Add image2video microservice (Stable Video Diffusion)

.github/workflows/docker/compose/image2video-compose-cd.yaml

@@ -0,0 +1,18 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+# this file should be run in the root of the repo
+# images used by GenAIExamples: image2video,svd,svd-gaudi
+services:
+  image2video:
+    build:
+      dockerfile: comps/image2video/Dockerfile
+    image: ${REGISTRY:-opea}/image2video:${TAG:-latest}
+  svd:
+    build:
+      dockerfile: comps/image2video/dependency/Dockerfile
+    image: ${REGISTRY:-opea}/svd:${TAG:-latest}
+  svd-gaudi:
+    build:
+      dockerfile: comps/image2video/dependency/Dockerfile.intel_hpu
+    image: ${REGISTRY:-opea}/svd-gaudi:${TAG:-latest}

comps/__init__.py

@@ -21,6 +21,9 @@
     GraphDoc,
     LVMDoc,
     LVMVideoDoc,
+    ImagePath,
+    ImagesPath,
+    VideoPath,
     ImageDoc,
     TextImageDoc,
     MultimodalDoc,

comps/cores/mega/constants.py

@@ -29,6 +29,8 @@ class ServiceType(Enum):
     LVM = 12
     KNOWLEDGE_GRAPH = 13
     WEB_RETRIEVER = 14
+    IMAGE2VIDEO = 15
+    TEXT2IMAGE = 16
 
 
 class MegaServiceEndpoint(Enum):

comps/cores/proto/docarray.py

@@ -249,3 +249,15 @@ class LVMVideoDoc(BaseDoc):
     chunk_duration: float
     prompt: str
     max_new_tokens: conint(ge=0, le=1024) = 512
+
+
+class ImagePath(BaseDoc):
+    image_path: str
+
+
+class ImagesPath(BaseDoc):
+    images_path: DocList[ImagePath]
+
+
+class VideoPath(BaseDoc):
+    video_path: str

comps/image2video/Dockerfile

@@ -0,0 +1,18 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+FROM python:3.11-slim
+
+# Set environment variables
+ENV LANG=en_US.UTF-8
+
+COPY comps /home/comps
+
+RUN pip install --no-cache-dir --upgrade pip && \
+    pip install --no-cache-dir -r /home/comps/image2video/requirements.txt
+
+ENV PYTHONPATH=$PYTHONPATH:/home
+
+WORKDIR /home/comps/image2video
+
+ENTRYPOINT ["python", "image2video.py"]

comps/image2video/README.md

@@ -0,0 +1,83 @@
+# Image-to-Video Microservice
+
+Image-to-Video is a task that generate video conditioning on the provided image(s). This microservice supports image-to-video task by using Stable Video Diffusion (SVD) model.
+
+# 🚀1. Start Microservice with Python (Option 1)
+
+## 1.1 Install Requirements
+
+```bash
+pip install -r requirements.txt
+pip install -r dependency/requirements.txt
+```
+
+## 1.2 Start SVD Service
+
+```bash
+# Start SVD service
+cd dependency/
+python svd_server.py
+```
+
+## 1.3 Start Image-to-Video Microservice
+
+```bash
+cd ..
+# Start the OPEA Microservice
+python image2video.py
+```
+
+# 🚀2. Start Microservice with Docker (Option 2)
+
+## 2.1 Build Images
+
+### 2.1.1 SVD Server Image
+
+Build SVD server image on Xeon with below command:
+
+```bash
+cd ../..
+docker build -t opea/svd:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/image2video/dependency/Dockerfile .
+```
+
+Build SVD server image on Gaudi with below command:
+
+```bash
+cd ../..
+docker build -t opea/svd-gaudi:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/image2video/dependency/Dockerfile.intel_hpu .
+```
+
+### 2.1.2 Image-to-Video Service Image
+
+```bash
+docker build -t opea/image2video:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/image2video/Dockerfile .
+```
+
+## 2.2 Start SVD and Image-to-Video Service
+
+### 2.2.1 Start SVD server
+
+Start SVD server on Xeon with below command:
+
+```bash
+docker run --ipc=host -p 9368:9368 -e http_proxy=$http_proxy -e https_proxy=$https_proxy opea/svd:latest
+```
+
+Start SVD server on Gaudi with below command:
+
+```bash
+docker run -p 9368:9368 --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --ipc=host -e http_proxy=$http_proxy -e https_proxy=$https_proxy opea/svd-gaudi:latest
+```
+
+### 2.2.2 Start Image-to-Video service
+
+```bash
+ip_address=$(hostname -I | awk '{print $1}')
+docker run -p 9369:9369 --ipc=host -e http_proxy=$http_proxy -e https_proxy=$https_proxy -e SVD_ENDPOINT=http://$ip_address:9368 opea/image2video:latest
+```
+
+### 2.2.3 Test
+
+```bash
+http_proxy="" curl http://localhost:9369/v1/image2video -XPOST -d '{"images_path":[{"image_path":"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"}]}' -H 'Content-Type: application/json'
+```

