diff --git a/README.md b/README.md
index 8bd105e6d..703789d97 100644
--- a/README.md
+++ b/README.md
@@ -1,6 +1,8 @@
 <div align="center"><img src="assets/logo.png" width="350"></div>
 <img src="assets/demo.png" >
 
+<a href="https://replicate.com/arielreplicate/yolox"><img src="https://replicate.com/arielreplicate/yolox/badge"></a>
+
 ## Introduction
 YOLOX is an anchor-free version of YOLO, with a simpler design but better performance! It aims to bridge the gap between research and industrial communities.
 For more details, please refer to our [report on Arxiv](https://arxiv.org/abs/2107.08430).
diff --git a/cog.yaml b/cog.yaml
new file mode 100644
index 000000000..15669a5ee
--- /dev/null
+++ b/cog.yaml
@@ -0,0 +1,19 @@
+build:
+  gpu: true
+  cuda: "11.2"
+  python_version: 3.8
+  system_packages:
+    - libgl1-mesa-glx
+    - libglib2.0-0
+  python_packages:
+    - numpy==1.21.6
+    - torch==1.13
+    - torchvision==0.14
+    - opencv-python==4.6.0.66
+    - loguru==0.6.0
+    - psutil==5.9.4
+    - tqdm==4.64.1
+    - pycocotools==2.0.6
+    - tabulate==0.9.0
+    - thop==0.1.1.post2209072238
+predict: "predict.py:Predictor"
diff --git a/predict.py b/predict.py
new file mode 100644
index 000000000..cef0797fc
--- /dev/null
+++ b/predict.py
@@ -0,0 +1,102 @@
+import json
+
+import cv2
+import torch
+from typing import Optional
+
+from tools.demo import get_exp
+from tools.demo import Predictor as YoloPredictor
+from yolox.data import COCO_CLASSES
+
+from cog import BasePredictor, Path, Input, BaseModel
+
+YOLO_VERSIONS = ["yolox-s", "yolox-m", "yolox-l", "yolox-x"]
+
+
+class Output(BaseModel):
+    img: Optional[Path]
+    json_str: Optional[str]
+
+
+class Predictor(BasePredictor):
+    def setup(self):
+        pass
+
+    def predict(
+        self,
+        input_image: Path = Input(description="Path to an image"),
+        model_name: str = Input(
+            description="Model name", default="yolox-s", choices=YOLO_VERSIONS
+        ),
+        conf: float = Input(description="Confidence threshold: Only detections with confidence higher are kept",
+                            default=0.3, ge=0., le=1.),
+        nms: float = Input(description="NMS threshold: NMS removes redundant detections. Detections with overlap "
+                                       "percentage (IOU) above this threshold are consider redundant to each other and "
+                                       "only one of them will be kept", default=0.3, ge=0., le=1.),
+        tsize: int = Input(description="Resize image to this size", default=640),
+
+        return_json: bool = Input(
+        description="Return results in json format", default=False
+        ),
+    ) -> Output:
+        input_image = str(input_image)
+        model_name = str(model_name)
+        tsize = int(tsize)
+
+        # Load model
+        exp = get_exp(None, model_name)
+        exp.test_conf = float(conf)
+        exp.nmsthre = float(nms)
+        exp.test_size = (tsize, tsize)
+        model = exp.get_model().cuda()
+        model.eval()
+
+        ckpt = f'models/{model_name.replace("-", "_")}.pth'
+        ckpt = torch.load(ckpt, map_location="cpu")
+        model.load_state_dict(ckpt["model"])
+
+        predictor = YoloPredictor(
+            model, exp, COCO_CLASSES, None, None, "gpu", None, False
+        )
+
+        # inference
+        outputs, img_info = predictor.inference(input_image)
+
+        if bool(return_json):
+            return Output(
+                json_str=get_output_str(
+                    outputs[0], img_info["ratio"], predictor.cls_names
+                )
+            )
+        else:
+            result_image = predictor.visual(outputs[0], img_info, predictor.confthre)
+            output_path = f"outputs.png"
+            cv2.imwrite(output_path, result_image)
+            return Output(img=Path(output_path))
+
+
+def get_output_str(outputs, ratio, class_names, cls_conf=0.35):
+    if outputs is None:
+        return json.dumps("")
+
+    output = outputs.cpu().numpy()
+
+    bboxes = output[:, 0:4]
+    bboxes /= ratio
+
+    cls = output[:, 6]
+    scores = output[:, 4] * output[:, 5]
+    output_dict = dict()
+    for i in range(len(bboxes)):
+        if scores[i] > cls_conf:
+            box = bboxes[i]
+            output_dict[f"Det-{i}"] = {
+                "x0": box[0],
+                "y0": box[1],
+                "x1": box[2],
+                "y1": box[3],
+                "score": scores[i],
+                "cls": class_names[int(cls[i])],
+            }
+
+    return json.dumps(str(output_dict))