Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

MiVisionForgeSVOR

v1.0.5

Video Object Remove with SVOR(stable video object removal).

2· 132·0 current·0 all-time
byFei Wang@wangfei1204
Security Scan
Capability signals
Requires sensitive credentials
These labels describe what authority the skill may exercise. They are separate from suspicious or malicious moderation verdicts.
VirusTotalVirusTotal
Suspicious
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
Name/description (video object removal) match the actual behavior: the script compresses and uploads videos to a remote SVOR endpoint and processes returned masks. Required binaries (python3, ffmpeg) and a single API key are appropriate for the described task.
Instruction Scope
SKILL.md and the script explicitly instruct uploading the full video to https://mipixgen-pre.ai.mioffice.cn and downloading processed results; this is necessary for a cloud-based removal service but is privacy-sensitive (the README also warns about not uploading sensitive content). No instructions attempt to read unrelated local secrets or system config.
Install Mechanism
No install spec (instruction-only plus an included script). Nothing is downloaded from arbitrary URLs or installed automatically; risk from installation mechanism is low.
Credentials
Only SVOR_API_KEY is required and is used as a Bearer token to authenticate to the documented service endpoint. No unrelated credentials or extra environment variables are requested.
Persistence & Privilege
The skill does not request persistent or system-wide presence (always is false). It writes temporary files in a temp directory under the video's folder, which is expected for media processing.
Assessment
This skill appears coherent but it uploads full videos to a remote service. Before installing, ensure you: (1) trust the endpoint (https://mipixgen-pre.ai.mioffice.cn) and the operator (declared as Xiaomi team), (2) do not upload videos containing sensitive or personally identifiable information, (3) keep your SVOR_API_KEY secret and avoid setting it in shared environments, (4) have ffmpeg and Python 3 installed and test the tool first with a small non-sensitive video, and (5) review any service terms or data-retention policy from the operator if you need stronger privacy guarantees.

Like a lobster shell, security has layers — review code before you run it.

Runtime requirements

Binspython3, ffmpeg
EnvSVOR_API_KEY
Primary envSVOR_API_KEY
latestvk976zhdngmvt4p49zfa261rsjd84yn4m
132downloads
2stars
6versions
Updated 3d ago
v1.0.5
MIT-0

视频目标分割与消除 (CLI 版)

基于 SVOR(stable video object removal) 实现视频目标分割和消除,通过命令行脚本 video_remove.py 执行。 支持最长3s视频的指定目标消除

基本信息

服务运营商小米大模型应用团队
服务域名https://mipixgen-pre.ai.mioffice.cn
开源仓库https://github.com/xiaomi-research/svor
API Key环境变量 SVOR_API_KEY(必需,可在开源仓库页面获取)
脚本路径scripts/video_remove.py

前置依赖

python --version      # Python 3.x
pip show requests     # requests 模块
ffmpeg -version       # ffmpeg 用于视频处理

工作流程

推荐两步法(复杂场景)

Step 1: 分割 → 获取 VLM 推荐 → 人工分析目标 ID Step 2: 使用目标 ID → mask 转换 → 合并 → 消除

一步法(简单场景,文本提示有效时)

直接指定 classes + targets → 分割 → mask 转换 → 消除

命令行用法

1. 仅分割(获取 VLM 推荐)

python scripts/video_remove.py --video "path/to/video.mp4" --segment --classes "label:text:remove_prompt"

输出 VLM 推荐的目标 ID 和边界框,据此决定哪些 ID 需要消除。

2. 跑消除(基于 VLM 推荐)

python scripts/video_remove.py --video "path/to/video.mp4" --skip-segment --targets "label:1,2,3"

3. 一步完成

python scripts/video_remove.py --video "path/to/video.mp4" --classes "sign:sign:remove the sign" --targets "sign:2,3"

4. 多类目标消除

# 第一步:分割两类目标
python scripts/video_remove.py --video "path/to/video.mp4" --segment \
  --classes "person:person:remove pedestrians" "car:car:remove cars"

# 第二步:根据 VLM 推荐填写 targets
python scripts/video_remove.py --video "path/to/video.mp4" --skip-segment \
  --targets "person:1,3,4" "car:2"

