diff --git a/README.md b/README.md
index bc5af8c..55e4f51 100644
--- a/README.md
+++ b/README.md
@@ -67,7 +67,7 @@ Update them any time with `sudo ./setup.sh configure`.
 |-------|---------|
 | `base` | `net-tools`, `ncdu`, `git`, `curl`, `wget`, `htop`, `tree`, `zip`/`unzip`, `ca-certificates`, `gnupg`, `jq`, `rsync`; `glow` (terminal markdown reader, Charm apt repo) |
 | `homelab` | `caddy`, `crowdsec`, `authelia`, `homeassistant`, `asterisk` |
-| `utilities` | `actualbudget`, `archivebox`, `changedetection`, `ddclient`, `filebrowser`, `fmd`, `gatus`, `homebox`, `joplin`, `koha`, `magicmirror`, `mail-archiver`, `mattermost`, `mealie`, `meshcentral`, `n8n`, `nextcloud`, `ntfy`, `onlyoffice`, `portainer`, `rustdesk`, `stirling-pdf`, `syncthing`, `traccar`, `unifi`, `uptimekuma`, `vaultwarden`, `watchyourlan`, `watchtower`, `wg-easy` |
+| `utilities` | `actualbudget`, `ai-gpu`, `archivebox`, `changedetection`, `ddclient`, `filebrowser`, `fmd`, `gatus`, `homebox`, `iopaint`, `joplin`, `koha`, `magicmirror`, `mail-archiver`, `mattermost`, `mealie`, `meshcentral`, `n8n`, `nextcloud`, `ntfy`, `onlyoffice`, `portainer`, `rustdesk`, `stirling-pdf`, `syncthing`, `traccar`, `unifi`, `uptimekuma`, `vaultwarden`, `watchyourlan`, `watchtower`, `wg-easy` |
 | `media` | `arm`, `audiobookshelf`, `calibre-web`, `emby`, `immich`, `jellyfin`, `lyrion` |
 | `cameras` | `frigate`, `frigate-audio`, `frigate-notify`, `sky-cam` |
 | `gaming` | `drum-rhythm-game`, `js99er`, `minecraft`, `wolf`, `wolf-pair` |
diff --git a/services/ai-gpu.sh b/services/ai-gpu.sh
new file mode 100644
index 0000000..edcb780
--- /dev/null
+++ b/services/ai-gpu.sh
@@ -0,0 +1,573 @@
+#!/bin/bash
+# services/ai-gpu.sh — GPU AI stack: InvokeAI image gen + Ollama/OpenWebUI LLM + swap portal.
+# Part of the modular post-install system (sourced by setup.sh).
+#
+# Can also be run standalone on any machine:
+#   sudo bash ai-gpu.sh
+# (Docker + nvidia-container-toolkit must already be installed)
+#
+# Clones https://github.com/outis1one/ai-6gb-gpu and installs three stacks:
+#   image-gen/  — InvokeAI (port 9090), nvidia GPU, optimised for 6 GB VRAM
+#   llm/        — Ollama (11434) + Open WebUI (3000) + SearXNG (internal)
+#   portal/     — Flask app (port 8080), mounts Docker socket, hot-swaps GPU between stacks
+#
+# Because a 6 GB GPU can only run ONE stack at a time, the portal handles the swap:
+# stop the active stack, start the requested one. Stop image-gen before starting llm, and vice versa.
+
+# ── Standalone bootstrap ──────────────────────────────────────────────────────
+if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
+    [[ "$(id -u)" == "0" ]] || { echo "Run with sudo: sudo bash $0"; exit 1; }
+
+    _SELF_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+    _COMMON="$_SELF_DIR/../lib/common.sh"
+
+    if [[ -f "$_COMMON" ]]; then
+        source "$_COMMON"
+    else
+        log_info()    { echo -e "\033[0;34m[INFO]\033[0m $*"; }
+        log_success() { echo -e "\033[0;32m[OK]\033[0m $*"; }
+        log_warning() { echo -e "\033[1;33m[WARN]\033[0m $*"; }
+        log_error()   { echo -e "\033[0;31m[ERROR]\033[0m $*" >&2; }
+
+        require_docker() {
+            command -v docker &>/dev/null || {
+                log_error "Docker not found. Install it first:"
+                log_error "  curl -fsSL https://get.docker.com | sudo sh"
+                return 1
+            }
+            docker compose version &>/dev/null || {
+                log_error "Docker Compose plugin missing:"
+                log_error "  sudo apt-get install -y docker-compose-plugin"
+                return 1
+            }
+        }
+
+        ensure_docker_dir_ownership() {
+            chown -R "$ACTUAL_USER:$ACTUAL_USER" "$@" 2>/dev/null || true
+        }
+
+        prompt_text() {
+            local _q="$1" _def="$2" _var="$3" _r
+            [[ "${UNATTENDED:-false}" == "true" ]] && { eval "$_var='$_def'"; return; }
+            read -r -p "  $_q " _r
+            eval "$_var='${_r:-$_def}'"
+        }
+
+        prompt_yn() {
+            local _q="$1" _def="$2" _var="$3" _r
+            [[ "${UNATTENDED:-false}" == "true" ]] && { eval "$_var='$_def'"; return; }
+            read -r -p "  $_q " _r
+            eval "$_var='${_r:-$_def}'"
+        }
+
+        configure_caddy_for_service() {
+            local _name="$1" _upstream="$2" _subdomain="$3" _extra="${4:-}"
+            local _caddy_dir="$DOCKER_DIR/caddy"
+            local _caddyfile="$_caddy_dir/Caddyfile"
+            local _display_port="${_upstream##*:}"
+
+            local _mode="none"
+            [[ -d "$_caddy_dir" ]] && _mode="local"
+            [[ -n "${CADDY_REMOTE_HOST:-}" ]] && [[ "$_mode" != "local" ]] && _mode="remote"
+            [[ "$_mode" == "none" ]] && {
+                log_info "Access $_name directly on port $_display_port."
+                return 0
+            }
+
+            echo ""
+            local _do_caddy=""
+            if [[ "$_mode" == "remote" ]]; then
+                log_info "Remote Caddy configured (${CADDY_REMOTE_HOST})."
+                log_info "A snippet file will be saved to ~/docker/caddy-snippets/."
