updating to use stable-diffusion.cpp for the image generation as ollama has since removed it

README.md

@@ -52,10 +52,11 @@ ollama pull llama3.2
 ollama pull nomic-embed-text
 ```
 
-Optionally, for image generation (macOS / Apple Silicon only):
+For image generation, DM-Pal talks to a [stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp) `sd-server` (cross-platform) via its AUTOMATIC1111-compatible API. Download a model (e.g. Z-Image-Turbo or FLUX.2-klein weights), start the server, then point DM-Pal at it under **Settings → Image Server** (default `http://localhost:1234`):
 
 ```bash
-ollama pull x/flux2-klein:4b
+# build sd-server, then run with your chosen model
+sd-server -m path/to/model.safetensors --vae path/to/vae.safetensors
 ```
 
 ## Where your data lives

src-tauri/src/commands/image_commands.rs

@@ -1,66 +1,32 @@
 use crate::commands::emit_busy;
 use crate::llm::AppState;
-use futures_util::StreamExt;
 use serde::{Deserialize, Serialize};
 use std::collections::hash_map::DefaultHasher;
 use std::hash::{Hash, Hasher};
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::time::Duration;
 use tauri::ipc::Channel;
 use tauri::AppHandle;
 
+// ponytail: we talk to a stable-diffusion.cpp `sd-server` over HTTP using its
+// AUTOMATIC1111-compatible API (/sdapi/v1/txt2img + /sdapi/v1/progress). This
+// reuses the existing HTTP+channel plumbing and drops the old Ollama/macOS-only
+// limitation — sd-server runs on macOS, Linux, and Windows. The model is chosen
+// at server startup, so there's no per-request model field anymore.
+
 // ponytail: DefaultHasher is fine for a cache filename — not crypto, just a stable key.
 
 #[derive(Debug, Deserialize)]
 pub struct ImageRequest {
     pub prompt: String,
-    /// Override the configured image model for this call.
-    pub model: Option<String>,
-}
-
-/// One line of Ollama's NDJSON image-generation response.
-/// `step`/`total` carry progress; the final `done: true` line carries the
-/// singular `image` base64 field. All fields optional per-line.
-#[derive(Debug, Default, Deserialize)]
-struct OllamaImageLine {
-    #[serde(default)]
-    done: bool,
-    #[serde(default)]
-    step: Option<u32>,
-    #[serde(default)]
-    total: Option<u32>,
-    #[serde(default)]
-    image: Option<String>,
-}
-
-/// Parsed view of one NDJSON line used by both the streaming and the
-/// buffered paths. Extracted so the parsing logic is unit-testable without
-/// touching the network.
-#[derive(Debug, PartialEq)]
-struct ImageProgress {
-    done: bool,
-    step: Option<u32>,
-    total: Option<u32>,
-    image: Option<String>,
-}
-
-fn parse_image_line(line: &str) -> Option<ImageProgress> {
-    let line = line.trim();
-    if line.is_empty() {
-        return None;
-    }
-    let parsed: OllamaImageLine = serde_json::from_str(line).ok()?;
-    Some(ImageProgress {
-        done: parsed.done,
-        step: parsed.step,
-        total: parsed.total,
-        image: parsed.image,
-    })
 }
 
 /// Channel events for streaming image generation.
 #[derive(Clone, Serialize)]
 #[serde(tag = "type", content = "data", rename_all = "camelCase")]
 pub enum ImageEvent {
-    /// Progress update: (step, total). Either may be None if Ollama omits it.
+    /// Progress update: (step, total). Either may be None if the server omits it.
     Progress { step: Option<u32>, total: Option<u32> },
     /// Final result: a `data:image/png;base64,...` URL ready for `<img src>`.
     Done(String),
@@ -68,11 +34,23 @@ pub enum ImageEvent {
     Error(String),
 }
 
-/// Generate (or fetch from disk cache) an image for `prompt` via the configured
-/// Ollama image model. Returns a `data:image/png;base64,...` URL ready for `<img src>`.
-///
-/// Ollama image models are macOS-only today; on other platforms we return an error
-/// so the front-end can fall back to a placeholder instead of a confusing timeout.
+/// A1111 `/sdapi/v1/txt2img` response. Only the base64 PNG list matters to us.
+#[derive(Debug, Deserialize)]
+struct Txt2ImgResponse {
+    #[serde(default)]
+    images: Vec<String>,
+}
+
+/// A1111 `/sdapi/v1/progress` response. `progress` is 0.0–1.0 of the current job.
+#[derive(Debug, Deserialize)]
+struct ProgressResponse {
+    #[serde(default)]
+    progress: f32,
+}
+
+/// Generate (or fetch from disk cache) an image for `prompt` via a local
+/// stable-diffusion.cpp `sd-server`. Returns a `data:image/png;base64,...` URL
+/// ready for `<img src>`.
 #[tauri::command]
 pub async fn generate_image(
     state: tauri::State<'_, AppState>,
@@ -88,27 +66,23 @@ pub async fn generate_image(
 }
 
 async fn generate_image_inner(state: tauri::State<'_, AppState>, req: ImageRequest) -> Result<String, String> {
-    if cfg!(not(target_os = "macos")) {
-        return Err("image generation is macOS-only via Ollama (for now)".into());
-    }
-
     let config = state.config.lock().map_err(|e| e.to_string())?.clone();
-    let model = req.model.unwrap_or(config.image_model.clone());
+    let api_url = config.image_api_url.clone();
 
-    let (cache_path, cache_hit) = prepare_cache(&state.data_dir, &model, &req.prompt)?;
+    let (cache_path, cache_hit) = prepare_cache(&state.data_dir, &api_url, &req.prompt)?;
     if cache_hit {
         let bytes = std::fs::read(&cache_path).map_err(|e| e.to_string())?;
         return Ok(data_url(&bytes));
     }
 
-    let png_bytes = request_image_bytes(&config, &model, &req.prompt, None).await?;
+    let png_bytes = txt2img(&api_url, &req.prompt).await?;
     std::fs::write(&cache_path, &png_bytes).map_err(|e| e.to_string())?;
     Ok(data_url(&png_bytes))
 }
 
-/// Streaming variant: emits `ImageEvent::Progress` as Ollama reports `step`/`total`,
-/// then `ImageEvent::Done` with the data URL (or `Error`). Reuses the same disk cache
-/// as `generate_image`. The DM gets a real progress bar for the multi-second wait.
