Show bounding boxes on the sample image card if we have them in the prompt

ui/src/components/BoundingBoxOverlay.tsx

@@ -0,0 +1,78 @@
+import classNames from 'classnames';
+
+// A single bbox parsed from an Ideogram-style caption/prompt. Stored coords are
+// normalized 0-1000 in [y1, x1, y2, x2] order (top-left origin).
+export interface OverlayBox {
+  y1: number;
+  x1: number;
+  y2: number;
+  x2: number;
+  label: string;
+  type: 'obj' | 'text';
+}
+
+// Returns the list of boxes if the text is an Ideogram bbox-JSON caption/prompt
+// with at least one bbox, otherwise null (normal captions/prompts get no overlay).
+export function parseBoundingBoxes(text: string): OverlayBox[] | null {
+  const trimmed = text.trim();
+  if (!trimmed.startsWith('{')) return null;
+  let data: any;
+  try {
+    data = JSON.parse(trimmed);
+  } catch {
+    return null;
+  }
+  const elements = data?.compositional_deconstruction?.elements;
+  if (!Array.isArray(elements)) return null;
+  const boxes: OverlayBox[] = [];
+  for (const el of elements) {
+    const bb = el?.bbox;
+    if (Array.isArray(bb) && bb.length === 4 && bb.every((n: any) => typeof n === 'number')) {
+      const isText = el.type === 'text';
+      const label = (isText ? el.text : el.desc) ?? '';
+      boxes.push({ y1: bb[0], x1: bb[1], y2: bb[2], x2: bb[3], label: `${label}`, type: isText ? 'text' : 'obj' });
+    }
+  }
+  return boxes.length > 0 ? boxes : null;
+}
+
+// Absolute overlay layer of boxes. Place inside a `relative` container that wraps
+// the image (e.g. inside react-zoom-pan-pinch's TransformComponent) so boxes
+// track the image during zoom/pan. Coords are percentages of the frame, so this
+// is resolution-independent. pointer-events-none keeps zoom/pan working through it.
+export default function BoundingBoxOverlay({ boxes }: { boxes: OverlayBox[] }) {
+  return (
+    <div className="absolute inset-0 pointer-events-none">
+      {boxes.map((b, i) => (
+        <div
+          key={i}
+          className={classNames('absolute border', {
+            'border-cyan-400': b.type === 'obj',
+            'border-amber-400': b.type === 'text',
+          })}
+          style={{
+            left: `${b.x1 / 10}%`,
+            top: `${b.y1 / 10}%`,
+            width: `${(b.x2 - b.x1) / 10}%`,
+            height: `${(b.y2 - b.y1) / 10}%`,
+          }}
+        >
+          {b.label && (
+            <span
+              title={b.label}
+              className={classNames(
+                'absolute top-0 left-0 max-w-full px-1 py-0.5 text-[9px] leading-tight font-medium whitespace-pre-line break-words line-clamp-3 text-gray-900',
+                {
+                  'bg-cyan-400/90': b.type === 'obj',
+                  'bg-amber-400/90': b.type === 'text',
+                },
+              )}
+            >
+              {b.label}
+            </span>
+          )}
+        </div>
+      ))}
+    </div>
+  );
+}

ui/src/components/DatasetImageViewer.tsx

@@ -10,42 +10,7 @@ import { apiClient } from '@/utils/api';
 import { isVideo, isAudio } from '@/utils/basic';
 import AudioPlayer from './AudioPlayer';
 import { TransformWrapper, TransformComponent } from 'react-zoom-pan-pinch';
-
+import BoundingBoxOverlay, { parseBoundingBoxes } from './BoundingBoxOverlay';
-// A single bbox parsed from an Ideogram-style caption. Stored coords are
-// normalized 0-1000 in [y1, x1, y2, x2] order (top-left origin).
-interface OverlayBox {
-  y1: number;
-  x1: number;
-  y2: number;
-  x2: number;
-  label: string;
-  type: 'obj' | 'text';
-}
-
-// Returns the list of boxes if the caption is an Ideogram bbox-JSON caption with
-// at least one bbox, otherwise null (normal captions get no overlay).
-function parseBoundingBoxes(caption: string): OverlayBox[] | null {
-  const trimmed = caption.trim();
-  if (!trimmed.startsWith('{')) return null;
-  let data: any;
-  try {
-    data = JSON.parse(trimmed);
-  } catch {
-    return null;
-  }
-  const elements = data?.compositional_deconstruction?.elements;
-  if (!Array.isArray(elements)) return null;
-  const boxes: OverlayBox[] = [];
-  for (const el of elements) {
-    const bb = el?.bbox;
-    if (Array.isArray(bb) && bb.length === 4 && bb.every((n: any) => typeof n === 'number')) {
-      const isText = el.type === 'text';
-      const label = (isText ? el.text : el.desc) ?? '';
-      boxes.push({ y1: bb[0], x1: bb[1], y2: bb[2], x2: bb[3], label: `${label}`, type: isText ? 'text' : 'obj' });
-    }
-  }
-  return boxes.length > 0 ? boxes : null;
-}
 
