Upload 15 files

inference_drag.py

@@ -0,0 +1,111 @@
+import torch
+from diffusers import MotionAdapter, AnimateDiffPipeline, DDIMScheduler
+from diffusers.utils import export_to_gif
+from PIL import Image
+import cv2
+import numpy as np
+import os
+
+# ==========================================
+# 1. 설정 (학습 때와 맞춰야 잘 나옵니다)
+# ==========================================
+BASE_MODEL = "runwayml/stable-diffusion-v1-5"
+MOTION_ADAPTER = "guoyww/animatediff-motion-adapter-v1-5-2"
+LORA_PATH = "output_drag_lora"  # 방금 학습 끝난 폴더
+TEST_IMAGE_PATH = "test_input.png"  # 준비한 이미지
+
+# 학습 때 썼던 프롬프트 그대로 사용 (중요!)
+PROMPT = "timelapse clouds moving in the sky, cinematic, high quality, 4k"
+NEGATIVE_PROMPT = "bad quality, worst quality, blurry, low resolution, distortion, watermark"
+
+
+# ==========================================
+# 2. 색상 보정 저장 함수 (파란색 방지)
+# ==========================================
+def save_video_fixed(frames, path, fps=8):
+    height, width, _ = np.array(frames[0]).shape
+    # OpenCV 비디오 작성기
+    out = cv2.VideoWriter(path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (width, height))
+
+    for frame in frames:
+        # PIL -> Numpy 변환
+        img_np = np.array(frame)
+        # RGB -> BGR 변환 (이게 핵심!)
+        img_bgr = cv2.cvtColor(img_np, cv2.COLOR_RGB2BGR)
+        out.write(img_bgr)
+
+    out.release()
+    print(f"✨ 영상 저장 완료: {path}")
+
+
+# ==========================================
+# 3. 메인 실행 함수
+# ==========================================
+def main():
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"Using device: {device}")
+
+    # 1. 모델 로드
+    print("Loading Base Model...")
+    adapter = MotionAdapter.from_pretrained(MOTION_ADAPTER)
+    pipe = AnimateDiffPipeline.from_pretrained(
+        BASE_MODEL,
+        motion_adapter=adapter,
+        torch_dtype=torch.float16
+    ).to(device)
+
+    # 스케줄러 설정
+    pipe.scheduler = DDIMScheduler.from_pretrained(
+        BASE_MODEL,
+        subfolder="scheduler",
+        clip_sample=False,
+        timestep_spacing="linspace",
+        beta_schedule="linear",
+        steps_offset=1
+    )
+
+    # 2. 학습한 LoRA 불러오기 (성적표 확인)
+    print(f"Loading LoRA from {LORA_PATH}...")
+    try:
+        pipe.unet.load_attn_procs(LORA_PATH)
+        print("✅ LoRA Load Success!")
+    except Exception as e:
+        print(f"❌ LoRA Load Failed: {e}")
+        return
+
+    # 3. 이미지 준비
+    if not os.path.exists(TEST_IMAGE_PATH):
+        print(f"⚠️ {TEST_IMAGE_PATH}가 없습니다! 검은 화면으로 테스트합니다.")
+        input_image = Image.new('RGB', (256, 256), color='black')
+    else:
+        input_image = Image.open(TEST_IMAGE_PATH).convert("RGB")
+        input_image = input_image.resize((256, 256))  # 학습 해상도 맞춤
+
+    # 4. 영상 생성 (Inference)
+    print("Generating Video... (약 1분 소요)")
+
+    # 시드 고정 (매번 똑같이 잘 나오게 하기 위해)
+    generator = torch.Generator(device=device).manual_seed(42)
+
+    output = pipe(
+        prompt=PROMPT,
+        negative_prompt=NEGATIVE_PROMPT,
+        num_frames=16,  # 2초 영상
+        guidance_scale=7.5,
+        num_inference_steps=25,  # 25번만 그려도 충분
+        generator=generator,
+        width=256,
+        height=256
+    )
+
+    frames = output.frames[0]
+
+    # 5. 저장 (GIF + MP4)
+    export_to_gif(frames, "final_result.gif")
+    save_video_fixed(frames, "final_result.mp4", fps=8)
+
+    print("🎉 모든 작업 완료! 'final_result.mp4'를 확인하세요.")
+
+
+if __name__ == "__main__":
+    main()

output_drag_lora/checkpoint-200/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dc451b5b80d8f93d87dc2a4804e24ed2ca8d950dcae1bd8973fdf7520fab237
+size 5269637960

output_drag_lora/checkpoint-200/optimizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd45978dd2af53b0e7fa1c34949f65ec801a87fb3d75b19420692cc61ef3b8ef
+size 32586114

output_drag_lora/checkpoint-200/random_states_0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9fe4cbbb6dd28903cee54423e7adeb58305fed600351ad550648a40b914db57
+size 14344

output_drag_lora/checkpoint-200/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c5e66cd0701a17acd6a6484ee4019ae1d960fac369fffe82987ea5dd7915828
+size 988

output_drag_lora/checkpoint-400/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55badd86b56c1084d6ff85506e77c95303c35a50ea4c27a476ee9ad1faafc05d
+size 5269637960

output_drag_lora/checkpoint-400/optimizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0a7048d044782e7341cbfb89b12b6d552d4d9914bfc2d2750c0f7ae9a73748c
+size 32586114

output_drag_lora/checkpoint-400/random_states_0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c24300fc53e165ac7325140a2cfe55dcd85aa244509ddf785381b83bc638c02b
+size 14344

output_drag_lora/checkpoint-400/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ed35716f7a9275082254cfbe153b60686b76c261de984a9e4d37dfb27c8ef24
+size 988

output_drag_lora/checkpoint-600/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52a22baa43787af67caba82823f0553d510813387a7e5d801e872fff37a2b781
+size 5269637960

output_drag_lora/checkpoint-600/optimizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92514ffaa22d74392ed8688a8a0535d92c1be145410e80ef12f44314a5465f24
+size 32586114

output_drag_lora/checkpoint-600/random_states_0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b68c4c937eb2a1de34c6c4c7a4892625c0541036db3d34daa8391c3bcc16dc9f
+size 14344

output_drag_lora/checkpoint-600/scaler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30cc72c17790183d209119d8c6a06b1b84e2e99ea44df74f41e2a84800dba3b3
+size 988

output_drag_lora/pytorch_lora_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e6d416718ea78e54b1921d3ca3df89c559743b712e8f1ac61253321d84229f5
+size 16128712