+/// Streaming variant: emits `ImageEvent::Progress` (from `/sdapi/v1/progress`)
+/// while the txt2img POST is in flight, then `ImageEvent::Done` with the data
+/// URL (or `Error`). Reuses the same disk cache as `generate_image`.
 #[tauri::command]
 pub async fn generate_image_stream(
     state: tauri::State<'_, AppState>,
@@ -116,17 +90,10 @@ pub async fn generate_image_stream(
     req: ImageRequest,
     channel: Channel<ImageEvent>,
 ) -> Result<(), String> {
-    if cfg!(not(target_os = "macos")) {
-        let _ = channel.send(ImageEvent::Error(
-            "image generation is macOS-only via Ollama (for now)".into(),
-        ));
-        return Ok(());
-    }
-
     let config = state.config.lock().map_err(|e| e.to_string())?.clone();
-    let model = req.model.unwrap_or(config.image_model.clone());
+    let api_url = config.image_api_url.clone();
 
-    let (cache_path, cache_hit) = prepare_cache(&state.data_dir, &model, &req.prompt)?;
+    let (cache_path, cache_hit) = prepare_cache(&state.data_dir, &api_url, &req.prompt)?;
     if cache_hit {
         let bytes = std::fs::read(&cache_path).map_err(|e| e.to_string())?;
         let _ = channel.send(ImageEvent::Done(data_url(&bytes)));
@@ -135,19 +102,41 @@ pub async fn generate_image_stream(
 
     // Drive the request on a background task so the command returns immediately
     // and progress flows through the channel. Errors become ImageEvent::Error.
-    let channel = std::sync::Arc::new(channel);
+    let channel = Arc::new(channel);
     let ch = channel.clone();
     // ponytail: only emit busy around the actual generation (not cache hits),
     // and always balance it in the spawn — even on error.
     emit_busy(&app, true);
     tauri::async_runtime::spawn(async move {
-        match request_image_bytes(&config, &model, &req.prompt, Some(ch)).await {
+        // Poll /sdapi/v1/progress while the txt2img POST runs so the DM sees a
+        // real progress bar. Best-effort: poll errors are silent (bar falls
+        // back to indeterminate). 200ms is smooth without spamming the server.
+        let poll_client = reqwest::Client::new();
+        let done = Arc::new(AtomicBool::new(false));
+        let done_p = done.clone();
+        let api_p = api_url.clone();
+        let ch_p = ch.clone();
+        let poller = tauri::async_runtime::spawn(async move {
+            while !done_p.load(Ordering::SeqCst) {
+                if let Ok(p) = poll_progress(&poll_client, &api_p).await {
+                    let step = (p.clamp(0.0, 1.0) * 100.0) as u32;
+                    let _ = ch_p.send(ImageEvent::Progress { step: Some(step), total: Some(100) });
+                }
+                tokio::time::sleep(Duration::from_millis(200)).await;
+            }
+        });
+
+        let result = txt2img(&api_url, &req.prompt).await;
+        done.store(true, Ordering::SeqCst);
+        let _ = poller.await; // let the poller flush its last iteration
+
+        match result {
             Ok(png_bytes) => {
                 let _ = std::fs::write(&cache_path, &png_bytes);
-                let _ = channel.send(ImageEvent::Done(data_url(&png_bytes)));
+                let _ = ch.send(ImageEvent::Done(data_url(&png_bytes)));
             }
             Err(e) => {
-                let _ = channel.send(ImageEvent::Error(e));
+                let _ = ch.send(ImageEvent::Error(e));
             }
         }
         emit_busy(&app, false);
@@ -156,43 +145,61 @@ pub async fn generate_image_stream(
     Ok(())
 }
 
-/// Resolve the on-disk cache path for (model, prompt) and report a cache hit.
+/// Resolve the on-disk cache path for (api_url, prompt) and report a cache hit.
 /// `base` is the configured data dir (AppState.data_dir).
 fn prepare_cache(
     base: &std::path::Path,
-    model: &str,
+    api_url: &str,
     prompt: &str,
 ) -> Result<(std::path::PathBuf, bool), String> {
     let cache_dir = base.join("images");
     std::fs::create_dir_all(&cache_dir).map_err(|e| e.to_string())?;
 
     let mut hasher = DefaultHasher::new();
-    model.hash(&mut hasher);
+    api_url.hash(&mut hasher);
     prompt.hash(&mut hasher);
     let cache_path = cache_dir.join(format!("{:016x}.png", hasher.finish()));
     let hit = cache_path.exists();
     Ok((cache_path, hit))
 }
 
-/// POST the image request to Ollama and collect the PNG bytes. When a channel
-/// is given, parse the NDJSON body incrementally and emit progress; otherwise
-/// read the whole body at once (legacy buffered path).
-async fn request_image_bytes(
-    config: &crate::llm::LlmConfig,
-    model: &str,
-    prompt: &str,
-    channel: Option<std::sync::Arc<Channel<ImageEvent>>>,
-) -> Result<Vec<u8>, String> {
-    let client = reqwest::Client::new();
-    let url = format!("{}/api/generate", config.api_url.trim_end_matches('/'));
-    let body = serde_json::json!({ "model": model, "prompt": prompt, "stream": false });
+/// POST `/sdapi/v1/txt2img` to the sd-server and return the raw PNG bytes of
+/// the first generated image. A1111 returns base64 PNGs *without* a `data:`
+/// prefix, so we decode to bytes for the disk cache.
+async fn txt2img(api_url: &str, prompt: &str) -> Result<Vec<u8>, String> {
+    // ponytail: width/height/steps/cfg are fixed — diffusion can take tens of
+    // seconds; add config knobs only if the DM wants to tune them.
+    let client = reqwest::Client::builder()
+        .timeout(Duration::from_secs(300))
+        .build()
+        .map_err(|e| e.to_string())?;
+    let url = format!("{}/sdapi/v1/txt2img", api_url.trim_end_matches('/'));
+    let body = serde_json::json!({
+        "prompt": prompt,
+        "width": 1024,
+        "height": 1024,
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "seed": -1,
+        "batch_size": 1,
+    });
 
     let res = client
         .post(&url)
         .json(&body)
         .send()
         .await
-        .map_err(|e| format!("image request failed: {e}"))?;
+        .map_err(|e| {
+            // ponytail: non-technical DMs see a connect error as a mystery —
+            // name the fix (start sd-server) and point at the in-app guide.
