diff --git a/doc/SAM31.md b/doc/SAM31.md
new file mode 100644
index 0000000..5545a98
--- /dev/null
+++ b/doc/SAM31.md
@@ -0,0 +1,13 @@
+# SAM 3.1 图片分割
+
+SAM 3.1 使用独立的 Python 环境运行，避免改变 SSUI 当前锁定的 PyTorch 2.4.1 和 CUDA 12.4 依赖。
+
+1. 按 [官方安装说明](https://github.com/facebookresearch/sam3#installation)创建 Python 3.12、PyTorch 2.7+、CUDA 12.6+ 的环境，并安装最新版 `facebookresearch/sam3`。
+2. 在 [facebook/sam3.1](https://huggingface.co/facebook/sam3.1)申请权重访问权限，并在独立环境中登录 Hugging Face。也可以下载 `sam3.1_multiplex.pt` 到本地，通过 `--checkpoint` 指定路径。
+3. 在仓库根目录运行：
+
+   ```bash
+   python scripts/sam31_image.py input.jpg "red car" output_masks
+   ```
+
+脚本写入 `results.json` 和每个实例的灰度 PNG 掩码。`--threshold` 可设置置信度阈值，默认 0.5。该入口只处理图片文本提示；视频追踪和 SSUI 工作流节点仍需独立接入。
diff --git a/scripts/sam31_image.py b/scripts/sam31_image.py
new file mode 100644
index 0000000..692b8ab
--- /dev/null
+++ b/scripts/sam31_image.py
@@ -0,0 +1,52 @@
+"""Run SAM 3.1 image segmentation in a separate, official SAM3 environment.
+
+Example: python scripts/sam31_image.py image.jpg "red car" output_dir
+"""
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from PIL import Image
+from sam3.model.sam3_image_processor import Sam3Processor
+from sam3.model_builder import build_sam3_image_model, download_ckpt_from_hf
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Segment an image using SAM 3.1")
+    parser.add_argument("image", type=Path)
+    parser.add_argument("prompt")
+    parser.add_argument("output_dir", type=Path)
+    parser.add_argument("--checkpoint", type=Path, help="Local SAM 3.1 checkpoint; downloads from Hugging Face if omitted")
+    parser.add_argument("--threshold", type=float, default=0.5)
+    args = parser.parse_args()
+
+    if not 0 <= args.threshold <= 1:
+        parser.error("--threshold must be between 0 and 1")
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    checkpoint = str(args.checkpoint) if args.checkpoint else download_ckpt_from_hf(version="sam3.1")
+    model = build_sam3_image_model(device=device, checkpoint_path=checkpoint, load_from_HF=False)
+    processor = Sam3Processor(model, device=device, confidence_threshold=args.threshold)
+
+    with Image.open(args.image) as source:
+        image = source.convert("RGB")
+    with torch.inference_mode():
+        result = processor.set_text_prompt(args.prompt, processor.set_image(image))
+
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+    instances = []
+    for index, (mask, box, score) in enumerate(zip(result["masks"], result["boxes"], result["scores"])):
+        mask_file = f"mask_{index:04d}.png"
+        mask_image = Image.fromarray((mask.squeeze().cpu().numpy() > 0.5).astype("uint8") * 255)
+        mask_image.save(args.output_dir / mask_file)
+        instances.append({"mask": mask_file, "box": box.cpu().tolist(), "score": float(score.cpu())})
+
+    (args.output_dir / "results.json").write_text(
+        json.dumps({"prompt": args.prompt, "image": str(args.image), "instances": instances}, indent=2),
+        encoding="utf-8",
+    )
+
+
+if __name__ == "__main__":
+    main()