+            fi
+            read -r -p "  Configure Caddy reverse proxy for $_name? [y/N]: " _do_caddy
+            [[ "${_do_caddy,,}" == "y" ]] || {
+                log_info "Skipping — access at: http://localhost:$_display_port"
+                return 0
+            }
+
+            local _default_domain=""
+            if [[ -n "${SITE_DOMAIN:-}" ]] && [[ "$SITE_DOMAIN" != "example.com" ]]; then
+                _default_domain="${_subdomain}.${SITE_DOMAIN}"
+                log_info "Default: $_default_domain"
+            fi
+            local _domain=""
+            read -r -p "  Domain [${_default_domain:-required}]: " _domain
+            _domain="${_domain:-$_default_domain}"
+            [[ -n "$_domain" ]] || { log_warning "No domain entered — skipping Caddy."; return 0; }
+
+            local _block_upstream="$_upstream"
+            if [[ "$_mode" == "remote" ]]; then
+                _block_upstream="${CADDY_REMOTE_HOST}:${_display_port}"
+            fi
+
+            local _site_block
+            _site_block="$(cat << CBLOCK
+
+# $_name
+${_domain} {
+    reverse_proxy ${_block_upstream}
+
+    header {
+        Strict-Transport-Security "max-age=31536000; includeSubDomains; preload"
+        X-Content-Type-Options "nosniff"
+        X-Frame-Options "SAMEORIGIN"
+        Referrer-Policy "strict-origin-when-cross-origin"
+    }
+
+    log {
+        output file /var/log/caddy/${_domain}.log
+        format json
+    }
+${_extra}
+}
+CBLOCK
+)"
+
+            if [[ "$_mode" == "local" ]]; then
+                if [[ -f "$_caddyfile" ]]; then
+                    local _bk="$_caddy_dir/Caddyfile.backup.$(date +%Y%m%d-%H%M%S)"
+                    cp "$_caddyfile" "$_bk"
+                    log_info "Backed up Caddyfile to $(basename "$_bk")"
+                else
+                    touch "$_caddyfile"
+                fi
+
+                if grep -q "^${_domain}" "$_caddyfile" 2>/dev/null; then
+                    log_warning "$_domain already in Caddyfile"
+                    local _ow=""
+                    read -r -p "  Overwrite? [y/N]: " _ow
+                    [[ "${_ow,,}" == "y" ]] || { log_info "Keeping existing entry."; return 0; }
+                    sed -i "/^${_domain}/,/^}/d" "$_caddyfile"
+                fi
+
+                printf '%s\n' "$_site_block" >> "$_caddyfile"
+                log_success "Added $_domain to Caddyfile"
+                docker exec caddy caddy fmt --overwrite /etc/caddy/Caddyfile 2>/dev/null || true
+                if docker exec caddy caddy reload --config /etc/caddy/Caddyfile 2>/dev/null; then
+                    log_success "$_name accessible at: https://$_domain"
+                else
+                    log_warning "Reload failed — check: docker logs caddy"
+                    log_info "Manual reload: docker exec caddy caddy reload --config /etc/caddy/Caddyfile"
+                fi
+            else
+                local _snippet_dir="$DOCKER_DIR/caddy-snippets"
+                local _snippet_file="$_snippet_dir/${_subdomain}.caddy"
+                mkdir -p "$_snippet_dir"
+                printf '%s\n' "$_site_block" > "$_snippet_file"
+                chown "$ACTUAL_USER:$ACTUAL_USER" "$_snippet_file" 2>/dev/null || true
+                log_success "Snippet saved: $_snippet_file"
+                log_info "Copy to Caddy machine:"
+                log_info "  scp $_snippet_file caddy-host:~/caddy-snippets/"
+                log_info "  rsync -av $_snippet_dir/ caddy-host:~/caddy-snippets/  (all at once)"
+            fi
+        }
+
+        write_readme() {
+            local _dir="$1"
+            mkdir -p "$_dir"
+            [[ "${DRY_RUN:-false}" == "true" ]] && return 0
+            cat > "$_dir/README.md"
+        }
+
+        generate_password() {
+            local _len="${1:-32}"
+            tr -dc 'A-Za-z0-9' < /dev/urandom | head -c "$_len"
+            echo
+        }
+    fi
+
+    ACTUAL_USER="${ACTUAL_USER:-${SUDO_USER:-$USER}}"
+    ACTUAL_HOME="$(getent passwd "$ACTUAL_USER" 2>/dev/null | cut -d: -f6 || echo "${HOME:-/root}")"
+    DOCKER_DIR="${DOCKER_DIR:-$ACTUAL_HOME/docker}"
+    DRY_RUN="${DRY_RUN:-false}"
+    UNATTENDED="${UNATTENDED:-false}"
+    SITE_TZ="${SITE_TZ:-$(cat /etc/timezone 2>/dev/null || echo UTC)}"
+    SITE_DOMAIN="${SITE_DOMAIN:-example.com}"
+    SITE_CADDY_NET="${SITE_CADDY_NET:-caddy_net}"
+    CADDY_REMOTE_HOST="${CADDY_REMOTE_HOST:-}"
+
+    register_service() { :; }
+    _RUN_STANDALONE=1
+fi
+# ─────────────────────────────────────────────────────────────────────────────
+
+register_service ai-gpu utilities "GPU AI stack — InvokeAI image gen + Ollama/OpenWebUI LLM (6 GB VRAM)" 9090
+
+install_ai_gpu() {
+    require_docker || return 1
+    log_info "Installing AI GPU stack (InvokeAI + Ollama/OpenWebUI + portal)..."
+    log_info "Requires: nvidia GPU with 6 GB+ VRAM, nvidia-container-toolkit installed."
+
+    local AI_DIR="$DOCKER_DIR/ai-gpu"
+    local REPO_URL="https://github.com/outis1one/ai-6gb-gpu.git"
+    local REPO_DIR="$AI_DIR/src"
+
+    if [ "$DRY_RUN" = true ]; then
+        echo "[DRY-RUN] Would clone $REPO_URL to $REPO_DIR"
+        echo "[DRY-RUN] Would create stacks: image-gen (InvokeAI:9090), llm (Ollama:11434 + OpenWebUI:3000), portal (8080)"
+        echo "[DRY-RUN] Would prompt for Ollama and InvokeAI model selection"
+        echo "[DRY-RUN] Would auto-pull selected Ollama models after LLM stack starts"
+        echo "[DRY-RUN] Would queue InvokeAI starter model via REST API"
+        return 0
+    fi
+
+    # ── Timezone ──────────────────────────────────────────────────────────────
+    local TZ_VAL="${SITE_TZ:-UTC}"
+    prompt_text "Timezone (e.g. America/New_York) [$TZ_VAL]:" "$TZ_VAL" TZ_VAL
+    TZ_VAL="${TZ_VAL:-UTC}"
+
+    # ── Ollama model selection ────────────────────────────────────────────────
+    echo ""
+    log_info "Ollama LLM models — select which to download (enter numbers separated by spaces):"
+    log_info "All models are quantized (Q4_K_M) and run comfortably on 6 GB VRAM."