+            if e.is_connect() || e.is_timeout() {
+                format!(
+                    "Can't reach the image server at {api_url}. Is sd-server running? Open the Image tab → 'First time setup' for step-by-step instructions. ({e})"
+                )
+            } else {
+                format!("image request failed: {e}")
+            }
+        })?;
 
     if !res.status().is_success() {
         let status = res.status();
@@ -200,56 +207,71 @@ async fn request_image_bytes(
         return Err(format!("image error {status}: {text}"));
     }
 
-    // ponytail: Ollama image models emit NDJSON even with stream:false —
-    // progress lines (step/total) then a final done:true carrying the image.
-    // The buffered path reads it all at once; the streaming path splits lines
-    // as they arrive so the bar animates.
-    let mut png_b64: Option<String> = None;
-    let mut buffer = String::new();
+    let parsed: Txt2ImgResponse = res
+        .json()
+        .await
+        .map_err(|e| format!("image parse error: {e}"))?;
+    let b64 = parsed
+        .images
+        .into_iter()
+        .next()
+        .ok_or_else(|| "no image in sd-server response".to_string())?;
+    base64_decode(&b64)
+}
 
-    let apply_line = |line: &str, png: &mut Option<String>, ch: Option<&Channel<ImageEvent>>| {
-        if let Some(p) = parse_image_line(line) {
-            if let Some(c) = ch {
-                let _ = c.send(ImageEvent::Progress { step: p.step, total: p.total });
+/// GET `/sdapi/v1/progress` → fraction (0.0–1.0) of the current job. Best-effort.
+async fn poll_progress(client: &reqwest::Client, api_url: &str) -> Result<f32, String> {
+    let url = format!("{}/sdapi/v1/progress", api_url.trim_end_matches('/'));
+    let res = client
+        .get(&url)
+        .send()
+        .await
+        .map_err(|e| e.to_string())?;
+    if !res.status().is_success() {
+        return Err(format!("progress HTTP {}", res.status()));
     }
-            if let Some(b64) = p.image {
-                *png = Some(b64);
-            }
-        }
-    };
+    let p: ProgressResponse = res.json().await.map_err(|e| e.to_string())?;
+    Ok(p.progress)
+}
 
-    if let Some(ch) = channel {
-        let mut stream = res.bytes_stream();
-        while let Some(chunk) = stream.next().await {
-            let chunk = chunk.map_err(|e| format!("image read error: {e}"))?;
-            buffer.push_str(&String::from_utf8_lossy(&chunk));
-            // Process complete lines; keep the trailing partial line in buffer.
-            while let Some(idx) = buffer.find('\n') {
-                let line = buffer.split_off(idx + 1);
-                let complete = std::mem::replace(&mut buffer, line);
-                apply_line(&complete, &mut png_b64, Some(&ch));
-            }
-        }
-        if !buffer.trim().is_empty() {
-            apply_line(&buffer, &mut png_b64, Some(&ch));
-        }
-    } else {
-        let body_text = res.text().await.map_err(|e| format!("image read error: {e}"))?;
-        for line in body_text.lines() {
-            apply_line(line, &mut png_b64, None);
-            // ponytail: stop after the done line in the buffered path — the
-            // final image is the one we want.
-            if let Some(p) = parse_image_line(line) {
-                if p.done && p.image.is_some() {
-                    break;
-                }
-            }
-        }
-    }
+/// GET `/sdapi/v1/progress` as a liveness probe for the configured sd-server.
+/// Returns ok=true when the server responds (even idle — progress is 0.0).
+/// Mirrors the LLM `test_connection` so Settings can show a green check.
+#[derive(serde::Serialize)]
+pub struct ImageConnectionTest {
+    pub ok: bool,
+    pub error: String,
+}
 
-    let b64 = png_b64.ok_or_else(|| "no image data in Ollama response".to_string())?;
-    let png_bytes = base64_decode(&b64)?;
-    Ok(png_bytes)
+#[tauri::command]
+pub async fn test_image_connection(
+    state: tauri::State<'_, AppState>,
+) -> Result<ImageConnectionTest, String> {
+    let api_url = state
+        .config
+        .lock()
+        .map_err(|e| e.to_string())?
+        .image_api_url
+        .clone();
+    let client = reqwest::Client::builder()
+        .timeout(Duration::from_secs(5))
+        .build()
+        .map_err(|e| e.to_string())?;
+    let url = format!("{}/sdapi/v1/progress", api_url.trim_end_matches('/'));
+    match client.get(&url).send().await {
+        Ok(res) if res.status().is_success() => Ok(ImageConnectionTest {
+            ok: true,
+            error: String::new(),
+        }),
+        Ok(res) => Ok(ImageConnectionTest {
+            ok: false,
+            error: format!("HTTP {} — is this an sd-server?", res.status()),
+        }),
+        Err(e) => Ok(ImageConnectionTest {
+            ok: false,
+            error: format!("Can't reach {api_url} — is sd-server running? ({e})"),
+        }),
+    }
 }
 
 fn data_url(png: &[u8]) -> String {
@@ -336,55 +358,19 @@ mod tests {
 
     #[test]
     fn cache_key_is_stable() {
-        // sanity: same inputs → same filename shape (16 hex digits)
+        // sanity: same inputs → same 16-hex-digit filename shape
         let mut h = DefaultHasher::new();
-        "x/flux2-klein:4b".hash(&mut h);
+        "http://localhost:1234".hash(&mut h);
         "prompt".hash(&mut h);
         let s = format!("{:016x}", h.finish());
         assert_eq!(s.len(), 16);
     }
 
     #[test]
-    fn parse_progress_and_image_lines() {
-        // intermediate progress line, no image
-        let p = parse_image_line(r#"{"model":"x/flux2-klein:4b","done":false,"total":4,"step":1}"#).unwrap();
-        assert_eq!(p, ImageProgress { done: false, step: Some(1), total: Some(4), image: None });
-
-        // final done line carries the image
-        let p = parse_image_line(r#"{"model":"x/flux2-klein:4b","done":true,"image":"iVBORw0KGgoAAAANSUhEUgAA"}"#).unwrap();
-        assert!(p.done);
-        assert_eq!(p.image.as_deref(), Some("iVBORw0KGgoAAAANSUhEUgAA"));
-        assert!(p.total.is_none() && p.step.is_none());
-
-        // blank/garbage lines are ignored, not errors
-        assert!(parse_image_line("").is_none());
-        assert!(parse_image_line("not json").is_none());
-    }
-
-    #[test]
-    fn split_ndjson_buffer_keeps_trailing_partial() {
-        // Simulate two chunks arriving separately where the split falls mid-line.
