Add /encode_image endpoint for compositional image cues

Same model and preprocessing as the dataset features (ViT-B-32 /
laion2b_s34b_b79k), so user-uploaded images can be compared directly
to the static site's image features in the same vector space.

Switched to gr.TabbedInterface so both /encode and /encode_image
endpoints are exposed.

app.py

@@ -1,12 +1,13 @@
 import gradio as gr
 import open_clip
 import torch
+from PIL import Image
 
 MODEL_NAME = "ViT-B-32"
 PRETRAINED = "laion2b_s34b_b79k"
 
 print(f"Loading {MODEL_NAME} / {PRETRAINED}...")
-model, _, _ = open_clip.create_model_and_transforms(MODEL_NAME, pretrained=PRETRAINED)
+model, _, preprocess = open_clip.create_model_and_transforms(MODEL_NAME, pretrained=PRETRAINED)
 model.eval()
 tokenizer = open_clip.get_tokenizer(MODEL_NAME)
 print("Model loaded.")
@@ -28,12 +29,54 @@ def encode(text: str):
     }
 
 
-demo = gr.Interface(
+def encode_image(image):
+    """Encode an image into a 512-d L2-normalized CLIP embedding.
+
+    Same vector space as encode(text), so text-image cosine similarity works.
+    Same vector space as the dataset features baked into the static site
+    (ViT-B-32 / laion2b_s34b_b79k), so user-uploaded image cues can be
+    compared directly to those features for compositional retrieval.
+    """
+    if image is None:
+        return {"error": "no image", "embedding": None}
+    if not isinstance(image, Image.Image):
+        try:
+            image = Image.fromarray(image)
+        except Exception as e:
+            return {"error": f"unsupported image input: {type(image).__name__}: {e}", "embedding": None}
+    image = image.convert("RGB")
+    with torch.no_grad():
+        x = preprocess(image).unsqueeze(0)
+        feat = model.encode_image(x)
+        feat = feat / feat.norm(dim=-1, keepdim=True)
+    return {
+        "model": MODEL_NAME,
+        "pretrained": PRETRAINED,
+        "dim": feat.shape[-1],
+        "embedding": feat[0].tolist(),
+    }
+
+
+text_iface = gr.Interface(
     fn=encode,
     inputs=gr.Textbox(label="Query", placeholder="a hotel bathroom with a walk-in shower"),
     outputs=gr.JSON(label="CLIP text embedding"),
-    title="CLIP text encoder (ViT-B-32 / LAION-2B)",
-    description="Returns a 512-d L2-normalized text embedding for CLIP-guided image search.",
+    title="CLIP text encoder",
+    description="Returns a 512-d L2-normalized text embedding (ViT-B-32 / LAION-2B). API endpoint: /encode",
+)
+
+image_iface = gr.Interface(
+    fn=encode_image,
+    inputs=gr.Image(type="pil", label="Image"),
+    outputs=gr.JSON(label="CLIP image embedding"),
+    title="CLIP image encoder",
+    description="Returns a 512-d L2-normalized image embedding (ViT-B-32 / LAION-2B). API endpoint: /encode_image",
+)
+
+demo = gr.TabbedInterface(
+    [text_iface, image_iface],
+    ["Text", "Image"],
+    title="CLIP encoder (ViT-B-32 / LAION-2B)",
 )
 
-demo.launch()
+demo.launch()