AI images at home without uploading family photos

By Emmanuel Sauvage · Published · Updated

This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.

You want to rework family photos, try avatars, or sketch room ideas while the files stay off an online image service (for example Midjourney or DALL·E, which run on a company’s servers). Local generation goes through ComfyUI (a node canvas that chains the steps) or Automatic1111 (A1111, a more traditional single-page web UI), with FLUX (a newer image-model family) or SD (Stable Diffusion, the open-source standard since 2022). ComfyUI fits modular workflows: inpainting (filling or changing part of a picture), ControlNet, and FLUX chains. A1111 stays a straighter path if you are following older SD 1.5 or SDXL tutorials.

This use guide sits next to the software tutorial ComfyUI and FLUX locally. The angle here is the household, privacy, and hardware, with a short install path and prompt examples that use no real photo.

Intermediate. You can install software on your PC and read English docs when you have to. You have yet to finish a first local image from start to finish.

Who What you need first A rough duration
A parent or a maker who refuses to upload albums A Windows or Linux PC with an NVIDIA GPU. 32 GB of RAM is the comfortable figure 2–4 hours the first time (drivers, ComfyUI, and an SD 1.5 test)
A household already comfortable with Ollama for text A clear split: pixels and chat are two stacks, even on one machine 30 minutes if the GPU and the disk are already sized (image VRAM, SSD)
New to local AI Start with a first local AI in 15 minutes, then come back This guide assumes you have already tried Ollama or something like it

Getting started. ComfyUI asks for more setup than Automatic1111 (A1111) at the beginning (a graph of nodes, saved workflows). In return it grows more cleanly toward FLUX, inpainting, and longer chains. An old all-in-one SD 1.5 tutorial still works in A1111. To stay aligned with LocAIhome, prefer ComfyUI.

Why the pictures stay home

Online service On your machine
An upload is a copy on someone else’s servers Files live on your SSD
Terms of use can change You choose the checkpoints
A subscription that renews The cost is the GPU plus electricity

Watch system cloud backups (OneDrive, iCloud, Google Drive). They can sync the output folder. Exclude the ComfyUI directory from sync. The steps are below.

Keep ComfyUI out of cloud sync (OneDrive, iCloud)

Service Steps to check (October 2026)
OneDrive (Windows 11) Windows Settings, then Accounts, then Windows backup / OneDrive, then Manage backup. Exclude the ComfyUI folder. Or put models and output on a drive that is unsynced, for example D:\ComfyUI\.
iCloud Drive (macOS) System Settings, then [your name], then iCloud, then iCloud Drive, then Options…. Uncheck sync for the folder that holds ComfyUI, or move the project out of ~/Library/Mobile Documents/com~apple~CloudDocs/.
Google Drive (on a PC) Google Drive for desktop preferences, then Folders from your computer, and remove the parent path of ComfyUI.

Also check Photos apps that automatically upload the output folder. The image model can stay local while an operating-system sync quietly copies the results.

Hardware you need first

Part Cue
NVIDIA GPU VRAM (video RAM, the memory on the graphics card): 8 GB minimum for SD (Stable Diffusion) 1.5; 12 GB and up for SDXL and FLUX. The steps are in VRAM for image models (SD, SDXL, FLUX) (the image section, separate from the text-chat tables)
SSD 50–100 GB free (SSD for models)
RAM 32 GB is the comfortable figure when the GPU is the limit

Mid-range cards (October 2026; check the current price at the store): if you are buying an RTX to move from SD 1.5 to SDXL or FLUX, read the VRAM table above and the hardware selector first (budget, noise, weekend kilowatt-hours). Store search: NVIDIA RTX cards on amazon.ca. Affiliate link. LocAIhome may earn a commission at no extra cost to you. LocAIhome has not lab-tested each GPU. Full disclosure on the affiliate page.

Software: where to begin

The long install stays in ComfyUI and FLUX locally (drivers, checkpoints, workflows). ComfyUI asks for more settings than A1111 in the first few hours. Finish the checklist below, then the five numbered steps, before you open outside tutorials.

Preflight checklist (CUDA, cloud sync, disk)

Check What you want to see
CUDA / NVIDIA drivers (nvidia-smi) The card name, and a “CUDA Version” line
ComfyUI folder outside OneDrive, iCloud, and Google Drive Sync stays off models/ and output/
Free SSD space 50 GB and up for a first checkpoint (sizing the disk)
Test workflow Default SD 1.5, 512×512. Leave FLUX for a later day

Five steps, before the full ComfyUI guide

# Action Success looks like
1 NVIDIA drivers, then nvidia-smi Card name, and a “CUDA Version” line
2 A models/checkpoints folder outside cloud sync The path written down (for example D:\ComfyUI\models\checkpoints\)
3 A local .safetensors checkpoint, license read The file is there, with 50 GB free on the SSD
4 Start ComfyUI, then open http://127.0.0.1:8188 The web page loads
5 Default SD 1.5 workflow, 512×512, Queue Prompt An image in output/, with no VRAM error

Each step is spelled out below. After that: SDXL, FLUX, or workflows imported from the ComfyUI and FLUX guide.