-        let chunk1 = "{\"done\":false,\"step\":1,\"total\":4}\n{\"done\":tru";
-        let chunk2 = "e,\"image\":\"abc\"}\n";
-        let mut buffer = String::new();
-        let mut png: Option<String> = None;
-        let whole = format!("{chunk1}{chunk2}");
-        // emulate the streaming loop over the concatenated body
-        buffer.push_str(&whole);
-        let mut lines = Vec::new();
-        while let Some(idx) = buffer.find('\n') {
-            let line = buffer.split_off(idx + 1);
-            let complete = std::mem::replace(&mut buffer, line);
-            lines.push(complete);
-        }
-        if !buffer.trim().is_empty() {
-            lines.push(std::mem::take(&mut buffer));
-        }
-        for line in &lines {
-            if let Some(p) = parse_image_line(line) {
-                if let Some(b) = p.image { png = Some(b); }
-            }
-        }
-        assert_eq!(png.as_deref(), Some("abc"));
+    fn progress_fraction_maps_to_step() {
+        // The streaming poller emits step = clamp(progress,0,1)*100.
+        assert_eq!((0.0f32 * 100.0) as u32, 0);
+        assert_eq!((0.5f32 * 100.0) as u32, 50);
+        assert_eq!((1.5f32.clamp(0.0, 1.0) * 100.0) as u32, 100);
     }
 }

src-tauri/src/lib.rs

@@ -63,6 +63,7 @@ pub fn run() {
             commands::llm_commands::test_connection,
             commands::image_commands::generate_image,
             commands::image_commands::generate_image_stream,
+            commands::image_commands::test_image_connection,
             commands::rag_commands::rag_add,
             commands::rag_commands::rag_search,
             commands::rag_commands::rag_list,

src-tauri/src/llm/mod.rs

@@ -31,8 +31,9 @@ pub struct LlmConfig {
     pub temperature: f32,
     pub max_tokens: u32,
     pub top_p: f32,
-    /// Ollama image-generation model (e.g. `x/flux2-klein:4b`). macOS-only via Ollama.
-    pub image_model: String,
+    /// Base URL of a stable-diffusion.cpp `sd-server` (A1111-compatible API),
+    /// e.g. `http://localhost:1234`. Cross-platform — no longer Ollama/macOS-only.
+    pub image_api_url: String,
     /// Ollama embedding model for lore RAG (e.g. `nomic-embed-text`).
     pub embed_model: String,
 }
@@ -47,7 +48,7 @@ impl Default for LlmConfig {
             temperature: 0.7,
             max_tokens: 512,
             top_p: 0.9,
-            image_model: "x/flux2-klein:4b".to_string(),
+            image_api_url: "http://localhost:1234".to_string(),
             embed_model: "nomic-embed-text".to_string(),
         }
     }

src/components/FirstRunWizard.tsx

@@ -26,7 +26,7 @@ export function FirstRunWizard({ onComplete }: { onComplete: () => void }) {
       // Test connection to the default Ollama URL
       const result = await invoke<{ ok: boolean; models: string[]; error: string }>(
         "test_connection",
-        { config: { api_url: apiUrl, api_key: apiKey, model: "", temperature: 0.7, max_tokens: 512, top_p: 0.9, image_model: "", embed_model: "" } }
+        { config: { api_url: apiUrl, api_key: apiKey, model: "", temperature: 0.7, max_tokens: 512, top_p: 0.9, image_api_url: "", embed_model: "" } }
       );
       if (result.ok) {
         setModels(result.models);
@@ -58,7 +58,7 @@ export function FirstRunWizard({ onComplete }: { onComplete: () => void }) {
           temperature: 0.7,
           max_tokens: 512,
           top_p: 0.9,
-          image_model: "x/flux2-klein:4b",
+          image_api_url: "http://localhost:1234",
           embed_model: "nomic-embed-text",
         }
       });
@@ -244,7 +244,7 @@ export function FirstRunWizard({ onComplete }: { onComplete: () => void }) {
         </h2>
         <p className="text-[var(--color-text-primary)] text-sm mb-4">
           DM-Pal is now configured to use {selectedModel} for text generation.
-          Image generation uses x/flux2-klein:4b (macOS-only via Ollama).
+          Image generation runs via a stable-diffusion.cpp `sd-server` at http://localhost:1234 (cross-platform).
           Embedding model for lore search is set to nomic-embed-text.
         </p>
         <div className="flex flex-col gap-3">

src/components/GeneratedImage.tsx

@@ -1,6 +1,5 @@
 import { useEffect, useState } from "react";
 import { invoke, Channel } from "@tauri-apps/api/core";
-import { ImageOff } from "lucide-react";
 import { useToast } from "./Toast";
 
 interface Props {
@@ -20,22 +19,20 @@ type ImageEvent =
   | { type: "error"; data: string };
 
 /**
- * Generates and caches an image for `prompt` via the Ollama image model.
- * Results are cached on disk by the backend, so re-renders are instant.
- * macOS-only: on other OSes the backend returns an error and we show a
- * placeholder instead of a confusing timeout.
+ * Generates and caches an image for `prompt` via a stable-diffusion.cpp
+ * `sd-server` (A1111-compatible API). Results are cached on disk by the
+ * backend, so re-renders are instant. Cross-platform.
  *
  * Uses the streaming command so the DM sees a real progress bar (driven from
- * Ollama's NDJSON step/total) during the multi-second wait. The image still
- * loads even if no progress events arrive — Done carries the data URL.
