Update app.py

app.py

@@ -222,6 +222,26 @@ SPACES: Dict[str, Dict[str, Any]] = {
 # ---------------------------------
 # Utility
 # ---------------------------------
+def _openpose_canvas_from_image(img_rgb: Image.Image) -> Image.Image:
+    try:
+        if PKG.get("OpenposeDetector") is None:
+            return Image.new("RGB", img_rgb.size, "black")
+        det = PKG["OpenposeDetector"]()
+        pose = det(img_rgb)
+        return pose.convert("RGB").resize(img_rgb.size)
+    except Exception:
+        return Image.new("RGB", img_rgb.size, "black")
+
+def _blend_pose_canvases(orig_pose: Image.Image, tpose: Image.Image, alpha: float = 0.4) -> Image.Image:
+    alpha = max(0.0, min(1.0, float(alpha)))
+    if orig_pose.size != tpose.size:
+        tpose = tpose.resize(orig_pose.size)
+    return Image.blend(orig_pose, tpose, alpha).convert("RGB")
+
+def _mean_brightness(img: Image.Image) -> float:
+    import numpy as np
+    return float(np.asarray(img.convert("L"), dtype=np.uint8).mean())
+
 def _save_png(img: Image.Image, path: Union[str,Path]) -> str:
     p = Path(path); p.parent.mkdir(parents=True, exist_ok=True); img.save(p); return str(p)
 
@@ -635,50 +655,47 @@ def step1_gpu_refine(
 
     # ---- T-포즈 (ControlNet/OpenPose)
     if enforce_tpose:
-        try:
-            from diffusers import ControlNetModel, StableDiffusionControlNetImg2ImgPipeline, EulerAncestralDiscreteScheduler
-
-            controlnet = ControlNetModel.from_pretrained(
-                "lllyasviel/control_v11p_sd15_openpose", torch_dtype=dtype
-            )
-            pipe_pose = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
-                "runwayml/stable-diffusion-v1-5",
-                controlnet=controlnet,
-                torch_dtype=dtype
-            )
-            # 안정 스케줄러
-            try:
-                pipe_pose.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe_pose.scheduler.config)
-            except Exception:
-                pass
-            if dev == "cuda":
-                pipe_pose.to("cuda")
-
-            img_rgb = _resize_to_multiple(img.convert("RGB"), multiple=8, max_side=512)
-            pose_canvas = _make_tpose_canvas_like(img_rgb).resize(img_rgb.size)
-            pose_canvas.save(OUT/"step1"/"dbg_02_pose_canvas.png")
-
-            POS = "T-pose, full body, same outfit, same colors, clean anime lines, consistent scale, white background"
-            NEG = "extra limbs, deformed hands, melted face, disfigured, distorted body, artifacts, blurry, noisy, watermark, text"
-
-            out = pipe_pose(
-                prompt=POS,
-                negative_prompt=NEG,
-                image=img_rgb,
-                control_image=pose_canvas,
-                strength=float(tpose_strength),
-                guidance_scale=float(tpose_guidance),
-                num_inference_steps=int(tpose_steps),
-                controlnet_conditioning_scale=0.45,
-                control_guidance_start=[0.0],
-                control_guidance_end=[0.6],
-            ).images[0]
-
-            img = out.convert("RGBA")
+        from diffusers import ControlNetModel, StableDiffusionControlNetImg2ImgPipeline, EulerAncestralDiscreteScheduler
+        import torch
+        dev = "cuda" if torch.cuda.is_available() else "cpu"
+        dtype = torch.float16 if dev == "cuda" else torch.float32
+    
+        controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_openpose", torch_dtype=dtype)
+        pipe_pose  = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
+            "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=dtype
+        )
+        try: pipe_pose.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe_pose.scheduler.config)
+        except: pass
+        if dev == "cuda": pipe_pose.to("cuda")
+    
+        img_rgb    = _resize_to_multiple(img.convert("RGB"), multiple=8, max_side=512)
+        pose_orig  = _openpose_canvas_from_image(img_rgb)
+        pose_tpose = _draw_tpose_openpose_canvas(size=min(img_rgb.size))  # 기존 T-포즈 가이드
+        pose_canvas= _blend_pose_canvases(pose_orig, pose_tpose, alpha=0.4)
+    
+        out = pipe_pose(
+            prompt="T-pose tendency, full body, same outfit and colors, clean anime lines, plain light background",
+            negative_prompt="extra limbs, deformed, melted face, distorted body, watermark, text, noisy",
+            image=img_rgb,
+            control_image=pose_canvas,
+            strength=float(tpose_strength),          # (클램프됨)
+            guidance_scale=float(tpose_guidance),
+            num_inference_steps=int(tpose_steps),
+            controlnet_conditioning_scale=0.30,      # ★ 낮춤
+            control_guidance_start=[0.0],
+            control_guidance_end=[0.50],             # ★ 중반까지만
+            guess_mode=True                          # ★ 배경/미세영역 과제어 방지
+        ).images[0]
+    
+        img = out.convert("RGBA")
+        if _mean_brightness(img) < 12:
+            logs.append("T-포즈 결과가 너무 어두움 → 원본으로 롤백")
+            img = Image.open(s1_path).convert("RGBA")
+        else:
+            (OUT/"step1"/"dbg_pose_orig.png").write_bytes(pose_orig.tobytes()) if False else None
             img.save(OUT/"step1"/"dbg_03_after_tpose.png")
-            logs.append("ControlNet(OpenPose) T-포즈 적용(안정화 설정)")
-        except Exception as e:
-            logs.append(f"T-포즈 ControlNet 실패: {e}")
+            logs.append("ControlNet(OpenPose) T-포즈(블렌드) 적용")
+
 
     # ---- (옵션) 리드로우(img2img)
     if do_redraw_flag: