feat(diffusion): native FLUX.2-dev text-to-image + /v1/images serving + GUI

Add the `lumen-diffusion` crate: an MLX-native FLUX.2 text-to-image pipeline,
fully separate from the autoregressive LLM stack. Components (each parity-
validated against mflux / Swift references):

- VAE decoder, Mistral-Small-3.2 text encoder (layers 10/20/30 -> [1,512,15360]),
  32B rectified-flow DiT (8 double + 48 single stream, 4D RoPE, AdaLN, guidance),
  flow-match Euler scheduler, dev pipeline, and the FLUX.2 prompt tokenizer.
- Both 4-bit (36 GB) and non-quantized bf16 (128 GB) weights are auto-detected
  per tensor (.scales presence), so the same loaders serve either footprint.

Serving (lumen-server):
- POST /v1/images/generations (OpenAI-compatible, base64 PNG).
- LUMEN_SERVE = chat | image | hybrid serve modes; image mode leaves the MLX
  memory governors at their defaults (the ~31 GB diffusion working set exceeds
  the LLM-tuned wired cap, which would OOM the load).

Desktop app (lumen-app):
- IMAGE-MODELS catalog card with a diagonal "IMAGE" corner badge, an image-
  generation panel (prompt / size / steps / seed / guidance), and image-mode
  launch wiring (+ en/ko strings).

Housekeeping:
- Resolve all FLUX.2 component repos from the local HuggingFace cache via
  `hf_cache` (no machine-specific hardcoded paths).
- Fix safe pre-existing warnings (unused imports / mut, minijinja serde feature,
  toml dep semver metadata).

chore(release): v0.11.0

Author: rabbitson87

Cargo.lock

@@ -4,6 +4,7 @@ members = [
     "crates/turboquant-cache",
     "crates/lumen-metal",
     "crates/lumen-mlx",
+    "crates/lumen-diffusion",
     "crates/lumen-model",
     "crates/lumen-server",
     "crates/paged-attention",
@@ -18,6 +19,7 @@ default-members = [
     "crates/turboquant-cache",
     "crates/lumen-metal",
     "crates/lumen-mlx",
+    "crates/lumen-diffusion",
     "crates/lumen-model",
     "crates/lumen-server",
     "crates/paged-attention",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lumen-app"
-version = "0.10.1"
+version = "0.11.0"
 edition.workspace = true
 rust-version.workspace = true
 license.workspace = true
@@ -36,7 +36,7 @@ reqwest = { version = "0.13.3", default-features = false, features = [
 # Config persistence (TOML at ~/Library/Application Support/Lumen/config.toml)
 serde = { workspace = true }
 serde_json = { workspace = true }
-toml = "1.1.2+spec-1.1.0"
+toml = "1.1.2"
 
 # Model registry — local scan only. Downloads use raw HTTPS via reqwest.
 walkdir = "2.5.0"

crates/lumen-app/Cargo.toml

@@ -1,7 +1,7 @@
 {
   "name": "lumen-app-frontend",
   "private": true,
-  "version": "0.10.1",
+  "version": "0.11.0",
   "type": "module",
   "scripts": {
     "predev": "cargo build --manifest-path ../../../Cargo.toml -p lumen-server --release",

crates/lumen-app/frontend/package.json

@@ -25,6 +25,7 @@
   import DoctorPanel from "./lib/DoctorPanel.svelte";
   import UpdatePanel from "./lib/UpdatePanel.svelte";
   import ApiTabs from "./lib/ApiTabs.svelte";
+  import ImagePanel from "./lib/ImagePanel.svelte";
   import { bytes, duration } from "./lib/format";
   import { i18n, t, LOCALE_LABELS, type Locale } from "./lib/i18n.svelte";
 
@@ -223,6 +224,24 @@
     return models.find((m) => m.id === config!.active_model) ?? null;
   });
 
+  // Text-to-image catalog entries. `image_models` is `#[serde(default)]` on the
+  // Rust side so it may be absent against an older server binary — default [].
+  let imageModels = $derived(catalog.image_models ?? []);
+
+  // The active model id, if it is a diffusion image model. Drives both the
+  // image-models card's ✓ state and the generation panel's visibility.
+  let activeImageModelId = $derived.by(() => {
+    const id = config?.active_model;
+    if (!id) return null;
+    return imageModels.some((m) => m.id === id) ? id : null;
+  });
+
+  // Show the generation panel only when an image model is active AND the
+  // server is up (image mode serves POST /v1/images/generations).
+  let imagePanelVisible = $derived(
+    activeImageModelId != null && status.state === "running",
+  );
+
   // Memory the model can realistically claim, reclaim-aware. macOS evicts /
   // swap-compresses inactive + compressor + other apps' active pages when a big
   // MLX allocation arrives, so the only genuinely-unavailable memory is WIRED
@@ -1465,6 +1484,49 @@
     {/if}
   </section>
 