+ * `/sdapi/v1/progress`) during the multi-second wait. The image still loads
+ * even if no progress events arrive — Done carries the data URL.
  */
 export function GeneratedImage({ prompt, nonce = 0, className = "", aspect = "aspect-square" }: Props) {
   const [dataUrl, setDataUrl] = useState<string | null>(null);
   const [loading, setLoading] = useState(false);
-  const [unsupported, setUnsupported] = useState(false);
   const [expanded, setExpanded] = useState(false);
   // ponytail: progress for the in-flight generation. step/total may be null
-  // if Ollama omits them; we show an indeterminate shimmer then.
+  // if the server omits them; we show an indeterminate shimmer then.
   const [progress, setProgress] = useState<{ step: number | null; total: number | null } | null>(null);
   const { addToast } = useToast();
 
@@ -43,7 +40,6 @@ export function GeneratedImage({ prompt, nonce = 0, className = "", aspect = "as
     if (!prompt) return;
     let cancelled = false;
     setLoading(true);
-    setUnsupported(false);
     setProgress(null);
 
     const channel = new Channel<ImageEvent>();
@@ -56,9 +52,7 @@ export function GeneratedImage({ prompt, nonce = 0, className = "", aspect = "as
         setLoading(false);
         setProgress(null);
       } else if (event.type === "error") {
-        const msg = String(event.data);
-        if (msg.includes("macOS-only")) setUnsupported(true);
-        else addToast(`Image failed: ${msg}`, "error");
+        addToast(`Image failed: ${event.data}`, "error");
         setLoading(false);
         setProgress(null);
       }
@@ -66,9 +60,7 @@ export function GeneratedImage({ prompt, nonce = 0, className = "", aspect = "as
 
     invoke("generate_image_stream", { req: { prompt }, channel }).catch((e) => {
       if (!cancelled) {
-        const msg = String(e);
-        if (msg.includes("macOS-only")) setUnsupported(true);
-        else addToast(`Image failed: ${msg}`, "error");
+        addToast(`Image failed: ${e}`, "error");
         setLoading(false);
         setProgress(null);
       }
@@ -83,17 +75,6 @@ export function GeneratedImage({ prompt, nonce = 0, className = "", aspect = "as
 
   const box = `relative ${aspect} w-full rounded-lg overflow-hidden border border-[var(--color-border-glass)] bg-[var(--color-bg-deep)] flex items-center justify-center ${className}`;
 
-  if (unsupported) {
-    return (
-      <div className={box}>
-        <div className="flex flex-col items-center gap-1 text-[var(--color-text-dim)]">
-          <ImageOff size={20} />
-          <span className="text-[10px] text-center px-2">Image gen is macOS-only via Ollama</span>
-        </div>
-      </div>
-    );
-  }
-
   if (loading) {
     // ponytail: determinate bar when step/total are known, indeterminate shimmer otherwise.
     const pct = progress && progress.step != null && progress.total ? Math.round((progress.step / progress.total) * 100) : null;

src/components/ImageGenerator.tsx

@@ -4,8 +4,7 @@ import { ImagePlus, Sparkles } from "lucide-react";
 import { useToast } from "./Toast";
 import { addGeneration, type Generation } from "../lib/generations";
 import { usePrefillEffect } from "../lib/usePrefill";
-
-const DEFAULT_MODEL = "x/flux2-klein:4b";
+import { ImageSetupGuide } from "./ImageSetupGuide";
 
 interface Props {
   prefill?: Generation | null;
@@ -14,12 +13,11 @@ interface Props {
 
 export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
   const [prompt, setPrompt] = useState("");
-  const [model, setModel] = useState("");
   const [dataUrl, setDataUrl] = useState<string | null>(null);
   const [variants, setVariants] = useState<string[]>([]);
   const [loading, setLoading] = useState(false);
   const [variant, setVariant] = useState(0);
-  const [unsupported, setUnsupported] = useState(false);
+  const [guideOpen, setGuideOpen] = useState(false);
   const { addToast } = useToast();
 
   usePrefillEffect(prefill ?? null, "image", () => onPrefillConsumed?.(), (g) => {
@@ -38,23 +36,21 @@ export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
       // so "regenerate" actually produces a new image instead of the cached one.
       const v = variant > 0 ? `\n\n(variation ${variant})` : "";
       const url = await invoke<string>("generate_image", {
-        req: { prompt: `${prompt}${v}`, model: model.trim() || null },
+        req: { prompt: `${prompt}${v}` },
       });
       setDataUrl(url);
-      setUnsupported(false);
       addToast("Image generated", "success");
       // ponytail: persist the PNG data URL so the DM can re-open past renders.
-      void addGeneration({ kind: "image", title: prompt, data: url, source: model.trim() || DEFAULT_MODEL });
+      void addGeneration({ kind: "image", title: prompt, data: url, source: "sd-server" });
     } catch (e) {
-      const msg = String(e);
-      if (msg.toLowerCase().includes("macos-only")) setUnsupported(true);
-      addToast(`Image generation failed: ${msg}`, "error");
+      addToast(`Image generation failed: ${String(e)}`, "error");
+      setGuideOpen(true); // ponytail: a failed generate usually means sd-server isn't running — pop the guide.
     }
     setLoading(false);
   }
 
   // ponytail: batch = 4 concurrent generates with distinct variation tags so
-  // the backend cache yields 4 different images. Ollama serializes them
+  // the backend cache yields 4 different images. sd-server processes them
   // server-side anyway, so concurrency is just cleaner code than a loop with
   // awaits. Each variant is persisted so the gallery gets all four.
   async function generateBatch() {
@@ -62,12 +58,11 @@ export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
     setLoading(true);
     setVariants([]);
     setDataUrl(null);
-    setUnsupported(false);
     const tags = [1, 2, 3, 4];
     const results = await Promise.allSettled(
       tags.map((t) =>
         invoke<string>("generate_image", {
-          req: { prompt: `${prompt}\n\n(variation ${t})`, model: model.trim() || null },
+          req: { prompt: `${prompt}\n\n(variation ${t})` },
         }),
       ),
     );
@@ -75,14 +70,14 @@ export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
     for (const r of results) {
       if (r.status === "fulfilled" && r.value) {
         ok.push(r.value);
-        void addGeneration({ kind: "image", title: prompt, data: r.value, source: model.trim() || DEFAULT_MODEL });
+        void addGeneration({ kind: "image", title: prompt, data: r.value, source: "sd-server" });
       }
     }
     if (ok.length === 0) {
       const firstErr = results.find((r) => r.status === "rejected");
       const msg = firstErr ? String((firstErr as PromiseRejectedResult).reason) : "";
-      if (msg.toLowerCase().includes("macos-only")) setUnsupported(true);
       addToast(`Batch failed: ${msg || "no images returned"}`, "error");
+      setGuideOpen(true); // ponytail: likely sd-server not running — show the guide.