comps/image2video/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0

comps/image2video/dependency/Dockerfile

@@ -0,0 +1,22 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+FROM python:3.11-slim
+
+# Set environment variables
+ENV LANG=en_US.UTF-8
+
+ARG ARCH="cpu"
+
+COPY comps /home/comps
+
+RUN apt-get update && apt-get install python3-opencv -y && \
+    pip install --no-cache-dir --upgrade pip && \
+    if [ ${ARCH} = "cpu" ]; then pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu; fi && \
+    pip install --no-cache-dir -r /home/comps/image2video/dependency/requirements.txt
+
+ENV PYTHONPATH=$PYTHONPATH:/home
+
+WORKDIR /home/comps/image2video/dependency
+
+ENTRYPOINT ["python", "svd_server.py"]

comps/image2video/dependency/Dockerfile.intel_hpu

@@ -0,0 +1,30 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+# HABANA environment
+# FROM vault.habana.ai/gaudi-docker/1.16.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.2:latest as hpu
+FROM opea/habanalabs:1.16.1-pytorch-installer-2.2.2 as hpu
+RUN useradd -m -s /bin/bash user && \
+    mkdir -p /home/user && \
+    chown -R user /home/user/
+
+COPY comps /home/user/comps
+
+RUN chown -R user /home/user/comps/image2video
+
+RUN rm -rf /etc/ssh/ssh_host*
+USER user
+# Set environment variables
+ENV LANG=en_US.UTF-8
+ENV PYTHONPATH=/home/user:/usr/lib/habanalabs/:/optimum-habana
+
+# Install requirements and optimum habana
+RUN pip install --no-cache-dir --upgrade pip && \
+    pip install --no-cache-dir -r /home/user/comps/image2video/dependency/requirements.txt && \
+    pip install --no-cache-dir optimum[habana]
+
+ENV PYTHONPATH=$PYTHONPATH:/home/user
+
+WORKDIR /home/user/comps/image2video/dependency
+
+ENTRYPOINT ["python", "svd_server.py", "--device", "hpu"]

comps/image2video/dependency/requirements.txt

@@ -0,0 +1,7 @@
+accelerate
+diffusers
+fastapi
+opencv-python
+torch
+transformers
+uvicorn