5. 指定标注帧

python scripts/video_remove.py --video "path/to/video.mp4" --frame 16 --segment \
  --classes "girl:girl in red:remove the girl in red"

6. 框标注(文本提示无效时)

python scripts/video_remove.py --video "path/to/video.mp4" --segment \
  --classes "obj:cup:remove the cup" --boxes "obj:0.15,0.7,0.3,0.25"

7. 点标注

python scripts/video_remove.py --video "path/to/video.mp4" --segment \
  --classes "obj:cup:remove the cup" --points "obj:0.25,0.75"

参数说明

参数说明示例
--video输入视频路径(必填)"D:\video.mp4"
--classes分割类别 label:text:prompt"person:person:remove people"
--targets目标 ID label:1,2,3"person:1,3,4"
--boxes框标注 label:x,y,w,h(归一化 0~1)"obj:0.15,0.7,0.3,0.25"
--points点标注 label:x,y(归一化 0~1)"obj:0.25,0.75"
--frame标注帧序号(0-based),默认 0--frame 16
--segment仅分割,打印 VLM 推荐后退出
--skip-segment跳过分割,使用缓存 mask
--tmp-dir临时目录(默认 ./temp/"C:\temp"

VLM 推荐分析规则

SAM3 返回的 suggest_obj 包含 VLM 对目标的分析。不要用正则表达式解析,应人工分析:

  • VLM 会说明哪些 obj_id 符合消除指令、哪些不符合
  • 关注"需要保留"和"需要消除"的区分
  • 结合 object_boxes 的坐标辅助判断位置

示例 VLM 输出分析:

VLM: 符合用户指令的有 obj_id 1、3、4(背景行人),需消除;
     obj_id 2 是前景人物,不需要消除。
→ --targets "person:1,3,4"

标注方式选择

场景推荐方式原因
常见物体(人、车、动物)文本提示简单直接
文本提示无法识别框标注比点标注更稳定
需要精确位置点标注定位精确
指定帧的目标--frame + 文本/框定位到特定帧

多类目标消除流程

SAM3 接口一次仅能分割一类目标,多类消除需要:

  1. 分别对每类目标调用 SAM3 分割
  2. 分析 VLM 推荐,确定每类的 obj_id
  3. 分别转换每类 mask(目标 ID → 255)
  4. blend=all_mode=lighten 合并所有 mask
  5. 用合并后的 mask 调用 MiEraser 消除

注意事项

  1. 上传限制:网关限制约 1000KB,脚本自动压缩
  2. 坐标系统points/bounding_boxes 用归一化坐标(0~1),object_boxes 用绝对像素坐标
  3. mask 格式:消除区域像素值必须为 255,背景为 0
  4. frame_index:选择目标清晰可见、无遮挡的帧
  5. 结果链接:预签名 URL,7 天过期
  6. 中文路径:ffmpeg 不支持中文路径,脚本自动拷贝到临时目录处理

错误码

状态码含义处理
400请求参数错误检查参数格式
401未提供 API Key检查 Headers
403API Key 无效检查 Key
413请求体过大降低视频分辨率
429超出速率限制等待后重试
502上游服务连接失败重试
504上游服务超时重试或降低分辨率

⚠️ 隐私与安全声明

  • 服务运营商:本技能的云端处理服务由小米大模型应用团队运营,服务端点为 https://mipixgen-pre.ai.mioffice.cn,与开源仓库 xiaomi-research/svor 对应
  • 数据上传:本技能会将完整视频文件上传至上述远程服务进行处理。请勿上传包含敏感个人信息或可识别身份信息的视频内容
  • 数据处理:上传的视频仅用于目标分割和消除处理,处理完成后结果通过预签名 URL 返回(7 天过期)。具体数据保留和访问控制政策请参阅小米大模型应用团队的服务条款
  • API 密钥安全:所需 SVOR_API_KEY 仅用于向该服务认证。请勿在共享环境中设置该密钥,也请勿将其粘贴到公共日志或版本控制中
  • 使用建议:首次使用前,建议在隔离环境中使用小型、不涉及敏感内容的测试视频验证功能

Comments

Loading comments...