azure OpenAI for example; for pictures, a GPU server running ComfyUI, for example at RunPod:
- Start a pod with the ComfyUI template. It is billed by running time, to the millisecond, about 0.50 to 5 USD an hour depending on the graphics card – a 24 GB card (RTX 4090, RTX PRO 4000) is enough for Qwen-Image and costs around 0.60–0.90 USD an hour.
- Keep the models on a network volume: it stays when the pod ends and costs about 0.05–0.07 USD per GB and month (100 GB ≈ 5–7 USD, 1 TB ≈ 50–70 USD). Without one, every new pod downloads the models again.
- Enter the pod's address (
https://…-8188.proxy.runpod.net) in Lumavra under Settings › Own AI as ComfyUI. - Stop the pod when you are done – stopped, the graphics card costs nothing, only the storage does (150 GB about 0.02 USD an hour). Terminating deletes the pod with its storage and the downloaded models; only what is on a network volume stays.
Important: a pod's address is open – anyone who knows it can make pictures on it, and you pay for the time. Don't share it and don't leave the pod running unattended. Keys for rented services stay in your iCloud Keychain.
Responsibility
Your own models have no provider's filter. Lumavra checks the words against the same block lists as with the service – for anything your models create beyond that, you are responsible. The rules in the Terms of Use apply here too.