comps/image2video/dependency/svd_server.py

@@ -0,0 +1,76 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+"""Stand-alone Stable Video Diffusion FastAPI Server."""
+
+import argparse
+import os
+import time
+
+import torch
+import uvicorn
+from diffusers import StableVideoDiffusionPipeline
+from diffusers.utils import export_to_video, load_image
+from fastapi import FastAPI, Request
+from fastapi.responses import JSONResponse, Response
+
+app = FastAPI()
+
+
+@app.post("/generate")
+async def generate(request: Request) -> Response:
+    print("SVD generation begin.")
+    request_dict = await request.json()
+    images_path = request_dict.pop("images_path")
+
+    start = time.time()
+    images = [load_image(img) for img in images_path]
+    images = [image.resize((1024, 576)) for image in images]
+
+    generator = torch.manual_seed(args.seed) if args.device == "cpu" else None
+    frames = pipe(images, decode_chunk_size=8, generator=generator).frames[0]
+    video_path = os.path.join(os.getcwd(), args.video_path)
+    export_to_video(frames, video_path, fps=7)
+    end = time.time()
+    print(f"SVD video output in {video_path}, time = {end-start}s")
+    return JSONResponse({"video_path": video_path})
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="0.0.0.0")
+    parser.add_argument("--port", type=int, default=9368)
+    parser.add_argument("--model_name_or_path", type=str, default="stabilityai/stable-video-diffusion-img2vid-xt")
+    parser.add_argument("--video_path", type=str, default="generated.mp4")
+    parser.add_argument("--use_hpu_graphs", default=False, action="store_true")
+    parser.add_argument("--device", type=str, default="cpu")
+    parser.add_argument("--seed", type=int, default=42)
+
+    args = parser.parse_args()
+    if args.device == "hpu":
+        from optimum.habana.diffusers import GaudiEulerDiscreteScheduler, GaudiStableVideoDiffusionPipeline
+        from optimum.habana.utils import set_seed
+
+        set_seed(args.seed)
+        scheduler = GaudiEulerDiscreteScheduler.from_pretrained(args.model_name_or_path, subfolder="scheduler")
+        kwargs = {
+            "scheduler": scheduler,
+            "use_habana": True,
+            "use_hpu_graphs": args.use_hpu_graphs,
+            "gaudi_config": "Habana/stable-diffusion",
+        }
+        pipe = GaudiStableVideoDiffusionPipeline.from_pretrained(
+            args.model_name_or_path,
+            **kwargs,
+        )
+    elif args.device == "cpu":
+        pipe = StableVideoDiffusionPipeline.from_pretrained(args.model_name_or_path)
+    else:
+        raise NotImplementedError(f"Only support cpu and hpu device now, device {args.device} not supported.")
+    print("Stable Video Diffusion model initialized.")
+
+    uvicorn.run(
+        app,
+        host=args.host,
+        port=args.port,
+        log_level="debug",
+    )

comps/image2video/image2video.py

@@ -0,0 +1,47 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+
+import json
+import os
+import time
+
+import requests
+
+from comps import (
+    ImagesPath,
+    ServiceType,
+    VideoPath,
+    opea_microservices,
+    register_microservice,
+    register_statistics,
+    statistics_dict,
+)
+
+
+@register_microservice(
+    name="opea_service@image2video",
+    service_type=ServiceType.IMAGE2VIDEO,
+    endpoint="/v1/image2video",
+    host="0.0.0.0",
+    port=9369,
+    input_datatype=ImagesPath,
+    output_datatype=VideoPath,
+)
+@register_statistics(names=["opea_service@image2video"])
+async def image2video(input: ImagesPath):
+    start = time.time()
+    images_path = [img.image_path for img in input.images_path]
+    inputs = {"images_path": images_path}
+    video_path = requests.post(url=f"{svd_endpoint}/generate", data=json.dumps(inputs), proxies={"http": None}).json()[
+        "video_path"
+    ]
+
+    statistics_dict["opea_service@image2video"].append_latency(time.time() - start, None)
+    return VideoPath(video_path=video_path)
+
+
+if __name__ == "__main__":
+    svd_endpoint = os.getenv("SVD_ENDPOINT", "http://localhost:9368")
+    print("Image2video server started.")
+    opea_microservices["opea_service@image2video"].start()

comps/image2video/requirements.txt

@@ -0,0 +1,11 @@
+datasets
+docarray[full]
+fastapi
+opentelemetry-api
+opentelemetry-exporter-otlp
+opentelemetry-sdk
+prometheus-fastapi-instrumentator
+pydantic==2.7.2
+pydub
+shortuuid
+uvicorn