Step 1: NVIDIA drivers and CUDA

CUDA (Compute Unified Device Architecture, NVIDIA’s layer that runs the math on the GPU) has to be supported by current graphics drivers. Quick check:

nvidia-smi

If the command prints your card’s name and a “CUDA Version” line, ComfyUI can see the GPU. Otherwise reinstall the driver from NVIDIA’s site, reboot, and reread the ComfyUI and FLUX guide. On AMD or Intel Arc, ComfyUI can still run. The FLUX workflows written up on LocAIhome mostly assume NVIDIA. See GPU, iGPU, and eGPU.

Step 2: a checkpoint on disk

  1. Create a local folder outside cloud sync (for example D:\ComfyUI\models\checkpoints\).
  2. Download an SD 1.5 or SDXL checkpoint from an official source or Hugging Face, and read the model’s license.
  3. Leave 50–100 GB free on the SSD (model storage).

Step 3: open the ComfyUI page

  1. Install ComfyUI using the local FLUX guide (Python, or a portable build, depending on your OS).
  2. Start the server. Open the browser at http://127.0.0.1:8188 (the usual local address).
  3. Load the default SD 1.5 workflow that ships with ComfyUI before you import heavy FLUX graphs.

Step 4: a first SD 1.5 workflow at 512 pixels

  1. In the interface, confirm the Load Checkpoint node points at your .safetensors file.
  2. Set the resolution to 512×512 for the first test (less VRAM).
  3. Click Queue Prompt and wait for the image in the output folder.
  4. If the image is black, or a VRAM error appears, use the “Common image errors” table further down.

FLUX and SD split like this: FLUX aims at a recent photographic look and often wants 12–16 GB of VRAM. SD 1.5 is the household classroom. See “Pick SD or FLUX” below.

VRAM and models

Model Indicative VRAM Household use
SD 1.5 6–8 GB Quick edits, first tests
SDXL 10–12 GB Posters, costume ideas. SDXL is the higher-resolution Stable Diffusion variant (typically 768–1024 px)
FLUX 12–16 GB A more recent photographic look

Household recommendation (October 2026): start with SD 1.5 at 512 px. Aim at FLUX when you have at least 12 GB of VRAM and a comfortable SSD. Otherwise stay on SDXL, or use an occasional online service only for visuals that contain no personal photos (see below).

If VRAM runs out, lower the resolution or use the optimized modes documented in ComfyUI.

Pick SD or FLUX for a first weekend

If you… Start with Why
Are new to this and have 8 GB of VRAM SD 1.5 A huge ecosystem, and light workflows
Want sharp posters at A3 size SDXL Higher definition, 10–12 GB of VRAM
Want a “recent photo” look FLUX Heavier (12–16 GB). See the FLUX guide
Only have a mini PC and no separate GPU Mini PC comparison Realistic image generation mostly wants a tower or a Mac set up for it

LocAIhome does not publish lab benchmarks per card. Rely on nvidia-smi, a 512 px test, then step the resolution up.

FLUX and SD, for a household

SD 1.5 / SDXL FLUX
VRAM 6–12 GB 12–16 GB
Community Large. LoRA (Low-Rank Adaptation: a small adapter that specializes an SD checkpoint without downloading it again) and lots of tutorials A growing ecosystem
Time to a first good image Simpler, via A1111 or the default SD 1.5 ComfyUI workflow Often more nodes, and more VRAM
Household use Posters, album variations A more recent photographic look

Ideas for the house (a real photo is optional)

The scenes below stay fully local. They stay off Midjourney and DALL·E unless you later choose those services yourself, and only for visuals that contain no personal data.

Idea Why privacy cares Hardware
Variations on a vacation album The service upload stays off SD 1.5 / SDXL, 8 GB VRAM
A child’s room poster from a sketch A real photo is optional SDXL, 10–12 GB
Mockups of a room before you buy furniture The plans stay home SD or FLUX, depending on the look you want

Rights in the photos. For album variations, inpainting, or room mockups, use your images, or content you have the rights to use. Photos found on the web, and other people’s faces used without agreement, stay out. A local pipeline sends nothing to an image cloud. A source photo you were not allowed to use is still a problem on your own disk.

A neutral example: a bedroom poster (text prompt only)

A local text-to-image workflow:

  1. Paste this prompt into the ComfyUI text node (SD 1.5, 512 px):
Wall poster for a child's room, soft watercolor, smiling farm animals, pastel blue and green palette, no readable text, no realistic human face, simple composition for an A3 frame.
  1. Negative prompt (the field meant for it):
realistic photo, child's face, text, logo, watermark, shocking content
  1. Generate. Change the colors in later tries. Leave an identifiable photo of a child out of the workflow.