+    echo ""
+    log_info "  1) llama3.2:3b       ~2.0 GB  Fast general chat. Great all-rounder.  ← Recommended"
+    log_info "  2) llama3.2:1b       ~1.3 GB  Ultra-fast. Light tasks, low latency."
+    log_info "  3) qwen2.5:7b        ~4.7 GB  Top code + math model. Strong reasoning."
+    log_info "  4) mistral:7b        ~4.1 GB  Solid all-rounder. Good at instruction follow."
+    log_info "  5) phi4-mini         ~2.5 GB  Microsoft Phi-4 mini. Excellent for coding."
+    log_info "  6) gemma3:4b         ~2.5 GB  Google Gemma 3. Well-rounded, multilingual."
+    log_info "  7) deepseek-r1:7b    ~4.7 GB  Strong reasoning and math. Think-step model."
+    log_info "  8) nomic-embed-text  ~274 MB  Embedding model — enables RAG/doc search."
+    log_info "                                 Recommended to add alongside a chat model."
+    echo ""
+    log_info "  Example: '1 8' pulls llama3.2:3b + nomic-embed-text"
+    log_info "  Enter '0' or leave blank to skip and pull models manually later."
+    echo ""
+
+    local OLLAMA_CHOICES=""
+    prompt_text "Models to download [1 8]:" "1 8" OLLAMA_CHOICES
+
+    declare -a OLLAMA_MODELS=()
+    for _n in $OLLAMA_CHOICES; do
+        case "$_n" in
+            1) OLLAMA_MODELS+=("llama3.2:3b") ;;
+            2) OLLAMA_MODELS+=("llama3.2:1b") ;;
+            3) OLLAMA_MODELS+=("qwen2.5:7b") ;;
+            4) OLLAMA_MODELS+=("mistral:7b") ;;
+            5) OLLAMA_MODELS+=("phi4-mini") ;;
+            6) OLLAMA_MODELS+=("gemma3:4b") ;;
+            7) OLLAMA_MODELS+=("deepseek-r1:7b") ;;
+            8) OLLAMA_MODELS+=("nomic-embed-text") ;;
+        esac
+    done
+
+    # ── InvokeAI model selection ──────────────────────────────────────────────
+    echo ""
+    log_info "InvokeAI image generation models (for 6 GB VRAM with partial GPU offload):"
+    log_info "InvokeAI uses VRAM=3 GB + 8 GB RAM cache, so all models below work on 6 GB."
+    echo ""
+    log_info "  1) stabilityai/stable-diffusion-v1-5   ~4 GB  SD 1.5 — fast, huge style/LoRA library."
+    log_info "                                                  Best starting model for most uses."
+    log_info "  2) stabilityai/sdxl-turbo              ~7 GB  SDXL Turbo — 4-step generation."
+    log_info "                                                  Fast, high quality, slightly slower on 6 GB."
+    log_info "  3) stabilityai/stable-diffusion-xl-base-1.0"
+    log_info "                                         ~7 GB  SDXL base — best quality at 1024px."
+    log_info "                                                  Slowest due to RAM offload on 6 GB."
+    log_info "  4) Skip — install models via the Model Manager at http://localhost:9090"
+    echo ""
+    log_info "  Tip: SD 1.5 (choice 1) is fastest and most compatible. Start here."
+    log_info "  HuggingFace token: required for some gated models (free at huggingface.co/settings/tokens)"
+    echo ""
+
+    local INVOKE_CHOICE=""
+    prompt_text "InvokeAI starter model [1]:" "1" INVOKE_CHOICE
+
+    local INVOKE_MODEL_SOURCE=""
+    local INVOKE_MODEL_NAME=""
+    case "$INVOKE_CHOICE" in
+        2) INVOKE_MODEL_SOURCE="stabilityai/sdxl-turbo"
+           INVOKE_MODEL_NAME="SDXL Turbo" ;;
+        3) INVOKE_MODEL_SOURCE="stabilityai/stable-diffusion-xl-base-1.0"
+           INVOKE_MODEL_NAME="SDXL Base" ;;
+        4) INVOKE_MODEL_SOURCE=""
+           INVOKE_MODEL_NAME="" ;;
+        *) INVOKE_MODEL_SOURCE="stabilityai/stable-diffusion-v1-5"
+           INVOKE_MODEL_NAME="SD 1.5" ;;
+    esac
+
+    local HF_TOKEN=""
+    if [ -n "$INVOKE_MODEL_SOURCE" ]; then
+        prompt_text "HuggingFace token (optional — needed for gated models, enter to skip):" "" HF_TOKEN
+    fi
+
+    # ── Clone / update repo ───────────────────────────────────────────────────
+    mkdir -p "$AI_DIR"
+    if [ -d "$REPO_DIR/.git" ]; then
+        log_info "Updating ai-6gb-gpu repo..."
+        git -C "$REPO_DIR" pull --ff-only 2>/dev/null \
+            && log_success "Repo updated" \
+            || log_warning "Could not pull latest — using existing version"
+    else
+        log_info "Cloning ai-6gb-gpu repo..."
