Update app/services/image_service.py

app/services/image_service.py

@@ -14,31 +14,31 @@ logger = logging.getLogger(__name__)
 
 # --- Virtual Try-On Instruction Template (Updated for 3 Inputs) ---
 VIRTUAL_TRY_ON_PROMPT = """
-You are an expert image-editing agent. Perform a high-fidelity virtual try-on using three inputs:
+You are an expert image synthesis agent. Your PRIMARY and MANDATORY directive is to generate a high-fidelity virtual try-on image using the three provided inputs.
 
-- Input 1 (dress reference): the exact garment image to be transferred.
-- Input 2 (person): the target person who must wear the garment.
-- Input 3 (shoes reference): the footwear to be worn.
+- Input 1 (dress reference): the garment to be transferred.
+- Input 2 (person): the target person.
+- Input 3 (shoes reference): the footwear.
+
+CRITICAL: You MUST generate an image output in every single response. Do not refuse to generate an image based on difficulty or perspective. Make your best attempt at a photorealistic composition.
 
 Positive instructions (Do):
-1. Produce exactly one photorealistic output image showing the person from Input 2 wearing BOTH the dress from Input 1 and the shoes from Input 3.
-2. Preserve the person's face, facial features, expression, hair, gender, skin tone, and body proportions — do NOT change identity.
-3. Reproduce the dress design, color, pattern, fabric texture, and major details (buttons, seams, prints) faithfully.
-4. Reproduce the shoes faithfully and ensure they fit the person's stance naturally (grounding, shadows).
-5. Match lighting, shadows, perspective, and scale so all items appear naturally worn by the person in the original scene.
-6. Also return a short text JSON summary with the form: {"success": true|false, "notes": "explain any limitations or changes"}.
+1. Fuse these inputs into a single photorealistic image showing the person (Input 2) wearing the dress (Input 1) and shoes (Input 3).
+2. Preserve the person's identity (face, hair, body shape) as closely as possible.
+3. Retain the exact textures and details of the dress and shoes.
+4. Adapt the lighting and shadows so the items look naturally integrated.
+5. If the pose is complex, prioritize completing the image generation over perfect physics—visual output is required.
+6. Return a short JSON summary alongside the image.
 
 Negative instructions (Do NOT):
-1. Do NOT modify the person's face, gender, or identifying features.
-2. Do NOT change the dress/shoe color, pattern, or main texture.
-3. Do NOT add extra clothing items, logos, watermarks, offensive symbols, or unrelated props.
-4. Do NOT produce empty responses, placeholders, or images containing text overlays.
-5. Do NOT blur, heavily distort, or crop important parts of the person, dress, or shoes.
+1. Do NOT return only text. An image is required.
+2. Do NOT change the person's face or gender.
+3. Do NOT add watermarks, text overlays, or multiple angles (collages).
+4. Do NOT alter the core design of the dress or shoes.
 
 Output requirements:
-- Return a single photorealistic image (same orientation as the person image) and a short text JSON summary.
-- No collages, no multiple thumbnail images, no visible watermarks.
-- If the try-on cannot be done reliably, set "success": false and explain why in "notes".
+- MANDATORY: 1 Photorealistic Image file.
+- Secondary: JSON text summary: {"success": true, "notes": "Image generated successfully"}.
 """