+  <!-- IMAGE MODELS (text-to-image / diffusion) -->
+  {#if imageModels.length > 0}
+  <section class="{cardBase} col-span-3">
+    <h2 class={cardH2}>{t("imageModels.title")} <span class="dim">{t("imageModels.titleHint")}</span></h2>
+    <div class="flex flex-col gap-2">
+      {#each imageModels as im}
+        {@const isActive = config?.active_model === im.id}
+        {@const fits = !systemInfo || im.min_ram_gb <= systemInfo.ram_gb}
+        <div
+          class={`image-card relative overflow-hidden flex flex-col gap-1.5 px-3 py-2.5 rounded-md bg-panel-2 border ${
+            isActive
+              ? "border-accent shadow-[0_0_0_1px_var(--color-accent)_inset] bg-accent/6"
+              : "border-transparent hover:border-border"
+          }`}
+        >
+          <!-- Diagonal corner ribbon marking this as an image / diffusion model -->
+          <span class="image-ribbon" aria-hidden="true">{t("imageModels.badge")}</span>
+          <div class="flex items-center gap-2 pl-5">
+            <span class="mono text-ok font-bold w-4">{isActive ? "✓" : ""}</span>
+            <div class="min-w-0 flex-1">
+              <div class="text-[13px] text-text">{im.label}</div>
+              <div class="dim mono text-[11px]">{im.id} · {im.approx_size_gb}GB · ≥{im.min_ram_gb}GB RAM</div>
+            </div>
+            <button
+              class="primary"
+              onclick={() => setActive(im.id)}
+              disabled={isActive}
+              title={fits ? "" : t("imageModels.action.title.overBudget")}
+            >{isActive ? t("imageModels.active") : t("action.use")}</button>
+          </div>
+          {#if !fits}
+            <div class="text-warn text-[11px] leading-normal">⚠ {t("imageModels.overBudget")}</div>
+          {/if}
+          <div class="dim text-[11px] leading-normal">{im.notes}</div>
+        </div>
+      {/each}
+    </div>
+    {#if activeImageModelId != null && status.state !== "running"}
+      <div class="dim mt-2.5 text-[11px] leading-normal">{t("imageModels.startHint")}</div>
+    {/if}
+  </section>
+  {/if}
+
   <!-- SERVER -->
   <section class="{cardBase} col-span-3">
     <h2 class={cardH2}>{t("server.title")}</h2>
@@ -1591,6 +1653,12 @@
     {/if}
   </section>
 
+  <!-- IMAGE GENERATION PANEL — only when a diffusion model is active and the
+       server is running in image mode (serves POST /v1/images/generations). -->
+  {#if imagePanelVisible && activeImageModelId}
+    <ImagePanel modelId={activeImageModelId} />
+  {/if}
+
   {/if}
 
   {#if activeTab === "debug"}
@@ -1959,4 +2027,41 @@
   details.mem-explainer[open] > summary::before {
     content: "▾ ";
   }
+
+  /* Diagonal "IMAGE" corner ribbon for text-to-image / diffusion model cards.
+     Classic technique: a wide banner pinned beyond the top-left corner and
+     rotated -45° so it crosses the corner at a diagonal. The parent card sets
+     `position: relative; overflow: hidden;` (Tailwind `relative overflow-hidden`)
+     so the banner is clipped to a neat triangle. Colors come from the app's
+     accent token — no hardcoded brand values. */
+  .image-ribbon {
+    position: absolute;
+    top: 9px;
+    left: -30px;
+    width: 96px;
+    transform: rotate(-45deg);
+    transform-origin: center;
+    text-align: center;
+    background: var(--accent);
+    color: var(--bg);
+    font-size: 8px;
+    font-weight: 700;
+    letter-spacing: 0.12em;
+    line-height: 14px;
+    text-transform: uppercase;
+    box-shadow: 0 1px 3px rgba(0, 0, 0, 0.45);
+    pointer-events: none;
+    user-select: none;
+    z-index: 1;
+  }
+
+  /* When this image model is the selected / active one, brighten the ribbon
+     so the active image model is unmistakably distinguished. */
+  .image-card.border-accent .image-ribbon {
+    background: var(--accent);
+    color: var(--bg);
+    box-shadow:
+      0 0 0 1px var(--accent),
+      0 1px 4px rgba(0, 0, 0, 0.55);
+  }
 </style>