 interface Props {
   imgPath: string | null; // current image path
@@ -373,40 +338,7 @@ export default function DatasetImageViewer({ imgPath, imageList, onChange, refre
                           draggable={false}
                           className="w-auto h-auto max-w-full sm:max-w-[95vw] max-h-[70vh] object-contain select-none !pointer-events-auto"
                         />
-                        {showBoxes && boundingBoxes && (
+                        {showBoxes && boundingBoxes && <BoundingBoxOverlay boxes={boundingBoxes} />}
-                          <div className="absolute inset-0 pointer-events-none">
-                            {boundingBoxes.map((b, i) => (
-                              <div
-                                key={i}
-                                className={classNames('absolute border', {
-                                  'border-cyan-400': b.type === 'obj',
-                                  'border-amber-400': b.type === 'text',
-                                })}
-                                style={{
-                                  left: `${b.x1 / 10}%`,
-                                  top: `${b.y1 / 10}%`,
-                                  width: `${(b.x2 - b.x1) / 10}%`,
-                                  height: `${(b.y2 - b.y1) / 10}%`,
-                                }}
-                              >
-                                {b.label && (
-                                  <span
-                                    title={b.label}
-                                    className={classNames(
-                                      'absolute top-0 left-0 max-w-full px-1 py-0.5 text-[9px] leading-tight font-medium whitespace-pre-line break-words line-clamp-3 text-gray-900',
-                                      {
-                                        'bg-cyan-400/90': b.type === 'obj',
-                                        'bg-amber-400/90': b.type === 'text',
-                                      },
-                                    )}
-                                  >
-                                    {b.label}
-                                  </span>
-                                )}
-                              </div>
-                            ))}
-                          </div>
-                        )}
                       </div>
                     </TransformComponent>
                   </TransformWrapper>

ui/src/components/SampleImageViewer.tsx

@@ -3,13 +3,15 @@ import { useState, useEffect, useMemo, useCallback, useRef } from 'react';
 import { createPortal } from 'react-dom';
 import { Dialog, DialogBackdrop, DialogPanel } from '@headlessui/react';
 import { SampleConfig, SampleItem } from '@/types';
-import { Cog } from 'lucide-react';
+import { Cog, SquareDashed } from 'lucide-react';
+import classNames from 'classnames';
 import { Menu, MenuButton, MenuItem, MenuItems } from '@headlessui/react';
 import { openConfirm } from './ConfirmModal';
 import { apiClient } from '@/utils/api';
 import { isVideo, isAudio } from '@/utils/basic';
 import AudioPlayer from './AudioPlayer';
 import { TransformWrapper, TransformComponent } from 'react-zoom-pan-pinch';
+import BoundingBoxOverlay, { parseBoundingBoxes } from './BoundingBoxOverlay';
 
 interface Props {
   imgPath: string | null; // current image path
@@ -31,6 +33,7 @@ export default function SampleImageViewer({
   const [mounted, setMounted] = useState(false);
   const [isOpen, setIsOpen] = useState(Boolean(imgPath));
   const [showingControlIdx, setShowingControlIdx] = useState<number | null>(null);
+  const [showBoxes, setShowBoxes] = useState<boolean>(false);
 
   useEffect(() => setMounted(true), []);
 
@@ -196,6 +199,20 @@ export default function SampleImageViewer({
     return imgPath;
   }, [showingControlIdx, controlImages, imgPath]);
 
