Jimeng AI Image Generator

即梦AI生图与视频生成工具。通过火山引擎API调用即梦AI，支持文生图、图片生成、视频生成。内置50+提示词模板覆盖电商、社交、商业设计、视频等场景。

Security Scan

The skill is mostly coherent with an image/video generation tool, but it reads sensitive credentials from your local secrets.json (jimeng AccessKeyId/SecretAccessKey) even though no credentials or config path are declared — that mismatch is a meaningful security concern.

Purpose & Capability

Name/description and the included Python scripts line up: they call Volcengine's visual.volcengineapi.com to submit image/video generation tasks. Requiring python is expected. However, the skill uses an on-disk secrets entry named 'jimeng' (AccessKeyId + SecretAccessKey) in ~/.openclaw/workspace/secrets.json; the registry metadata declares no required credentials or config paths. This is an incoherence: a networked cloud API key is required but not declared.

Instruction Scope

SKILL.md explicitly instructs the agent to use the token-manager 'jimeng' key and to run the included scripts. The scripts read ~/.openclaw/workspace/secrets.json, sign requests, call Volcengine, poll for results, and save image/video outputs to local files (temp_output.png, output.png). Those operations are within the stated purpose, but they access local secrets that were not declared and write files under the skill/agent workspace — the instructions therefore require access to sensitive local data without that being surfaced in the skill metadata.

Install Mechanism

No install spec — instruction- and script-based. That minimizes supply-chain install risk (no external downloads or package installs). The skill uses only Python stdlib urllib/hmac/hashlib; nothing fetched at install time.

Credentials

The code requires AccessKeyId and SecretAccessKey for Volcengine stored in a local secrets.json, but the skill metadata declares no required env vars or config paths. Requesting cloud API credentials for this purpose is reasonable, but failing to declare them is a transparency/permission problem. The credentials are sensitive (can be used to call Volcengine) and the skill will read them from disk automatically when invoked.

Persistence & Privilege

The skill is not always:true and does not attempt to modify other skills or global settings. It writes output files to the agent workspace and skill-relative paths (temp_output.png, output.png) which is expected behaviour for a generator; no persistent installation or background services are created.

What to consider before installing

This skill otherwise behaves like a normal Volcengine/Jimeng client, but it will automatically read credentials from ~/.openclaw/workspace/secrets.json under the key name 'jimeng' (AccessKeyId + SecretAccessKey) even though the registry doesn't declare any required credentials. Before installing or running it, verify: (1) the 'jimeng' entry in your token-manager/secrets.json is intended for this tool and has minimal permissions; (2) you trust the skill source (owner/homepage not provided); (3) inspect the secrets.json file and the scripts yourself if possible. If you don't want the skill to access your saved keys, do not install/run it or move/remove the 'jimeng' entry. Also note the skill will write image/video files to the agent workspace (temp_output.png, output.png).

即梦AI (Jimeng AI)

通过火山引擎 API 调用即梦AI，支持文生图、图生视频。

凭证

使用 token-manager 中保存的 jimeng 密钥：

• key: jimeng
• 内容: AccessKeyId + SecretAccessKey

快速使用

文生图（推荐）

# 生图并保存到文件
python skills/jimeng-ai/scripts/generate.py --prompt "描述" --output image.png

# 生图输出 MEDIA_URL（用于 OpenClaw 内联显示）
python skills/jimeng-ai/scripts/generate.py --prompt "描述"

查询任务

python skills/jimeng-ai/scripts/query.py <task_id> [req_key]

视频生成

python skills/jimeng-ai/scripts/video.py "描述" --duration 5

提示词模板库

内置50+即梦提示词模板，详见：

📋 references/prompts-templates.md

分类	模板数	场景
🛒 电商促销海报	15个	双11/618/春节/520/3C数码/美妆/运动等
📱 社交媒体图文	10个	小红书/朋友圈/知识科普/美食/健身/职场
🎨 商业设计	8个	Logo/IP形象/UI图标/VI系统
🎬 视频生成	10个	Vlog/品牌广告/国潮/创意短片

万能公式

主体 + 动作/状态 + 场景环境 + 风格美学 + 光影氛围 + 镜头视角 + 画质参数

示例：

• ❌ "一只猫在沙发上。"
• ✅ "一只橘色的胖猫慵懒地躺在棕色的真皮沙发上，午后阳光从落地窗斜射进来，空气中有细小的灰尘颗粒在舞动，电影质感，浅景深，8K超高清。"

参数说明

文生图 (req_key: jimeng_t2i_v40)

参数	说明
--prompt, -p	图像描述（必填）
--aspect	宽高比: 1:1(默认), 3:4, 4:3, 16:9, 9:16
--seed	随机种子，-1 表示随机
--output, -o	保存路径（不填则输出 MEDIA_URL）

文生视频 (req_key: jimeng_i2v_first_v30_1080)

参数	说明
--prompt	视频描述（必填）
--duration	时长(秒): 5-10（默认5）
--seed	随机种子

输出格式

Task submitted: <task_id>
IMAGE_SAVED: /path/to/image.png    # --output 模式
MEDIA_URL: https://...             # 无 --output 模式

进阶技巧

1. 角色定位 — 提示词开头加"你是一位资深电商平面设计师"
2. 锁定元素 — 引号标出必出内容，如"标题为'盛夏狂欢节'"
3. 风格先行 — 越早写"风格"效果越好
4. 分阶段生成 — 先简短提示词验证构图，满意后再加细节
5. 参考图 — 支持参考图+提示词双保险

注意事项

1. 生图任务通常 5-10 秒完成，视频任务 30-60 秒
2. 免费账户每日有积分限制（约66次生图）
3. 轮询间隔建议 3-5 秒，避免频繁请求
4. 图片通过 binary_data_base64 返回，无需处理过期URL