+        git clone --depth 1 "$REPO_URL" "$REPO_DIR" \
+            || { log_error "Clone failed — check network and git access"; return 1; }
+    fi
+
+    # ── Image-gen stack (InvokeAI) ────────────────────────────────────────────
+    local IMAGE_GEN_DIR="$AI_DIR/image-gen"
+    mkdir -p "$IMAGE_GEN_DIR"
+    if [ -d "$REPO_DIR/ai-image-gen" ]; then
+        cp -rn "$REPO_DIR/ai-image-gen/." "$IMAGE_GEN_DIR/" 2>/dev/null || true
+        find "$IMAGE_GEN_DIR" -name "docker-compose.yml" -exec \
+            sed -i "s|America/New_York|$TZ_VAL|g" {} \;
+    fi
+
+    cat > "$IMAGE_GEN_DIR/.env" << IMGENV
+# InvokeAI — image generation
+TZ=${TZ_VAL}
+# VRAM cap: 3 GB leaves headroom on a 6 GB card; remaining model layers go to RAM
+INVOKEAI_vram=3
+# RAM cache size for model layer offload
+INVOKEAI_ram=8
+${HF_TOKEN:+HUGGING_FACE_HUB_TOKEN=${HF_TOKEN}}
+CADDY_NET=${SITE_CADDY_NET}
+IMGENV
+    chmod 600 "$IMAGE_GEN_DIR/.env"
+
+    # ── LLM stack (Ollama + Open WebUI + SearXNG) ─────────────────────────────
+    local LLM_DIR="$AI_DIR/llm"
+    mkdir -p "$LLM_DIR"
+    if [ -d "$REPO_DIR/ai-llm" ]; then
+        cp -rn "$REPO_DIR/ai-llm/." "$LLM_DIR/" 2>/dev/null || true
+        find "$LLM_DIR" -name "docker-compose.yml" -exec \
+            sed -i "s|America/New_York|$TZ_VAL|g" {} \;
+    fi
+
+    local WEBUI_SECRET
+    WEBUI_SECRET="$(generate_password 32)"
+
+    cat > "$LLM_DIR/.env" << LLMENV
+# Ollama + Open WebUI + SearXNG
+TZ=${TZ_VAL}
+WEBUI_SECRET_KEY=${WEBUI_SECRET}
+# Open WebUI: enable SearXNG for web search in chats
+ENABLE_RAG_WEB_SEARCH=true
+RAG_WEB_SEARCH_ENGINE=searxng
+SEARXNG_QUERY_URL=http://searxng:8080/search?q=<query>&format=json
+CADDY_NET=${SITE_CADDY_NET}
+LLMENV
+    chmod 600 "$LLM_DIR/.env"
+
+    # ── Portal stack (Flask GPU swap controller) ──────────────────────────────
+    local PORTAL_DIR="$AI_DIR/portal"
+    mkdir -p "$PORTAL_DIR"
+    if [ -d "$REPO_DIR/ai-portal" ]; then
+        cp -rn "$REPO_DIR/ai-portal/." "$PORTAL_DIR/" 2>/dev/null || true
+        if [ -f "$PORTAL_DIR/docker-compose.yml" ]; then
+            sed -i \
+                "s|/home/[^/]*/docker:|${ACTUAL_HOME}/docker:|g" \
+                "$PORTAL_DIR/docker-compose.yml"
+            sed -i "s|America/New_York|$TZ_VAL|g" "$PORTAL_DIR/docker-compose.yml"
+        fi
+    fi
+
+    cat > "$PORTAL_DIR/.env" << PORTALENV
+# AI Portal — GPU stack swap controller
+TZ=${TZ_VAL}
+# Paths inside the container (/docker maps to ${ACTUAL_HOME}/docker via volume mount)
+IMAGE_STACK=/docker/ai-gpu/image-gen
+LLM_STACK=/docker/ai-gpu/llm
+CADDY_NET=${SITE_CADDY_NET}
+PORTALENV
+    chmod 600 "$PORTAL_DIR/.env"
+
+    ensure_docker_dir_ownership "$AI_DIR"
+
+    echo ""
+    log_success "AI GPU stacks configured under $AI_DIR"
+    log_warning "Only ONE GPU stack can run at a time on a 6 GB card."
+    log_info "Use the portal (port 8080) to hot-swap between image-gen and llm."
+    echo ""
+
+    # ── Caddy ─────────────────────────────────────────────────────────────────
+    configure_caddy_for_service "AI Portal" "ai-portal:8080" "localai"
+    configure_caddy_for_service "InvokeAI" "invokeai:9090" "images"
+
+    # ── README ────────────────────────────────────────────────────────────────
+    write_readme "$AI_DIR" << MD
+# AI GPU Stack
+
+Three Docker stacks optimised for a 6 GB VRAM nvidia GPU.
+Source: https://github.com/outis1one/ai-6gb-gpu
+
+## Stacks
+
+| Stack | Service | Port | Notes |
+|-------|---------|------|-------|
+| \`image-gen/\` | InvokeAI | 9090 | Image generation (SD 1.5, SDXL, Flux…) |
+| \`llm/\` | Ollama | 11434 | LLM inference engine |
+| \`llm/\` | Open WebUI | 3000 | Chat UI — models, RAG, web search |
+| \`llm/\` | SearXNG | internal | Web search backend for RAG in OpenWebUI |
+| \`portal/\` | AI Portal | 8080 | GPU swap controller — start/stop stacks |
+
+## GPU time-sharing (important)
+
+A 6 GB GPU can only run one AI stack at a time.
+Use the portal at http://localhost:8080 to swap.
+
+Manual swap:
+\`\`\`bash
+docker compose -f $AI_DIR/llm/docker-compose.yml down
+docker compose -f $AI_DIR/image-gen/docker-compose.yml up -d
+\`\`\`
+
+## InvokeAI — add more models
+
+Models installed at setup are in the Model Manager. To add more:
+1. Open http://localhost:9090 → Model Manager → Add Model
+2. Paste a HuggingFace repo ID (e.g. \`stabilityai/stable-diffusion-2-1\`)
+3. Or import a local .safetensors file
+
+For gated models, add \`HUGGING_FACE_HUB_TOKEN=xxx\` to \`image-gen/.env\`.
+
+Recommended models for 6 GB (with partial GPU offload):
+
+| Model | Source | Notes |
+|-------|--------|-------|
+| SD 1.5 | \`stabilityai/stable-diffusion-v1-5\` | Fast, huge LoRA library |
+| SDXL Turbo | \`stabilityai/sdxl-turbo\` | 4-step, good quality |
+| SDXL Base | \`stabilityai/stable-diffusion-xl-base-1.0\` | Best quality, slower |
+| SD 2.1 | \`stabilityai/stable-diffusion-2-1\` | Good mid-size choice |
+
+## Ollama — add more models
+
+\`\`\`bash
+# Pull any model while llm stack is running
+docker exec ollama ollama pull llama3.2:3b
+docker exec ollama ollama pull nomic-embed-text    # RAG embeddings
+docker exec ollama ollama list                      # see installed models
+\`\`\`
+
+Browse models at: https://ollama.com/library
+For 6 GB cards, stick to 7B or smaller with Q4_K_M quantisation (~4.5 GB).
+
+## Open WebUI — first login
+
+Open http://localhost:3000 and create your admin account on first visit.
+Models pulled into Ollama appear automatically in the model dropdown.
+Enable web search: Settings → Admin → Web Search (SearXNG is pre-configured).
+
+## Manage individual stacks
+\`\`\`bash
+cd $AI_DIR/image-gen && docker compose up -d    # start InvokeAI
+cd $AI_DIR/llm       && docker compose up -d    # start Ollama + OpenWebUI
+cd $AI_DIR/portal    && docker compose up -d    # start portal
+docker compose -f $AI_DIR/image-gen/docker-compose.yml logs -f invokeai
+docker compose -f $AI_DIR/llm/docker-compose.yml       logs -f openwebui
+\`\`\`
+
+## Update
+\`\`\`bash
+cd $REPO_DIR && git pull
+cd $AI_DIR/image-gen && docker compose pull && docker compose up -d
+cd $AI_DIR/llm       && docker compose pull && docker compose up -d
+cd $AI_DIR/portal    && docker compose build --pull && docker compose up -d
+\`\`\`
+MD
+
+    # ── Start stacks + pull models ────────────────────────────────────────────
+    echo ""
+    log_info "What would you like to start now?"