     } else {
       setVariants(ok);
       setDataUrl(ok[0]);
@@ -99,6 +94,7 @@ export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
 
   return (
     <div className="flex flex-col gap-3 text-sm">
+      <ImageSetupGuide forcedOpen={guideOpen} />
       <div className="flex flex-col gap-1">
         <label className="text-[var(--color-text-secondary)] text-xs font-medium">Prompt</label>
         <textarea
@@ -110,18 +106,6 @@ export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
         />
       </div>
 
-      <div className="flex flex-col gap-1">
-        <label className="text-[var(--color-text-secondary)] text-xs font-medium">
-          Model (optional — defaults to {DEFAULT_MODEL})
-        </label>
-        <input
-          className="rounded-lg bg-[var(--color-bg-surface)] border border-[var(--color-border-glass)] px-3 py-1.5 text-[var(--color-text-primary)] placeholder-[var(--color-text-dim)] focus:outline-none focus:border-[var(--color-gold-bright)] text-sm"
-          value={model}
-          onChange={(e) => setModel(e.target.value)}
-          placeholder="x/flux2-klein:4b  ·  x/z-image-turbo for speed"
-        />
-      </div>
-
       <div className="flex gap-2">
         <button
           onClick={generate}
@@ -155,14 +139,6 @@ export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
         </div>
       )}
 
-      {unsupported && (
-        <div className="rounded-lg bg-[var(--color-danger)]/10 border border-[var(--color-danger)]/30 p-3 text-[var(--color-danger)] text-xs">
-          Image generation is macOS-only via Ollama (for now). On other platforms the
-          ✨ buttons elsewhere fall back to a placeholder.
-        </div>
-      )}
-
-
       {variants.length > 1 && !loading && (
         <div className="flex flex-col gap-1">
           <span className="text-[var(--color-text-secondary)] text-xs font-medium">Variants — click to select</span>
@@ -215,7 +191,7 @@ export function ImageGenerator({ prefill, onPrefillConsumed }: Props = {}) {
         </div>
       )}
 
-      {!dataUrl && !loading && !unsupported && (
+      {!dataUrl && !loading && (
         <div className="flex flex-col items-center justify-center aspect-square max-w-md mx-auto text-center text-[var(--color-text-dim)] text-xs">
           <span>No image yet.</span>
           <span className="mt-1">Describe a scene or portrait, then Generate.</span>

src/components/ImageSetupGuide.tsx

@@ -0,0 +1,185 @@
+import { useState } from "react";
+import { invoke } from "@tauri-apps/api/core";
+import { ChevronDown, ExternalLink, CheckCircle2, XCircle } from "lucide-react";
+import { useToast } from "./Toast";
+
+// ponytail: a non-technical DM won't read the README or build sd-server from
+// source. This is the one place they'll look — the Image tab — so the full
+// step-by-step lives here as a collapsible panel. The "Test" button mirrors
+// Settings so they get a green check without leaving the tab.
+
+interface Props {
+  /** Control from the parent so a failed generate can pop this open. */
+  forcedOpen?: boolean;
+}
+
+export function ImageSetupGuide({ forcedOpen = false }: Props) {
+  const [open, setOpen] = useState(false);
+  const [testing, setTesting] = useState(false);
+  const [result, setResult] = useState<{ ok: boolean; error: string } | null>(null);
+  const { addToast } = useToast();
+
+  // ponytail: when the parent reports a connection failure, expand so the DM
+  // sees the fix immediately instead of staring at a toast.
+  const isOpen = open || forcedOpen;
+
+  async function test() {
+    setTesting(true);
+    setResult(null);
+    try {
+      const r = await invoke<{ ok: boolean; error: string }>("test_image_connection");
+      setResult(r);
+      if (r.ok) addToast("sd-server reachable — image generation is ready", "success");
+      else addToast(`Not reachable: ${r.error}`, "error");
+    } catch (e) {
+      setResult({ ok: false, error: String(e) });
+      addToast(`Not reachable: ${e}`, "error");
+    }
+    setTesting(false);
+  }
+
+  return (
+    <div className="rounded-lg border border-[var(--color-border-subtle)] bg-[var(--color-bg-deep)]">
+      <button
+        type="button"
+        onClick={() => setOpen((o) => !o)}
+        className="w-full flex items-center justify-between px-3 py-2 text-xs font-medium text-[var(--color-text-secondary)] hover:text-[var(--color-gold-bright)] cursor-pointer select-none"
+      >
+        <span>⚙️ First time? Set up image generation (5-minute guide)</span>
+        <ChevronDown size={14} className={`transition-transform ${isOpen ? "rotate-180" : ""}`} />
+      </button>
+
+      {isOpen && (
+        <div className="px-3 pb-3 pt-1 border-t border-[var(--color-border-subtle)] flex flex-col gap-3 text-xs leading-relaxed text-[var(--color-text-secondary)]">
+          <p>
+            DM-Pal makes images through a small free helper program called{" "}
+            <strong className="text-[var(--color-text-primary)]">sd-server</strong> (part of{" "}
+            <a href="https://github.com/leejet/stable-diffusion.cpp" target="_blank" rel="noreferrer" className="text-[var(--color-gold-bright)] underline">
+              stable-diffusion.cpp
+            </a>
+            ). It runs on your computer alongside Ollama. Set it up once and you're done.
