Update app.py

app.py

@@ -103,11 +103,6 @@ os.environ['PYTORCH_NO_CUDA_MEMORY_CACHING'] = '1'
 # CUDA 초기화 방지
 torch.set_grad_enabled(False)
 
-
-
-# Gradio 인터페이스에 주기적 정리 추가
-demo.load(periodic_cleanup, every=5)  # 5초마다 정리
-
 # Hugging Face 토큰 설정
 HF_TOKEN = os.getenv("HF_TOKEN")
 if HF_TOKEN is None:
@@ -246,12 +241,12 @@ def image_to_3d(trial_id: str, seed: int, randomize_seed: bool, ss_guidance_stre
                     formats=["gaussian", "mesh"],
                     preprocess_image=False,
                     sparse_structure_sampler_params={
-                        "steps": min(ss_sampling_steps, 8),  # 스텝 수 제한
+                        "steps": min(ss_sampling_steps, 8),
                         "cfg_strength": ss_guidance_strength,
-                        "batch_size": 1  # 배치 크기 명시적 제한
+                        "batch_size": 1
                     },
                     slat_sampler_params={
-                        "steps": min(slat_sampling_steps, 8),  # 스텝 수 제한
+                        "steps": min(slat_sampling_steps, 8),
                         "cfg_strength": slat_guidance_strength,
                         "batch_size": 1
                     },
@@ -263,8 +258,8 @@ def image_to_3d(trial_id: str, seed: int, randomize_seed: bool, ss_guidance_stre
                 # 비디오 렌더링 최적화
                 video = render_utils.render_video(
                     outputs['gaussian'][0],
-                    num_frames=30,  # 프레임 수 감소
-                    resolution=384  # 해상도 제한
+                    num_frames=30,
+                    resolution=384
                 )['color']
                 
                 video_geo = render_utils.render_video(
@@ -278,7 +273,7 @@ def image_to_3d(trial_id: str, seed: int, randomize_seed: bool, ss_guidance_stre
                 video_geo = [v.cpu().numpy() for v in video_geo]
                 clear_gpu_memory()
                 
-                # 나머지 처리
+                # 비디오 생성 및 저장
                 video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
                 new_trial_id = str(uuid.uuid4())
                 video_path = f"{TMP_DIR}/{new_trial_id}.mp4"
@@ -294,20 +289,11 @@ def image_to_3d(trial_id: str, seed: int, randomize_seed: bool, ss_guidance_stre
                 
     except Exception as e:
         print(f"Error in image_to_3d: {str(e)}")
-        g.trellis_pipeline.to('cpu')
+        if hasattr(g.trellis_pipeline, 'to'):
+            g.trellis_pipeline.to('cpu')
         clear_gpu_memory()
         return None, None
 
-        return generated_image
-    except Exception as e:
-        print(f"Error in image generation: {str(e)}")
-        return None
-    finally:
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            torch.cuda.synchronize()
-        gc.collect()
-
 def clear_gpu_memory():
     """GPU 메모리를 더 철저하게 정리하는 함수"""
     try:
@@ -377,7 +363,6 @@ def text_to_image(prompt: str, height: int, width: int, steps: int, scales: floa
             return any(ord('가') <= ord(c) <= ord('힣') for c in text)
         
         if contains_korean(prompt):
-            # Helsinki-NLP/opus-mt-ko-en 모델을 사용하여 번역
             translated = g.translator(prompt)[0]['translation_text']
             prompt = translated
             print(f"Translated prompt: {prompt}")
@@ -418,16 +403,6 @@ def text_to_image(prompt: str, height: int, width: int, steps: int, scales: floa
             torch.cuda.empty_cache()
             torch.cuda.synchronize()
         gc.collect()
-
-        return generated_image
-    except Exception as e:
-        print(f"Error in image generation: {str(e)}")
-        return None
-    finally:
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            torch.cuda.synchronize()
-        gc.collect()
     
 css = """
 footer {