+    log_info "  1) Portal only           — start the swap controller, configure the rest later"
+    log_info "  2) Portal + LLM stack    — start Ollama/OpenWebUI and pull selected models"
+    log_info "  3) Portal + image-gen    — start InvokeAI and queue the starter model download"
+    log_info "  4) None                  — start manually later"
+    echo ""
+
+    local START_CHOICE=""
+    prompt_text "Choice [2]:" "2" START_CHOICE
+
+    # Always start portal if any stack is starting
+    if [[ "$START_CHOICE" =~ ^[123]$ ]]; then
+        docker compose -f "$PORTAL_DIR/docker-compose.yml" up -d \
+            && log_success "Portal started — http://localhost:8080" \
+            || log_warning "Portal start failed — check: docker compose -f $PORTAL_DIR/docker-compose.yml logs"
+    fi
+
+    if [[ "$START_CHOICE" == "2" ]]; then
+        # Start LLM stack
+        docker compose -f "$LLM_DIR/docker-compose.yml" up -d \
+            && log_success "LLM stack started" \
+            || { log_warning "LLM stack start failed"; START_CHOICE="0"; }
+
+        # Pull Ollama models if any were selected
+        if [ ${#OLLAMA_MODELS[@]} -gt 0 ] && [[ "$START_CHOICE" == "2" ]]; then
+            log_info "Waiting for Ollama to be ready..."
+            local _w=0
+            while ! curl -sf "http://localhost:11434/api/version" &>/dev/null; do
+                sleep 3; _w=$((_w+3))
+                [[ $_w -ge 90 ]] && { log_warning "Ollama not responding after 90s — pull models manually later"; break; }
+            done
+
+            if curl -sf "http://localhost:11434/api/version" &>/dev/null; then
+                for _m in "${OLLAMA_MODELS[@]}"; do
+                    log_info "Pulling $_m (this may take a while)..."
+                    docker exec ollama ollama pull "$_m" \
+                        && log_success "$_m ready" \
+                        || log_warning "Pull failed for $_m — retry: docker exec ollama ollama pull $_m"
+                done
+                log_success "Open WebUI ready at: http://localhost:3000"
+                log_info "Create your admin account on the first visit."
+            fi
+        fi
+    fi
+
+    if [[ "$START_CHOICE" == "3" ]]; then
+        # Start image-gen stack
+        docker compose -f "$IMAGE_GEN_DIR/docker-compose.yml" up -d \
+            && log_success "InvokeAI started" \
+            || { log_warning "InvokeAI start failed — check: docker compose -f $IMAGE_GEN_DIR/docker-compose.yml logs"; START_CHOICE="0"; }
+
+        # Queue starter model via InvokeAI REST API
+        if [ -n "$INVOKE_MODEL_SOURCE" ] && [[ "$START_CHOICE" == "3" ]]; then
+            log_info "Waiting for InvokeAI to be ready (model database initialises on first start)..."
+            local _w=0
+            while ! curl -sf "http://localhost:9090/api/v1/app/version" &>/dev/null; do
+                sleep 5; _w=$((_w+5))
+                [[ $_w -ge 180 ]] && { log_warning "InvokeAI not responding after 3 min"; break; }
+            done
+
+            if curl -sf "http://localhost:9090/api/v1/app/version" &>/dev/null; then
+                log_info "Queuing $INVOKE_MODEL_NAME download..."
+                local _resp
+                _resp=$(curl -s -X POST "http://localhost:9090/api/v2/models/install" \
+                    -H "Content-Type: application/json" \
+                    -d "{\"source\": \"${INVOKE_MODEL_SOURCE}\"}" 2>/dev/null)
+                if echo "$_resp" | grep -q '"id"'; then
+                    log_success "$INVOKE_MODEL_NAME queued — downloading in background"
+                    log_info "Track progress: http://localhost:9090 → Model Manager → In Progress"
+                else
+                    log_warning "Could not queue via API. Install manually:"
+                    log_info "  Open http://localhost:9090 → Model Manager → Add Model"
+                    log_info "  Source: $INVOKE_MODEL_SOURCE"
+                fi
+            fi
+        fi
+    fi
+
+    if [[ "$START_CHOICE" == "4" ]] || [[ "$START_CHOICE" == "0" ]]; then
+        log_info "Start when ready:"
+        log_info "  docker compose -f $PORTAL_DIR/docker-compose.yml up -d"
+        log_info "  docker compose -f $LLM_DIR/docker-compose.yml up -d"
+        log_info "  docker compose -f $IMAGE_GEN_DIR/docker-compose.yml up -d"
+    fi
+
+    echo ""
+    echo "  Portal:       http://localhost:8080  (GPU swap controller)"
+    echo "  InvokeAI:     http://localhost:9090  (image-gen stack)"
+    echo "  Open WebUI:   http://localhost:3000  (llm stack)"
+    echo "  Ollama API:   http://localhost:11434 (llm stack)"
+    echo "  Source repo:  $REPO_DIR"
+    echo ""
+    if [ ${#OLLAMA_MODELS[@]} -gt 0 ]; then
+        echo "  Ollama models queued: ${OLLAMA_MODELS[*]}"
+    fi
+    if [ -n "$INVOKE_MODEL_NAME" ]; then
+        echo "  InvokeAI starter:     $INVOKE_MODEL_NAME ($INVOKE_MODEL_SOURCE)"
+    fi
+    echo ""
+}
+
+[[ "${_RUN_STANDALONE:-0}" == 1 ]] && install_ai_gpu
diff --git a/services/iopaint.sh b/services/iopaint.sh
new file mode 100644
index 0000000..b457572
--- /dev/null
+++ b/services/iopaint.sh
@@ -0,0 +1,475 @@
+#!/bin/bash
+# services/iopaint.sh — AI image editing: erase objects, fill regions, replace with text prompt.
+# Part of the modular post-install system (sourced by setup.sh).
+#
+# Can also be run standalone on any machine:
+#   sudo bash iopaint.sh
+# (Docker must already be installed when run standalone)
+#
+# Three use cases:
+#   Erase/remove  — mask an object, AI fills the gap (LaMa, CPU-safe)
+#   Inpaint/fill  — restore damaged areas, remove watermarks (multiple models)
+#   Replace       — mask + text prompt → AI draws new content (PowerPaint, GPU required)
+#
+# IOPaint is local-only: it cannot call a remote GPU or InvokeAI on another machine.
+# For text-guided replacement the GPU must be on this same machine.
+# IOPaint has no built-in auth — protect with Authelia via Caddy.
