Upload 7 files

app.py

@@ -14,18 +14,18 @@ pipe         = None
 def load_model():
     global pipe, MODEL_LOADED, LOAD_ERROR
     try:
-        print("📥 Loading virtual try-on model...")
+        print("📥 Loading model on CPU...")
         from diffusers import StableDiffusionInpaintPipeline
         pipe = StableDiffusionInpaintPipeline.from_pretrained(
             "runwayml/stable-diffusion-inpainting",
-            torch_dtype=torch.float32,
+            torch_dtype=torch.float32,   # CPU needs float32
             safety_checker=None,
             requires_safety_checker=False,
         )
+        # CPU ONLY — no .to("cuda"), no cpu_offload
         pipe.enable_attention_slicing()
-        pipe.enable_sequential_cpu_offload() if not torch.cuda.is_available() else pipe.to("cuda")
         MODEL_LOADED = True
-        print("✅ Model ready!")
+        print("✅ Model ready on CPU!")
     except Exception as e:
         LOAD_ERROR = str(e)
         print(f"❌ {e}")
@@ -37,14 +37,13 @@ def pil_to_b64(img):
     img.save(buf, format="PNG")
     return base64.b64encode(buf.getvalue()).decode()
 
-def make_upper_body_mask(size):
+def make_mask(size):
     w, h = size
     mask = Image.new("L", size, 0)
     draw = ImageDraw.Draw(mask)
-    draw.ellipse([w*0.3, h*0.08, w*0.7, h*0.22], fill=255)   # neck/head area avoid
-    draw.rectangle([w*0.05, h*0.20, w*0.95, h*0.65], fill=255) # torso
-    draw.rectangle([w*0.0,  h*0.20, w*0.15, h*0.60], fill=255) # left arm
-    draw.rectangle([w*0.85, h*0.20, w*1.0,  h*0.60], fill=255) # right arm
+    draw.rectangle([w*0.05, h*0.18, w*0.95, h*0.68], fill=255)
+    draw.rectangle([w*0.0,  h*0.18, w*0.15, h*0.58], fill=255)
+    draw.rectangle([w*0.85, h*0.18, w*1.0,  h*0.58], fill=255)
     return mask.convert("RGB")
 
 @app.get("/", response_class=HTMLResponse)
@@ -58,24 +57,14 @@ async def status():
 @app.post("/tryon")
 async def tryon(person: UploadFile = File(...), garment: UploadFile = File(...)):
     if not MODEL_LOADED:
-        return JSONResponse({"status":"loading","message":"Model loading, retry in 30s"}, status_code=503)
+        return JSONResponse({"status":"loading","message":"Model still loading, please wait and retry."}, status_code=503)
     try:
         SIZE = (512, 768)
-        person_img  = Image.open(io.BytesIO(await person.read())).convert("RGB").resize(SIZE)
-        garment_img = Image.open(io.BytesIO(await garment.read())).convert("RGB").resize((SIZE[0], SIZE[0]))
-        mask_img    = make_upper_body_mask(SIZE)
+        person_img = Image.open(io.BytesIO(await person.read())).convert("RGB").resize(SIZE)
+        mask_img   = make_mask(SIZE)
 
-        # Build a prompt from garment image colors
-        g = np.array(garment_img)
-        dominant = g.mean(axis=(0,1)).astype(int)
-        color_hint = f"rgb({dominant[0]},{dominant[1]},{dominant[2]})"
-
-        prompt = (
-            "A person wearing a clean fitted garment, "
-            "photorealistic fashion photo, high quality, "
-            "same pose, same background, sharp details"
-        )
-        negative = "deformed, blurry, bad anatomy, extra limbs, watermark, text"
+        prompt   = "Person wearing a clean stylish garment, photorealistic, high quality fashion photo, same pose, same background"
+        negative = "nude, deformed, blurry, bad anatomy, extra limbs, watermark, logo, text, disfigured"
 
         result = pipe(
             prompt=prompt,
@@ -84,9 +73,9 @@ async def tryon(person: UploadFile = File(...), garment: UploadFile = File(...))
             mask_image=mask_img,
             height=SIZE[1],
             width=SIZE[0],
-            num_inference_steps=30,
+            num_inference_steps=25,
             guidance_scale=7.5,
-            strength=0.99,
+            strength=0.95,
         ).images[0]
 
         return JSONResponse({"status":"ok","image": pil_to_b64(result)})