feat(#805): MV-Adapter slice 1 — 4-view AI texture bake for image→3D meshes

qml/PropertiesPanel.qml

@@ -2111,9 +2111,17 @@ Rectangle {
                         // image was picked (ready by now — no UI-blocking
                         // captioning here). Empty falls back to a neutral prompt
                         // inside generateMeshTextureMultiView.
+                        //
+                        // Views: front/back/left/right (MV-Adapter #805 slice 1
+                        // — its 6-view orthographic layout minus the two poles).
+                        // 4 equatorial views give full horizontal coverage +
+                        // seam overlap for the baker's cross-view blend, a clear
+                        // step up from the old 2 (front+back left the sides to
+                        // stretch/blur) without the ~3× cost of all 6. The two
+                        // poles (top/bottom) rarely help typical subjects.
                         MaterialEditorQML.generateMeshTextureMultiView(
                             MeshGenController.caption, 512, 512, 0.9,
-                            ["front", "back"],
+                            ["front", "back", "left", "right"],
                             "",                       // no photo pinning
                             mgPbr.checked)
                     }

src/ImageTo3D/ImageCaptioner.cpp

@@ -16,6 +16,7 @@
 #include <mtmd.h>
 #include <mtmd-helper.h>
 
+#include <mutex>
 #include <string>
 #include <vector>
 #endif
@@ -117,6 +118,16 @@
     if (image.isNull() || !modelsPresent())
         return {};
 
+    // Captioning runs on a detached worker thread (MeshGenController), and the
+    // user can pick a new image while a previous caption is still in flight —
+    // spawning a second concurrent call. llama.cpp model load + inference is NOT
+    // safe to run from two threads at once (global ggml backend state); doing so
+    // produced garbage captions (e.g. a person still described from the earlier
+    // image). Serialise: only one caption() runs at a time. The caller's
+    // per-path guard (setCaptionResult) drops any now-stale result.
+    static std::mutex s_captionMutex;
+    std::lock_guard<std::mutex> lock(s_captionMutex);
+
     // ---- Load the text model + its mtmd (vision projector) context ----------
     llama_model_params mparams = llama_model_default_params();
     // Vision projector runs on CPU/accelerator via ggml; the small LM can put a
@@ -168,10 +179,16 @@
         // ---- Build the prompt with the media marker + tokenize --------------
         const QString instruction = prompt.isEmpty()
             ? QString::fromLatin1(kDefaultPrompt) : prompt;
-        // SmolVLM (idefics3) chat format; the media marker is replaced by the
-        // image chunk during tokenization.
+        // SmolVLM (idefics3) chat format, matching the GGUF's embedded Jinja
+        // template EXACTLY. For a user turn whose FIRST content item is an image
+        // the template emits "User:" with NO trailing space (a text-first turn
+        // would use "User: "), then the image, then the text:
+        //   <|im_start|>User:<image>{instruction}<end_of_utterance>\nAssistant:
+        // The media marker is substituted for the real image chunk during
+        // tokenization. (The earlier "User: " + marker form had a stray space
+        // that mis-framed the turn.)
         const std::string text =
-            std::string("<|im_start|>User: ") + mtmd_default_marker()
+            std::string("<|im_start|>User:") + mtmd_default_marker()
             + instruction.toStdString() + "<end_of_utterance>\nAssistant:";
 
         mtmd_input_text itext;
@@ -210,6 +227,8 @@
             char buf[256];
             const int n = llama_token_to_piece(vocab, tok, buf, sizeof(buf), 0, true);
             if (n > 0) out.append(buf, n);
+            // pos == nullptr here → llama_decode auto-assigns positions from the
+            // KV state (seq_pos_max + 1), continuing correctly past the image.
             llama_batch nb = llama_batch_get_one(const_cast<llama_token*>(&tok), 1);
             if (llama_decode(lctx, nb) != 0) break;
         }

src/ImageTo3D/ImageCaptioner.h

@@ -26,16 +26,24 @@
 namespace ImageCaptioner {
 
 // True only when built with ENABLE_LOCAL_LLM.
-// A DETAIL-seeking prompt: SmolVLM-500M defaults to terse answers ("a
-// rabbit"), which makes a weak texture prompt. Explicitly ask for a rich,
-// comma-separated visual description (colours, materials, markings, surface)
-// and forbid the one-word reply, so the caption feeds SD with real texture cues.
+// A DETAIL-seeking prompt: SmolVLM-500M defaults to terse answers ("a rabbit"),
+// which makes a weak texture prompt, so we ask for a rich comma-separated
+// visual description (colours, materials, markings, surface).
+// IMPORTANT: do NOT include a concrete worked EXAMPLE here. SmolVLM-500M is
+// small enough that it parroted a detailed example ("a fluffy brown and white
+// rabbit, …") verbatim as its answer, ignoring the actual image entirely —
+// which is exactly how a person's photo came back described as a rabbit. Keep
+// the instruction abstract so the model has to look at the picture.
 static constexpr const char* kDefaultPrompt =
-    "Describe this object in detail as a comma-separated visual texture prompt: "
-    "its type, all colours, materials, surface texture, patterns and markings. "
-    "Be specific and descriptive, not a single word. Example: 'a fluffy brown "
-    "and white rabbit, soft short fur, pink inner ears, dark eyes'. "
-    "Answer with only the description.";
+    "Describe ONLY the main subject of this image (ignore the background and "
+    "any scenery) as a detailed, comma-separated visual description for "
+    "texturing a 3D model of it. START by naming WHAT the subject is (for "
+    "example: a woman, a man, a rabbit, a car, a chair), then describe its "
+    "surface appearance: for a person or character give skin tone, hair colour "
+    "and style, eye colour, makeup, and clothing colours/materials; for an "
+    "animal or object give its colours, materials, surface texture, and any "
+    "patterns or markings. Be specific, not a single word, and do not mention "
+    "the background. Answer with only the description.";
 
 bool isAvailable();