+
+# ── Standalone bootstrap ──────────────────────────────────────────────────────
+if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
+    [[ "$(id -u)" == "0" ]] || { echo "Run with sudo: sudo bash $0"; exit 1; }
+
+    _SELF_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+    _COMMON="$_SELF_DIR/../lib/common.sh"
+
+    if [[ -f "$_COMMON" ]]; then
+        source "$_COMMON"
+    else
+        log_info()    { echo -e "\033[0;34m[INFO]\033[0m $*"; }
+        log_success() { echo -e "\033[0;32m[OK]\033[0m $*"; }
+        log_warning() { echo -e "\033[1;33m[WARN]\033[0m $*"; }
+        log_error()   { echo -e "\033[0;31m[ERROR]\033[0m $*" >&2; }
+
+        require_docker() {
+            command -v docker &>/dev/null || {
+                log_error "Docker not found. Install it first:"
+                log_error "  curl -fsSL https://get.docker.com | sudo sh"
+                return 1
+            }
+            docker compose version &>/dev/null || {
+                log_error "Docker Compose plugin missing:"
+                log_error "  sudo apt-get install -y docker-compose-plugin"
+                return 1
+            }
+        }
+
+        ensure_docker_dir_ownership() {
+            chown -R "$ACTUAL_USER:$ACTUAL_USER" "$@" 2>/dev/null || true
+        }
+
+        prompt_text() {
+            local _q="$1" _def="$2" _var="$3" _r
+            [[ "${UNATTENDED:-false}" == "true" ]] && { eval "$_var='$_def'"; return; }
+            read -r -p "  $_q " _r
+            eval "$_var='${_r:-$_def}'"
+        }
+
+        prompt_yn() {
+            local _q="$1" _def="$2" _var="$3" _r
+            [[ "${UNATTENDED:-false}" == "true" ]] && { eval "$_var='$_def'"; return; }
+            read -r -p "  $_q " _r
+            eval "$_var='${_r:-$_def}'"
+        }
+
+        configure_caddy_for_service() {
+            local _name="$1" _upstream="$2" _subdomain="$3" _extra="${4:-}"
+            local _caddy_dir="$DOCKER_DIR/caddy"
+            local _caddyfile="$_caddy_dir/Caddyfile"
+            local _display_port="${_upstream##*:}"
+
+            local _mode="none"
+            [[ -d "$_caddy_dir" ]] && _mode="local"
+            [[ -n "${CADDY_REMOTE_HOST:-}" ]] && [[ "$_mode" != "local" ]] && _mode="remote"
+            [[ "$_mode" == "none" ]] && {
+                log_info "Access $_name directly on port $_display_port."
+                return 0
+            }
+
+            echo ""
+            local _do_caddy=""
+            if [[ "$_mode" == "remote" ]]; then
+                log_info "Remote Caddy configured (${CADDY_REMOTE_HOST})."
+                log_info "A snippet file will be saved to ~/docker/caddy-snippets/."
+            fi
+            read -r -p "  Configure Caddy reverse proxy for $_name? [y/N]: " _do_caddy
+            [[ "${_do_caddy,,}" == "y" ]] || {
+                log_info "Skipping — access at: http://localhost:$_display_port"
+                return 0
+            }
+
+            local _default_domain=""
+            if [[ -n "${SITE_DOMAIN:-}" ]] && [[ "$SITE_DOMAIN" != "example.com" ]]; then
+                _default_domain="${_subdomain}.${SITE_DOMAIN}"
+                log_info "Default: $_default_domain"
+            fi
+            local _domain=""
+            read -r -p "  Domain [${_default_domain:-required}]: " _domain
+            _domain="${_domain:-$_default_domain}"
+            [[ -n "$_domain" ]] || { log_warning "No domain entered — skipping Caddy."; return 0; }
+
+            local _block_upstream="$_upstream"
+            if [[ "$_mode" == "remote" ]]; then
+                _block_upstream="${CADDY_REMOTE_HOST}:${_display_port}"
+            fi
+
+            local _site_block
+            _site_block="$(cat << CBLOCK
+
+# $_name
+${_domain} {
+    reverse_proxy ${_block_upstream}
+
+    header {
+        Strict-Transport-Security "max-age=31536000; includeSubDomains; preload"
+        X-Content-Type-Options "nosniff"
+        X-Frame-Options "SAMEORIGIN"
+        Referrer-Policy "strict-origin-when-cross-origin"
+    }
+
+    log {
+        output file /var/log/caddy/${_domain}.log
+        format json
+    }
+${_extra}
+}
+CBLOCK
+)"
+
+            if [[ "$_mode" == "local" ]]; then
+                if [[ -f "$_caddyfile" ]]; then
+                    local _bk="$_caddy_dir/Caddyfile.backup.$(date +%Y%m%d-%H%M%S)"
+                    cp "$_caddyfile" "$_bk"
+                    log_info "Backed up Caddyfile to $(basename "$_bk")"
+                else
+                    touch "$_caddyfile"
+                fi
+
+                if grep -q "^${_domain}" "$_caddyfile" 2>/dev/null; then
+                    log_warning "$_domain already in Caddyfile"
+                    local _ow=""
+                    read -r -p "  Overwrite? [y/N]: " _ow
+                    [[ "${_ow,,}" == "y" ]] || { log_info "Keeping existing entry."; return 0; }
+                    sed -i "/^${_domain}/,/^}/d" "$_caddyfile"
+                fi
+
+                printf '%s\n' "$_site_block" >> "$_caddyfile"
+                log_success "Added $_domain to Caddyfile"
+                docker exec caddy caddy fmt --overwrite /etc/caddy/Caddyfile 2>/dev/null || true
+                if docker exec caddy caddy reload --config /etc/caddy/Caddyfile 2>/dev/null; then
+                    log_success "$_name accessible at: https://$_domain"
+                else
+                    log_warning "Reload failed — check: docker logs caddy"
+                    log_info "Manual reload: docker exec caddy caddy reload --config /etc/caddy/Caddyfile"
+                fi
+            else
+                local _snippet_dir="$DOCKER_DIR/caddy-snippets"
+                local _snippet_file="$_snippet_dir/${_subdomain}.caddy"
+                mkdir -p "$_snippet_dir"
+                printf '%s\n' "$_site_block" > "$_snippet_file"
+                chown "$ACTUAL_USER:$ACTUAL_USER" "$_snippet_file" 2>/dev/null || true
+                log_success "Snippet saved: $_snippet_file"
+                log_info "Copy to Caddy machine:"
+                log_info "  scp $_snippet_file caddy-host:~/caddy-snippets/"
+                log_info "  rsync -av $_snippet_dir/ caddy-host:~/caddy-snippets/  (all at once)"
+            fi
+        }
+
+        write_readme() {
+            local _dir="$1"
+            mkdir -p "$_dir"
+            [[ "${DRY_RUN:-false}" == "true" ]] && return 0
+            cat > "$_dir/README.md"
+        }
+    fi
+
+    ACTUAL_USER="${ACTUAL_USER:-${SUDO_USER:-$USER}}"
+    ACTUAL_HOME="$(getent passwd "$ACTUAL_USER" 2>/dev/null | cut -d: -f6 || echo "${HOME:-/root}")"
+    DOCKER_DIR="${DOCKER_DIR:-$ACTUAL_HOME/docker}"
+    DRY_RUN="${DRY_RUN:-false}"
+    UNATTENDED="${UNATTENDED:-false}"
+    SITE_TZ="${SITE_TZ:-$(cat /etc/timezone 2>/dev/null || echo UTC)}"
+    SITE_DOMAIN="${SITE_DOMAIN:-example.com}"
+    SITE_CADDY_NET="${SITE_CADDY_NET:-caddy_net}"
+    CADDY_REMOTE_HOST="${CADDY_REMOTE_HOST:-}"
+
+    register_service() { :; }
+    _RUN_STANDALONE=1
+fi
+# ─────────────────────────────────────────────────────────────────────────────
+
+register_service iopaint utilities "AI image inpainting — erase objects, fill, restore (IOPaint)" 8100
+
+install_iopaint() {
+    require_docker || return 1
+    log_info "Installing IOPaint..."