Write the prompts with a local language model (Ollama)

This section is separate from the household ideas above. Here a language model writes the prompt text, and you paste it into ComfyUI. Ollama does not paint the pixels.

Use Ollama or Open WebUI to draft descriptions, then paste them into ComfyUI’s text node. Example instruction in the local chat: “Write an SD 1.5 prompt for a pastel children’s poster, with no realistic face, 80 words max.”

Cost and electricity

A GPU session draws more than a chat mini PC. Estimate it with the electricity budget, and compare the 3-year cost if you are deciding whether to buy an RTX.

Privacy

Pictures of children call for the strictest care, including where you share them. Read privacy when the AI stays home.

Common image errors

Problem Likely cause What to try
Black image Checkpoint and workflow disagree Start again from a default SD 1.5 workflow
Out of VRAM Resolution too high, or FLUX too heavy Drop to 512 px, or use SD 1.5
Photos on iCloud Cloud sync Exclude ComfyUI/output
Extremely slow CPU instead of GPU NVIDIA drivers, and nvidia-smi during a generation

Text and image, as two steps

Write the prompt in Open WebUI or Ollama, then paste it into ComfyUI. That separates the text model from the image model. If the household only makes a few pictures a month, the 3-year cost comparison helps you decide whether a dedicated RTX is worth it next to an occasional online subscription that never receives your albums.

In one page

Fully local images means ComfyUI (or A1111), checkpoints on an SSD, enough VRAM, folders outside cloud sync, and prompts tried first at 512 px. For the step-by-step install (FLUX workflows, VRAM troubleshooting), continue with the local ComfyUI and FLUX guide. That page is the install. This one is household use and the minimum hardware.

Before you buy a card, cross choosing hardware, the VRAM steps for image models, GPU, iGPU, or an OCuLink eGPU if you are building a tower or an external box, the 2026 mini PC comparison if you are choosing between a tower and a small box, and the hardware selector for the power bill and the noise when the tower runs on a weekend. Nearby software: Whisper sometimes shares the same GPU. Schedule it with ComfyUI.

For family chat, with the pixels left out, the private family assistant, Home Assistant and Ollama, and n8n with a local model are text or automation paths. Image generation stays with ComfyUI. Beginners can prove the basics with a first local AI in 15 minutes before stacking ComfyUI. As soon as faces or sensitive documents enter the pipeline, even with no image service involved, read privacy when the AI stays home and estimate the GPU session with the electricity budget. Store links (the RTX search, for example) are explained on the affiliate page.

When an online service still has a place

An occasional image subscription can sit beside a privacy habit if you keep family albums off it and only try public-domain or rights-free visuals. The moment you retouch photos of children or personal documents, the local ComfyUI and FLUX pipeline is the one to use.

An online service can be enough Prefer the machine at home
How often A few visuals a month, generic prompts Several sessions a week
How sensitive Nothing identifiable, and a child’s face Family albums, personal documents, faces
Hardware budget You have no GPU tower, and an RTX is off the shopping list An RTX or a Mac is already paid for, and the 3-year cost favors local
  • How often: an occasional online service for ideas that contain no personal data. Local, once you iterate on the same files.
  • How sensitive: US cloud image services (Midjourney, DALL·E) come with their own terms and hosting. See privacy when the AI stays home.
  • Budget: compare a renewing subscription with an RTX plus the kilowatt-hour budget, using the hardware selector (the kilowatt-hour price is editable).

Cross-check choosing hardware before you buy an RTX only for images.

Common questions

Do my personal photos stay in the house?

Yes, when the pipeline is fully local and you turn off system cloud sync for the working folder.

FLUX or Stable Diffusion?

FLUX is the newer look. Stable Diffusion has the larger ecosystem: tutorials, and LoRA files (small adapters that specialize a checkpoint without downloading the whole model again).

What hardware do I need at a minimum?

An NVIDIA GPU with 8 GB of VRAM is tight for SD 1.5. 12 GB and up is more comfortable for SDXL and FLUX. Use the image-model section of the VRAM guide (SD, SDXL, FLUX), which is separate from the chat tables in that same hardware article.

Do I need Ollama to make images?

The pixels come from the image model, and Ollama stays out of that step. Ollama can still help you write the text prompt beforehand.

Is CUDA required?

On NVIDIA, yes, for a speed you will actually wait for. Recent drivers turn CUDA on. Check with nvidia-smi. With no NVIDIA GPU, ComfyUI can run on the CPU, and SDXL or FLUX will feel very slow.

Can Ollama and ComfyUI share one machine?

Yes, when VRAM and RAM are enough. Skip a FLUX generation while a large language model is occupying the GPU. Schedule the two jobs, or put chat and images on two hosts.

Sources

Narrow it down with the hardware selector · French version of this section