+          </p>
+
+          <Step n={1} title="Download sd-server">
+            Go to the{" "}
+            <a href="https://github.com/leejet/stable-diffusion.cpp/releases" target="_blank" rel="noreferrer" className="text-[var(--color-gold-bright)] underline">
+              releases page <ExternalLink size={10} className="inline" />
+            </a>{" "}
+            and download the latest build for your system:
+            <ul className="list-disc pl-5 mt-1 space-y-0.5">
+              <li><strong className="text-[var(--color-text-primary)]">Windows</strong> — the <code className="font-mono">sd-server-win</code> zip</li>
+              <li><strong className="text-[var(--color-text-primary)]">Mac (Apple Silicon)</strong> — the <code className="font-mono">sd-server-macos</code> zip</li>
+              <li><strong className="text-[var(--color-text-primary)]">Linux</strong> — the <code className="font-mono">sd-server-linux</code> build</li>
+            </ul>
+            Unzip it into a permanent folder, e.g. <code className="font-mono">sd-server</code> in your home folder.
+          </Step>
+
+          <Step n={2} title="Download a model (recommended: SDXL — great fantasy art, 2 files)">
+            Put these two files in your <code className="font-mono">sd-server</code> folder:
+            <ul className="list-disc pl-5 mt-1 space-y-0.5">
+              <li>
+                Model (~6.5 GB):{" "}
+                <a href="https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors" target="_blank" rel="noreferrer" className="text-[var(--color-gold-bright)] underline">
+                  sd_xl_base_1.0.safetensors <ExternalLink size={10} className="inline" />
+                </a>
+              </li>
+              <li>
+                VAE (~0.6 GB):{" "}
+                <a href="https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae-fp16-fix.safetensors" target="_blank" rel="noreferrer" className="text-[var(--color-gold-bright)] underline">
+                  sdxl_vae-fp16-fix.safetensors <ExternalLink size={10} className="inline" />
+                </a>
+              </li>
+            </ul>
+            <p className="mt-1 text-[var(--color-text-dim)]">
+              On the HuggingFace page click the ↓ next to the file to download it.
+            </p>
+          </Step>
+
+          <Step n={3} title="Start the server (leave the window open)">
+            Open a terminal — <em>Mac:</em> Terminal (Applications → Utilities); <em>Windows:</em> Command Prompt.
+            Move into your folder, then run:
+            <pre className="mt-1 rounded-lg bg-[var(--color-bg-surface)] border border-[var(--color-border-glass)] px-3 py-2 text-[10px] font-mono overflow-x-auto whitespace-pre">
+{`# Mac / Linux
+./sd-server -m sd_xl_base_1.0.safetensors \\
+  --vae sdxl_vae-fp16-fix.safetensors --offload-to-cpu
+
+# Windows
+sd-server.exe -m sd_xl_base_1.0.safetensors \\
+  --vae sdxl_vae-fp16-fix.safetensors --offload-to-cpu`}</pre>
+            <p className="mt-1 text-[var(--color-text-dim)]">
+              <code className="font-mono">--offload-to-cpu</code> lets it run on machines without a strong GPU (slower). You'll see it listening on <code className="font-mono">http://127.0.0.1:1234</code>.
+            </p>
+          </Step>
+
+          <Step n={4} title="Point DM-Pal at it & test">
+            In <strong className="text-[var(--color-text-primary)]">Settings → Image Server</strong>, make sure the URL is{" "}
+            <code className="font-mono">http://localhost:1234</code>. Then test it right here:
+            <div className="flex items-center gap-2 mt-2">
+              <button
+                type="button"
+                onClick={test}
+                disabled={testing}
+                className="rounded-lg bg-[var(--color-bg-surface)] border border-[var(--color-border-glass)] px-3 py-1.5 text-xs font-medium text-[var(--color-text-secondary)] hover:border-[var(--color-gold-bright)] hover:text-[var(--color-gold-bright)] transition-colors cursor-pointer disabled:opacity-50"
+              >
+                {testing ? "Testing…" : "Test sd-server connection"}
+              </button>
+              {result && (
+                <span className={`flex items-center gap-1 text-xs ${result.ok ? "text-[var(--color-success)]" : "text-[var(--color-danger)]"}`}>
+                  {result.ok ? <CheckCircle2 size={14} /> : <XCircle size={14} />}
+                  {result.ok ? "Connected — ready to generate" : result.error}
+                </span>
+              )}
+            </div>
+          </Step>
+
+          <details className="rounded-lg border border-[var(--color-border-subtle)] px-3 py-2">
+            <summary className="text-[var(--color-text-dim)] cursor-pointer hover:text-[var(--color-text-secondary)] select-none">
+              Slower laptop? Try the lightweight option (Z-Image-Turbo, ~4 GB VRAM, 3 files)
+            </summary>
+            <div className="mt-2 flex flex-col gap-1.5">
+              <p>Download these three files into your <code className="font-mono">sd-server</code> folder:</p>
+              <ul className="list-disc pl-5 space-y-0.5">
+                <li>Diffusion model:{" "}
+                  <a href="https://huggingface.co/leejet/Z-Image-Turbo-GGUF" target="_blank" rel="noreferrer" className="text-[var(--color-gold-bright)] underline">
+                    z_image_turbo-Q4_K_M.gguf <ExternalLink size={10} className="inline" />
+                  </a>
+                </li>
+                <li>VAE:{" "}
+                  <a href="https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/ae.sft" target="_blank" rel="noreferrer" className="text-[var(--color-gold-bright)] underline">
+                    ae.sft <ExternalLink size={10} className="inline" />
+                  </a>
+                </li>
+                <li>Text encoder:{" "}
+                  <a href="https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF" target="_blank" rel="noreferrer" className="text-[var(--color-gold-bright)] underline">
+                    Qwen3-4B-Instruct-2507-Q4_K_M.gguf <ExternalLink size={10} className="inline" />
+                  </a>
+                </li>
+              </ul>
+              <pre className="mt-1 rounded-lg bg-[var(--color-bg-surface)] border border-[var(--color-border-glass)] px-3 py-2 text-[10px] font-mono overflow-x-auto whitespace-pre">
+{`./sd-server --diffusion-model z_image_turbo-Q4_K_M.gguf \\
+  --vae ae.sft --llm Qwen3-4B-Instruct-2507-Q4_K_M.gguf \\
+  --cfg-scale 1.0 --diffusion-fa --offload-to-cpu`}</pre>
+              <p className="text-[var(--color-text-dim)]">
+                Z-Image-Turbo is built for speed (few steps), so it generates faster than SDXL on modest hardware.
