Fix Blend QR image

app.py

@@ -45,47 +45,63 @@ def make_qr(url: str = "http://www.mybirdfire.com", size: int = 512, border: int
 
 # ========= ControlNet Stylizer (SD1.5 + sd15-canny) =========
 _cn = {"pipe": None}
-def _load_controlnet():
+def _load_controlnet_dual():
     if _cn["pipe"] is None:
         from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
-        from controlnet_aux import CannyDetector
         from diffusers.schedulers.scheduling_euler_discrete import EulerDiscreteScheduler
+        from controlnet_aux import CannyDetector
+
+        canny = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=dtype)
+        tile  = ControlNetModel.from_pretrained("lllyasviel/control_v11f1e_sd15_tile", torch_dtype=dtype)
 
-        controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=dtype)
         pipe = StableDiffusionControlNetPipeline.from_pretrained(
             "runwayml/stable-diffusion-v1-5",
-            controlnet=controlnet,
+            controlnet=[canny, tile],            # <— dual CN
             torch_dtype=dtype,
             safety_checker=None
         ).to(device)
-        # Use Euler to match your sampler preference
+
         pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
         pipe.enable_attention_slicing()
         pipe.enable_vae_slicing()
-        _cn["pipe"] = pipe
-        _cn["canny"] = CannyDetector()
-    return _cn["pipe"], _cn["canny"]
 
-def stylize_qr(prompt: str, negative: str, steps: int, guidance: float, seed: int, canny_low: int, canny_high: int):
-    qr_image = make_qr("http://www.mybirdfire.com", size=512, border=4)
-    pipe, canny = _load_controlnet()
+        _cn["pipe"]  = pipe
+        _cn["cannyx"] = CannyDetector()
+    return _cn["pipe"], _cn["cannyx"]
+
+NEG_DEFAULT = "lowres, low contrast, blurry, washed out, jpeg artifacts, worst quality"
+
+
+def stylize_qr(prompt: str, negative: str, steps: int, guidance: float, seed: int,
+               canny_low: int, canny_high: int):
+    qr_image = make_qr("http://www.mybirdfire.com", size=512, border=6)
+    pipe, canny = _load_controlnet_dual()
+
+    # edge map for canny CN; tile CN uses the raw QR
     edges = canny(qr_image, low_threshold=int(canny_low), high_threshold=int(canny_high))
 
     gen = torch.Generator(device=device).manual_seed(int(seed)) if int(seed) != 0 else None
+
+    # weights per controlnet: [canny_weight, tile_weight]
+    cn_scales = [1.20, 0.60]   # stronger structure, lighter texture
+
     def run():
         return pipe(
             prompt=str(prompt),
-            negative_prompt=negative or "",
-            image=edges,
+            negative_prompt=negative or NEG_DEFAULT,
+            image=[edges, qr_image],
+            controlnet_conditioning_scale=cn_scales,
             num_inference_steps=int(steps),
             guidance_scale=float(guidance),
             generator=gen
         ).images[0]
+
     if device in ("cuda", "mps"):
         with torch.autocast(device):
             return run()
     return run()
 
+
 # ========= UI =========
 with gr.Blocks() as demo:
     gr.Markdown("## Stable Diffusion + QR Code + ControlNet")