+
+    local IOPAINT_DIR="$DOCKER_DIR/iopaint"
+
+    if [ "$DRY_RUN" = true ]; then
+        echo "[DRY-RUN] Would create $IOPAINT_DIR"
+        echo "[DRY-RUN] Would prompt for model and GPU (CUDA) support"
+        echo "[DRY-RUN] Would write docker-compose.yml and .env"
+        echo "[DRY-RUN] Would offer Authelia SSO"
+        return 0
+    fi
+
+    mkdir -p "$IOPAINT_DIR"
+    ensure_docker_dir_ownership "$IOPAINT_DIR"
+    cd "$IOPAINT_DIR" || return 1
+
+    # ── GPU ──────────────────────────────────────────────────────────────────
+    log_info "IOPaint can:"
+    log_info "  • Erase / remove  — mask an object, AI fills the gap (works CPU-only)"
+    log_info "  • Inpaint / fill  — restore damaged areas, remove watermarks"
+    log_info "  • Replace         — mask + type what goes there → AI draws it (needs GPU)"
+    log_info "  Note: inference is always local — IOPaint cannot use a GPU on another machine."
+    echo ""
+    local USE_GPU=""
+    prompt_yn "Enable CUDA GPU support? Requires nvidia-container-toolkit (y/n):" "n" USE_GPU
+
+    # ── Model selection ───────────────────────────────────────────────────────
+    echo ""
+    log_info "Select default model (you can switch models in the UI without restarting):"
+    echo ""
+    log_info "  ── CPU-safe — work on any machine ─────────────────────────────────────────"
+    log_info "  1) lama                    Erase / removal. Intelligent gap fill. ~200 MB  ← Recommended"
+    log_info "  2) cv2                     OpenCV fill. No download, instant. Rough quality."
+    log_info "  3) zits                    Portrait & face restoration. ~200 MB."
+    log_info "  4) manga                   Comic/manga text bubble removal. ~100 MB."
+    echo ""
+    if [[ "$USE_GPU" =~ ^[Yy]$ ]]; then
+        log_info "  ── GPU-accelerated fill ───────────────────────────────────────────────────"
+        log_info "  5) migan                   MiGAN: fast GPU inpainting. ~50 MB."
+        log_info "  6) fcf                     FcF: high-quality contextual fill. ~600 MB."
+        log_info "  7) mat                     MAT: large missing region fill. ~300 MB."
+        log_info "  8) ldm                     LDM: latent diffusion texture fill. ~1.2 GB."
+        echo ""
+        log_info "  ── Text-guided REPLACEMENT (GPU + Stable Diffusion) ───────────────────────"
+        log_info "  9) Sanster/PowerPaint-V2-filling"
+        log_info "                             Mask + type 'a red barn' → AI draws it. ~4 GB."
+        log_info "                             Modes: text-guided, shape-guided, erase, outpaint."
+        log_info " 10) runwayml/stable-diffusion-inpainting"
+        log_info "                             Classic SD 1.5 inpaint. Huge LoRA/style library. ~4 GB."
+        echo ""
+    fi
+
+    local MODEL_NUM=""
+    prompt_text "Model choice [1=lama]:" "1" MODEL_NUM
+
+    local IOPAINT_MODEL="lama"
+    case "$MODEL_NUM" in
+        2)  IOPAINT_MODEL="cv2" ;;
+        3)  IOPAINT_MODEL="zits" ;;
+        4)  IOPAINT_MODEL="manga" ;;
+        5)  IOPAINT_MODEL="migan" ;;
+        6)  IOPAINT_MODEL="fcf" ;;
+        7)  IOPAINT_MODEL="mat" ;;
+        8)  IOPAINT_MODEL="ldm" ;;
+        9)  IOPAINT_MODEL="Sanster/PowerPaint-V2-filling" ;;
+        10) IOPAINT_MODEL="runwayml/stable-diffusion-inpainting" ;;
+        *)  IOPAINT_MODEL="lama" ;;
+    esac
+
+    local DEVICE_VAL="cpu"
+    [[ "$USE_GPU" =~ ^[Yy]$ ]] && DEVICE_VAL="cuda"
+
+    # ── docker-compose.yml ────────────────────────────────────────────────────
+    # MODEL and DEVICE come from .env — change them there and restart to switch.
+    # Volume ./models:/root/.cache persists ALL model caches:
+    #   /root/.cache/torch/hub/checkpoints/  (LaMa, CV2, ZITS, etc.)
+    #   /root/.cache/huggingface/             (SD, PowerPaint, LDM, etc.)
