Create app.py

app.py

@@ -0,0 +1,93 @@
+import os
+import time
+import gradio as gr
+import torch
+from PIL import Image
+
+from diffusers import StableVideoDiffusionPipeline
+import imageio.v2 as imageio
+
+
+# -------- paths --------
+ROOT = "/data" if os.path.isdir("/data") else "/home/user"
+MODEL_DIR = os.path.join(ROOT, "models", "svd-xt")
+OUT_DIR = os.path.join(ROOT, "outputs")
+os.makedirs(OUT_DIR, exist_ok=True)
+
+pipe = None
+
+def get_pipe():
+    global pipe
+    if pipe is not None:
+        return pipe
+
+    if not os.path.isdir(MODEL_DIR):
+        raise gr.Error(f"Model not found at {MODEL_DIR}. postBuild didn't download it.")
+
+    dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+    pipe = StableVideoDiffusionPipeline.from_pretrained(
+        MODEL_DIR,
+        torch_dtype=dtype,
+        local_files_only=True,   # <-- запрет докачки в рантайме
+    )
+
+    if torch.cuda.is_available():
+        pipe = pipe.to("cuda")
+        try:
+            pipe.enable_xformers_memory_efficient_attention()
+        except Exception:
+            pass
+
+    return pipe
+
+
+def run(image: Image.Image, motion: int, fps: int, frames: int, steps: int, seed: int):
+    if image is None:
+        raise gr.Error("Upload an image first.")
+
+    if seed < 0:
+        seed = int(time.time()) % 10_000_000
+
+    generator = torch.Generator(device="cuda" if torch.cuda.is_available() else "cpu").manual_seed(int(seed))
+
+    pipe = get_pipe()
+
+    # SVD любит 1024 ширину по умолчанию, но лучше держать умеренно для VRAM
+    # Можно подстроить под фото, но начнём с безопасного.
+    img = image.convert("RGB")
+
+    # В diffusers для SVD параметры могут называться немного по-разному между версиями,
+    # но обычно работают: num_frames, fps, motion_bucket_id, num_inference_steps
+    out = pipe(
+        image=img,
+        num_frames=int(frames),
+        fps=int(fps),
+        motion_bucket_id=int(motion),
+        num_inference_steps=int(steps),
+        generator=generator,
+    )
+
+    video_frames = out.frames[0]  # list[PIL]
+    out_path = os.path.join(OUT_DIR, f"svd_{int(time.time())}.mp4")
+    imageio.mimsave(out_path, video_frames, fps=int(fps))
+    return out_path
+
+
+with gr.Blocks(title="SVD img2vid XT (local)") as demo:
+    gr.Markdown("## Stable Video Diffusion (img2vid-xt) — local in Space")
+
+    with gr.Row():
+        inp = gr.Image(type="pil", label="Input image")
+        out = gr.Video(label="Output video")
+
+    with gr.Accordion("Settings", open=False):
+        motion = gr.Slider(1, 255, value=127, step=1, label="motion_bucket_id (higher = more motion)")
+        fps = gr.Slider(6, 30, value=12, step=1, label="fps")
+        frames = gr.Slider(8, 30, value=14, step=1, label="num_frames")
+        steps = gr.Slider(10, 50, value=25, step=1, label="steps")
+        seed = gr.Number(value=-1, precision=0, label="seed (-1 random)")
+
+    btn = gr.Button("Generate", variant="primary")
+    btn.click(run, [inp, motion, fps, frames, steps, seed], out)
+
+demo.queue().launch()