Update app.py

app.py

@@ -2,10 +2,15 @@ import spaces
 import time
 import os
 
+# ONNX Runtime에서 CUDA provider 우선 사용 (insightface → ORT 세션에 반영)
+os.environ.setdefault("INSIGHTFACE_ONNX_PROVIDERS", "CUDAExecutionProvider,CPUExecutionProvider")
+os.environ.setdefault("ORT_LOG_severity_level", "3")  # ORT 로그 최소화
+
 import gradio as gr
 import torch
 from einops import rearrange
 from PIL import Image
+import numpy as np
 
 from flux.cli import SamplingOptions
 from flux.sampling import denoise, get_noise, get_schedule, prepare, unpack
@@ -160,14 +165,34 @@ def generate_image(
         torch.cuda.empty_cache()
 
     t1 = time.perf_counter()
-
     print(f"Done in {t1 - t0:.1f}s.")
+
     # bring into PIL format
     x = x.clamp(-1, 1)
     x = rearrange(x[0], "c h w -> h w c")
-
-    img = Image.fromarray((127.5 * (x + 1.0)).cpu().byte().numpy())
-    return img, str(opts.seed), flux_generator.pulid_model.debug_img_list
+    img = Image.fromarray((127.5 * (x + 1.0)).cpu().byte().numpy()).convert("RGB")
+
+    # Gallery용 디버그 이미지들을 PIL 리스트로 강제 변환
+    debug = []
+    for it in (flux_generator.pulid_model.debug_img_list or []):
+        try:
+            if isinstance(it, Image.Image):
+                pil = it.convert("RGB")
+            else:
+                if hasattr(it, "detach"):
+                    arr = it.detach().cpu().numpy()
+                else:
+                    arr = np.array(it)
+                if arr.ndim == 3 and arr.shape[0] in (1, 3):  # C,H,W → H,W,C
+                    arr = np.transpose(arr, (1, 2, 0))
+                if arr.dtype != np.uint8:
+                    arr = np.clip(arr, 0, 255).astype(np.uint8)
+                pil = Image.fromarray(arr).convert("RGB")
+            debug.append(pil)
+        except Exception:
+            continue
+
+    return img, str(opts.seed), debug
 
 
 def create_demo(args, model_name: str, device: str = "cuda" if torch.cuda.is_available() else "cpu", offload: bool = False):
@@ -189,7 +214,7 @@ def create_demo(args, model_name: str, device: str = "cuda" if torch.cuda.is_ava
         with gr.Row():
             with gr.Column():
                 prompt = gr.Textbox(label="Prompt", value="portrait, color, cinematic")
-                id_image = gr.Image(label="ID Image")
+                id_image = gr.Image(label="ID Image", type="numpy")
                 id_weight = gr.Slider(0.0, 3.0, 1, step=0.05, label="id weight")
 
                 width = gr.Slider(256, 1536, 896, step=16, label="Width")
@@ -214,9 +239,14 @@ def create_demo(args, model_name: str, device: str = "cuda" if torch.cuda.is_ava
                 generate_btn = gr.Button("Generate")
 
             with gr.Column():
-                output_image = gr.Image(label="Generated Image")
+                output_image = gr.Image(label="Generated Image", type="pil", format="png")
                 seed_output = gr.Textbox(label="Used Seed")
-                intermediate_output = gr.Gallery(label="Output", elem_id="gallery", visible=args.dev)
+                intermediate_output = gr.Gallery(
+                    label="Output",
+                    elem_id="gallery",
+                    visible=args.dev,
+                    allow_preview=True,
+                )
 
         with gr.Row(), gr.Column():
             gr.Markdown("## Examples")
@@ -296,4 +326,6 @@ if __name__ == "__main__":
         huggingface_hub.login(hf_token)
 
     demo = create_demo(args, args.name, args.device, args.offload)
-    demo.launch()
+    # SSR 비활성화: 업로드/탭 이동 중 ClientDisconnect 노이즈 감소 및 렌더 안정화
+    demo.launch(ssr_mode=False)
+