+    if [[ "$USE_GPU" =~ ^[Yy]$ ]]; then
+        cat > docker-compose.yml << 'IOPAINT_GPU'
+name: iopaint
+
+services:
+  iopaint:
+    image: cwq1913/iopaint:latest
+    container_name: iopaint
+    hostname: iopaint
+    restart: unless-stopped
+    command: >-
+      iopaint start
+      --model=${MODEL:-lama}
+      --device=${DEVICE:-cuda}
+      --port=8080
+      --host=0.0.0.0
+    ports:
+      - "8100:8080"
+    env_file: .env
+    volumes:
+      - ./models:/root/.cache
+      - ./input:/app/input
+      - ./output:/app/output
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    networks:
+      - caddy_net
+
+networks:
+  caddy_net:
+    external: true
+    name: ${CADDY_NET:-caddy_net}
+IOPAINT_GPU
+    else
+        cat > docker-compose.yml << 'IOPAINT_CPU'
+name: iopaint
+
+services:
+  iopaint:
+    image: cwq1913/iopaint:latest
+    container_name: iopaint
+    hostname: iopaint
+    restart: unless-stopped
+    command: >-
+      iopaint start
+      --model=${MODEL:-lama}
+      --device=${DEVICE:-cpu}
+      --port=8080
+      --host=0.0.0.0
+    ports:
+      - "8100:8080"
+    env_file: .env
+    volumes:
+      - ./models:/root/.cache
+      - ./input:/app/input
+      - ./output:/app/output
+    networks:
+      - caddy_net
+
+networks:
+  caddy_net:
+    external: true
+    name: ${CADDY_NET:-caddy_net}
+IOPAINT_CPU
+    fi
+
+    # ── .env ─────────────────────────────────────────────────────────────────
+    cat > .env << IOPAINT_ENV
+# IOPaint — change MODEL and restart to switch (no need to edit docker-compose.yml)
+
+# Current model (set during install — see README for full model list)
+MODEL=${IOPAINT_MODEL}
+
+# Device: cpu or cuda
+# For GPU: also requires the deploy: block in docker-compose.yml
+DEVICE=${DEVICE_VAL}
+
+# Caddy network
+CADDY_NET=${SITE_CADDY_NET}
+IOPAINT_ENV
+    chmod 600 .env
+
+    mkdir -p models input output
+    chown -R "$ACTUAL_USER:$ACTUAL_USER" "$IOPAINT_DIR"
+
+    echo ""
+    log_success "IOPaint configured — model: $IOPAINT_MODEL | device: $DEVICE_VAL"
+    if [[ "$IOPAINT_MODEL" == *"PowerPaint"* ]] || [[ "$IOPAINT_MODEL" == *"stable-diffusion"* ]]; then
+        log_info "SD-based model selected (~4 GB). It downloads from HuggingFace on first start."
+        log_info "If the download fails, try: HF_TOKEN=your_token docker compose up -d"
+    else
+        log_info "Model downloads automatically on first start (LaMa ~200 MB)."
+    fi
+    log_info "Switch models any time by editing MODEL= in .env and restarting."
+
+    # No built-in auth — offer Authelia SSO protection
+    local EXTRA_BLOCK=""
+    if [ -d "$DOCKER_DIR/authelia" ]; then
+        local _use_auth=""
+        prompt_yn "Protect IOPaint with Authelia SSO? (y/n):" "y" _use_auth
+        [[ "$_use_auth" =~ ^[Yy]$ ]] && EXTRA_BLOCK="    import authelia"
+    fi
+
+    configure_caddy_for_service "IOPaint" "iopaint:8080" "inpaint" "$EXTRA_BLOCK"
+
+    write_readme "$IOPAINT_DIR" << 'MD'
+# IOPaint
+
+AI-powered image editing:
+- **Erase / remove** — mask an object, AI fills the background (LaMa, works CPU-only)
+- **Inpaint / restore** — fix damaged areas, remove watermarks
+- **Replace with AI** — mask something + type what goes there → AI draws it (PowerPaint, GPU)
+
+Note: IOPaint is **local only** — all inference runs on this machine.
+For text-guided replacement a CUDA GPU on this machine is required.
+
+## Access
+- URL: http://localhost:8100
+- No built-in login — protect via Authelia SSO if exposed
+
+## Switching models
+Edit `MODEL=` in `.env` and restart — no need to touch `docker-compose.yml`:
+```bash
+cd ~/docker/iopaint
+nano .env          # change MODEL= line
+docker compose restart
+```
+
+## Model reference
+
+| # | Model | Type | Size | Best for |
+|---|-------|------|------|---------|
+| 1 | `lama` | CPU-safe | ~200 MB | **Object erase/removal** (default) |
+| 2 | `cv2` | CPU-safe | built-in | Basic fill, no download |
+| 3 | `zits` | CPU-safe | ~200 MB | Portrait & face restoration |
+| 4 | `manga` | CPU-safe | ~100 MB | Comic/manga text bubble removal |
+| 5 | `migan` | GPU | ~50 MB | Fast GPU inpainting |
+| 6 | `fcf` | GPU | ~600 MB | High-quality contextual fill |
+| 7 | `mat` | GPU | ~300 MB | Large missing region fill |
+| 8 | `ldm` | GPU | ~1.2 GB | Latent diffusion texture fill |
+| 9 | `Sanster/PowerPaint-V2-filling` | GPU+SD | ~4 GB | **Text-guided replacement** |
+| 10 | `runwayml/stable-diffusion-inpainting` | GPU+SD | ~4 GB | SD 1.5 inpaint, large LoRA library |
+
+SD-based models (9, 10) download from HuggingFace on first start.
+If a gated model needs a token: add `HF_TOKEN=xxx` to `.env`.
+
+## How to use
+1. Open http://localhost:8100
+2. Upload an image (or drag & drop)
+3. Paint a mask over the area to change
+4. For erase models: click Run → gap fills automatically
+5. For SD models (PowerPaint): type a text prompt → AI draws it into the masked area
+
+## GPU acceleration
+Requires `nvidia-container-toolkit`. The GPU compose adds a `deploy:` block.
+Re-run the installer with GPU=y to regenerate docker-compose.yml, or manually add:
+```yaml
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+```
+Then change `DEVICE=cuda` in `.env` and restart.
+
+## Manage
+```bash
+cd ~/docker/iopaint
+docker compose up -d
+docker compose down
+docker compose logs -f
+docker compose pull && docker compose down && docker compose up -d
+```
+
+## Files
+- docker-compose.yml — stack (MODEL and DEVICE come from .env)
+- .env               — model and device config
+- models/            — all cached model weights (torch + HuggingFace)
+- input/, output/    — optional file staging
+MD
+
+    local START_IO=""
+    prompt_yn "Start IOPaint now? (y/n):" "y" START_IO
+    if [ "$START_IO" = "y" ] || [ "$START_IO" = "Y" ]; then
+        docker compose up -d \
+            && log_success "IOPaint started — model downloads on first use" \
+            || log_warning "Start failed — check: docker compose logs"
+    fi
+
+    echo ""
+    echo "  URL:      http://localhost:8100"
+    echo "  Model:    $IOPAINT_MODEL"
+    echo "  Device:   $DEVICE_VAL"
+    echo "  Switch:   edit MODEL= in $IOPAINT_DIR/.env and restart"
+    echo ""
+}
+
+[[ "${_RUN_STANDALONE:-0}" == 1 ]] && install_iopaint