+  // The sample's prompt is what generated it; if it's an Ideogram bbox-JSON we can
+  // overlay the boxes on the generated image. Only on the main image (not controls).
+  const boundingBoxes = useMemo(
+    () => (sampleItem?.prompt ? parseBoundingBoxes(sampleItem.prompt) : null),
+    [sampleItem],
+  );
+  const canShowBoxes = Boolean(
+    boundingBoxes &&
+      showingControlIdx === null &&
+      displayedImgPath &&
+      !isAudio(displayedImgPath) &&
+      !isVideo(displayedImgPath),
+  );
+
   // keyboard events while open
   useEffect(() => {
     const handleKeyDown = (event: KeyboardEvent) => {
@@ -326,12 +343,15 @@ export default function SampleImageViewer({
                     }}
                   >
                     <TransformComponent>
-                      <img
+                      <div className="relative">
-                        src={`/api/img/${encodeURIComponent(displayedImgPath)}`}
+                        <img
-                        alt="Sample Image"
+                          src={`/api/img/${encodeURIComponent(displayedImgPath)}`}
-                        draggable={false}
+                          alt="Sample Image"
-                        className="w-auto h-auto max-w-full sm:max-w-[95vw] max-h-[82vh] object-contain select-none !pointer-events-auto"
+                          draggable={false}
-                      />
+                          className="w-auto h-auto max-w-full sm:max-w-[95vw] max-h-[82vh] object-contain select-none !pointer-events-auto"
+                        />
+                        {showBoxes && canShowBoxes && boundingBoxes && <BoundingBoxOverlay boxes={boundingBoxes} />}
+                      </div>
                     </TransformComponent>
                   </TransformWrapper>
                 ))}
@@ -386,33 +406,48 @@ export default function SampleImageViewer({
                 </div>
               </div>
             </div>
-            <div className="absolute top-2 right-2 bg-gray-900 rounded-full p-1 leading-[0px] opacity-50 hover:opacity-100 z-20">
+            <div className="absolute top-2 right-2 flex items-center gap-2 z-20">
-              <Menu>
+              {canShowBoxes && (
-                <MenuButton>
+                <button
-                  <Cog />
+                  type="button"
-                </MenuButton>
+                  onClick={() => setShowBoxes(v => !v)}
-                <MenuItems
+                  title={showBoxes ? 'Hide bounding boxes' : 'Show bounding boxes'}
-                  anchor="bottom end"
+                  className={classNames('bg-gray-900 rounded-full p-1 leading-[0px] hover:opacity-100', {
-                  className="bg-gray-900 border border-gray-700 rounded shadow-lg w-48 px-2 py-2 mt-1 z-50"
+                    'opacity-100 text-blue-400': showBoxes,
+                    'opacity-50': !showBoxes,
+                  })}
                 >
-                  {imgPath && isAudio(imgPath) && (
+                  <SquareDashed />
+                </button>
+              )}
+              <div className="bg-gray-900 rounded-full p-1 leading-[0px] opacity-50 hover:opacity-100">
+                <Menu>
+                  <MenuButton>
+                    <Cog />
+                  </MenuButton>
+                  <MenuItems
+                    anchor="bottom end"
+                    className="bg-gray-900 border border-gray-700 rounded shadow-lg w-48 px-2 py-2 mt-1 z-50"
+                  >
+                    {imgPath && isAudio(imgPath) && (
+                      <MenuItem>
+                        <a
+                          className="cursor-pointer px-4 py-1 hover:bg-gray-800 rounded block"
+                          href={`/api/img/${encodeURIComponent(imgPath)}`}
+                          download={imgPath.replace(/^.*[\\/]/, '')}
+                        >
+                          Download
+                        </a>
+                      </MenuItem>
+                    )}
                     <MenuItem>
-                      <a
+                      <div className="cursor-pointer px-4 py-1 hover:bg-gray-800 rounded" onClick={handleDelete}>
-                        className="cursor-pointer px-4 py-1 hover:bg-gray-800 rounded block"
+                        Delete Sample
-                        href={`/api/img/${encodeURIComponent(imgPath)}`}
+                      </div>
-                        download={imgPath.replace(/^.*[\\/]/, '')}
-                      >
-                        Download
-                      </a>
                     </MenuItem>
-                  )}
+                  </MenuItems>
-                  <MenuItem>
+                </Menu>
-                    <div className="cursor-pointer px-4 py-1 hover:bg-gray-800 rounded" onClick={handleDelete}>
+              </div>
-                      Delete Sample
-                    </div>
-                  </MenuItem>
-                </MenuItems>
-              </Menu>
             </div>
           </DialogPanel>
         </div>