+              </p>
+            </div>
+          </details>
+        </div>
+      )}
+    </div>
+  );
+}
+
+function Step({ n, title, children }: { n: number; title: string; children: React.ReactNode }) {
+  return (
+    <div className="flex flex-col gap-1">
+      <p className="font-medium text-[var(--color-text-primary)]">
+        <span className="inline-flex items-center justify-center w-5 h-5 rounded-full bg-[var(--color-gold-bright)] text-[var(--color-bg-deep)] text-[10px] font-bold mr-1.5 align-middle">
+          {n}
+        </span>
+        {title}
+      </p>
+      <div className="pl-[26px]">{children}</div>
+    </div>
+  );
+}

src/components/SettingsPanel.tsx

@@ -10,7 +10,7 @@ interface LlmConfig {
   temperature: number;
   max_tokens: number;
   top_p: number;
-  image_model: string;
+  image_api_url: string;
   embed_model: string;
 }
 
@@ -30,6 +30,9 @@ export function SettingsPanel() {
   const [conn, setConn] = useState<{ ok: boolean; models: string[]; error: string } | null>(null);
   const [dataDir, setDataDir] = useState<DataDirInfo | null>(null);
   const [relocating, setRelocating] = useState(false);
+  // ponytail: probe the sd-server so a non-technical DM gets a green check, same pattern as the LLM test.
+  const [imgTesting, setImgTesting] = useState(false);
+  const [imgConn, setImgConn] = useState<{ ok: boolean; error: string } | null>(null);
   // ponytail: advanced SQLite viewer — read-only, two DBs (lore.db / generations.db).
   const [sqlDb, setSqlDb] = useState<"lore" | "generations">("generations");
   const [sqlInput, setSqlInput] = useState("SELECT name, type FROM sqlite_master WHERE type='table' ORDER BY name");
@@ -153,6 +156,24 @@ export function SettingsPanel() {
     setTesting(false);
   }
 
+  // ponytail: probe the configured sd-server. Save first so it uses the latest URL.
+  async function testImage() {
+    if (!config) return;
+    setImgTesting(true);
+    setImgConn(null);
+    try {
+      await invoke("set_llm_config", { config });
+      const r = await invoke<{ ok: boolean; error: string }>("test_image_connection");
+      setImgConn(r);
+      if (r.ok) addToast("sd-server reachable", "success");
+      else addToast(`Image server not reachable: ${r.error}`, "error");
+    } catch (e) {
+      setImgConn({ ok: false, error: String(e) });
+      addToast(`Image server not reachable: ${e}`, "error");
+    }
+    setImgTesting(false);
+  }
+
   // ponytail: provider presets fill in the API URL pattern + default model.
   const PRESETS: { label: string; url: string; model: string; key?: boolean }[] = [
     { label: "Ollama", url: "http://localhost:11434", model: "llama3.2" },
@@ -177,7 +198,7 @@ export function SettingsPanel() {
     temperature: 0.7,
     max_tokens: 512,
     top_p: 0.9,
-    image_model: "x/flux2-klein:4b",
+    image_api_url: "http://localhost:1234",
     embed_model: "nomic-embed-text",
   };
   function resetDefaults() {
@@ -386,17 +407,39 @@ export function SettingsPanel() {
         </div>
       )}
 
-      <h3 className="font-heading text-[var(--color-gold-bright)] text-xs font-semibold uppercase tracking-wider border-b border-[var(--color-border-subtle)] pb-1 mb-1">Image Model</h3>
+      <h3 className="font-heading text-[var(--color-gold-bright)] text-xs font-semibold uppercase tracking-wider border-b border-[var(--color-border-subtle)] pb-1 mb-1">Image Server</h3>
       <div className="flex flex-col gap-1">
         <label className="text-[var(--color-text-secondary)] text-xs font-medium">
-          Image model
+          sd-server URL
         </label>
+        <p className="text-[10px] text-[var(--color-text-dim)] leading-relaxed">
+          Base URL of a <a href="https://github.com/leejet/stable-diffusion.cpp" target="_blank" rel="noreferrer" className="underline">stable-diffusion.cpp</a> `sd-server` (A1111-compatible API). Start it with your chosen model, e.g.<br /><code className="font-mono">sd-server -m model.safetensors</code>
+        </p>
         <input
           className="rounded-lg bg-[var(--color-bg-surface)] border border-[var(--color-border-glass)] px-3 py-2 text-[var(--color-text-primary)] placeholder-[var(--color-text-dim)] focus:outline-none focus:border-[var(--color-gold-bright)]"
-          value={config.image_model}
-          onChange={(e) => setConfig({ ...config, image_model: e.target.value })}
-          placeholder="x/flux2-klein:4b"
+          value={config.image_api_url}
+          onChange={(e) => setConfig({ ...config, image_api_url: e.target.value })}
+          placeholder="http://localhost:1234"
         />
+        {/* ponytail: mirror the LLM test button so a non-technical DM gets a green check */}
+        <div className="flex items-center gap-2 mt-1">
+          <button
+            type="button"
+            onClick={testImage}
+            disabled={imgTesting}
+            className="rounded-lg bg-[var(--color-bg-surface)] border border-[var(--color-border-glass)] px-3 py-1.5 text-xs font-medium text-[var(--color-text-secondary)] hover:border-[var(--color-gold-bright)] hover:text-[var(--color-gold-bright)] transition-colors cursor-pointer disabled:opacity-50"
+          >
+            {imgTesting ? "Testing…" : "Test connection"}
+          </button>
+          {imgConn && (
+            <span className={`text-xs ${imgConn.ok ? "text-[var(--color-success)]" : "text-[var(--color-danger)]"}`}>
+              {imgConn.ok ? "✓ sd-server reachable" : `✕ ${imgConn.error}`}
+            </span>
+          )}
+        </div>
+        <p className="text-[10px] text-[var(--color-text-dim)]">
+          New to this? Open the <strong>Image</strong> tab → "First time? Set up image generation" for a full step-by-step.
+        </p>
       </div>
 
       <h3 className="font-heading text-[var(--color-gold-bright)] text-xs font-semibold uppercase tracking-wider border-b border-[var(--color-border-subtle)] pb-1 mb-1">Embedding Model (Lore RAG)</h3>