tl-image

TensorsLab Image Generation

Overview

This skill enables AI-powered image generation through TensorsLab's API, supporting both text-to-image and image-to-image workflows. The agent enhances user prompts with detailed visual descriptions before calling the API, ensuring high-quality outputs.

Script Path

The Python scripts for this skill are located in the scripts/ subdirectory relative to this SKILL.md file. Always use the absolute path when executing scripts. Determine the absolute path based on where this skill is installed.

For example, if this SKILL.md is at /path/to/skills/tl-image/SKILL.md, then:

• Auth script: python "/path/to/skills/tl-image/scripts/tensorslab_auth.py"
• Image script: python "/path/to/skills/tl-image/scripts/tensorslab_image.py"

When executing, construct the command using the resolved absolute path:

python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" <args>

Authorization

BEFORE any image generation, you must ensure you are authorized with TensorsLab.

1. Automatic Authorization

The authorization script will automatically check if an API key already exists in the TENSORSLAB_API_KEY environment variable or in ~/.tensorslab/.env before proceeding. (Note: When you need to verify the environment variable, ONLY check if it exists. NEVER display or print the actual API key value.)

Run:

python "<absolute_path_to_skill_dir>/scripts/tensorslab_auth.py"

On Windows/macOS: This will open a browser for authorization. Wait for "Authorization Successful!" before proceeding.

On Linux: Browser auth is automatically skipped. The script will print instructions to get your API key manually from TensorsLab Console. Set it via:

export TENSORSLAB_API_KEY=your_api_key_here

After authorization, the API key is stored in ~/.tensorslab/.env and you don't need to re-authorize unless the key expires.

2. Key Persistence

When the API key is provided via export TENSORSLAB_API_KEY=xxx or --api-key argument and a generation task succeeds, the key is automatically saved to ~/.tensorslab/.env. Future sessions will pick it up without needing to export again.

Models

Model	Description	Best For
seedreamv4	Standard model (Default)	Fast generation, good quality
seedreamv45	Enhanced model	Higher quality than v4, faster than v5
seedreamv5	Seedream V5 Lite model (Generates slowly)	Highest quality, but slowest and most credits
zimage	Alternative model	Specific artistic styles
quickedit	Image instruction editing	Fast color/style/object editing

提示：默认使用 seedreamv4 模型以保证生成速度。推荐使用 seedreamv45 获得更好的质量。注意 seedreamv5 实际为 Seedream V5 Lite 版本，质量最高但生成速度显著较慢且耗费更多积分。

Workflow

For additional scenarios beyond basic generation (avatar generation, watermark removal, object erasure, face replacement), see references/scenarios.md.

1. Text-to-Image Generation

User request: "画一个在月球上吃热狗的宇航员"

Constraints:

• Do NOT pass sourceImage or imageUrl for text-to-image generation.

Agent processing:

1. Extract the core subject and action
2. Enhance prompt with details (lighting, composition, style, atmosphere)
3. Call API with enriched prompt
4. Monitor progress with heartbeat updates
5. Download to ./tensorslab_output/

Example enhanced prompt:

An astronaut sitting on the lunar surface, eating a hot dog with mustard,
cinematic lighting, Earth visible in the background, highly detailed,
photorealistic, 8k quality, dramatic shadows from the low sun angle

2. Image-to-Image Generation

User request: "把 cat.png 的背景换成太空" or "参考 sketch.png 渲染成 3D 模型"

Agent processing:

1. Extract image file paths (absolute or relative to current directory)
2. Enhance prompt with transformation instructions
3. Upload source images with prompt
4. Monitor and download results

Parameters for image-to-image:

• sourceImage: Array of image files (for local upload)
• imageUrl: URL of source image (Must be a standard HTTP/HTTPS URL. Do NOT use local paths like /tmp/xxx.png here)
• prompt: Description of desired transformation

3. Image Editing (General Purpose)

General-purpose editing for any local image modifications.

User request examples:

• "把这张图的天空改成日落色"
• "给人物加上墨镜"
• "把头发颜色染成粉色"

Agent processing:

1. Extract image file path
2. Parse the specific editing instruction (what to change, where)
3. Build enhanced prompt with precise editing guidance
4. Call API with source image and editing prompt
5. Save result to ./tensorslab_output/

Example enhanced prompt:

Change the sky to sunset colors with warm orange and pink gradients,
matching the existing lighting conditions and atmospheric perspective,
seamless blend at the horizon line

For avatar generation, watermark removal, object erasure, and face replacement scenarios, see references/scenarios.md.

4. Resolution Options

Supported formats:

• Aspect ratios: 9:16, 16:9, 3:4, 4:3, 1:1, 2:3, 3:2
• Resolution levels: 2K, 4K
• Specific dimensions: WxH format (e.g., 2048x2048, 1920x1080)
  • Constraint: Total pixels must be between 3,686,400 and 16,777,216

Using the Script

依赖：脚本需要 requests 和 pyyaml 库，首次使用前执行：

pip install requests pyyaml

重要：提示词必须用引号包裹。 传给脚本的 prompt 参数必须用双引号 " 包裹，避免 shell 解析问题：

# 正确
python "scripts/tensorslab_image.py" "a cat on the moon"
python "scripts/tensorslab_image.py" "把主体改为蓝色" --source image.png --model quickedit

# 错误 - 提示词没有引号
python "scripts/tensorslab_image.py" a cat on the moon

Execute the Python script directly:

# Text-to-image
python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" "a cat on the moon"

# With specific resolution
python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" "sunset over mountains" --resolution 16:9

# Image-to-image with local file
python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" "watercolor style" --source cat.png

# Image-to-image with URL
python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" "watercolor style" --image-url https://example.com/cat.jpg

# Specify model
python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" "cyberpunk city" --model seedreamv5

# Custom output directory
python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" "a beautiful landscape" --output-dir ./my_images

# Quick editing (Fast instructions)
python "<absolute_path_to_skill_dir>/scripts/tensorslab_image.py" "把主体改为蓝色" --source image.png --model quickedit

Task Status Flow

Status	Code	Meaning
Queued	1	Task waiting in queue
Processing	2	Currently generating
Completed	3	Done, images ready
Failed	4	Error occurred

Error Handling

Translate API errors to user-friendly messages:

Error Code	Meaning	User Message
9000	Insufficient credits	"亲，积分用完啦，请前往 https://tensorai.tensorslab.com"/ 充值"
9999	General error	Show the specific error message

Output

All images are saved to output directory with naming pattern:

• Default: ./tensorslab_output/ (current working directory)
• Custom: Use --output-dir or -o to specify a different path
• Naming: {task_id}_{index}.{ext} - e.g., abcd_1234567890_0.png

URL mapping: The script also saves file-to-URL mappings in ./tensorslab_output/urls.yaml. This file tracks the original URLs for each downloaded file and accumulates entries across multiple runs. When you need the original URL of a generated image, read this file.

# Example urls.yaml content
abcd_1234567890_0.png: https://tensorai.tensorslab.com/images/abcd_1234567890_0.png
abcd_1234567890_1.png: https://tensorai.tensorslab.com/images/abcd_1234567890_1.png

After completion, the script outputs both the local file path and the remote URL. Inform user with both:

🎉 您的图片处理完毕！
   - File: ./tensorslab_output/{filename}
   - URL: {remote_url}

Resources

• scripts/tensorslab_image.py: Main API client with full CLI support
• references/api_reference.md: Detailed API documentation
• references/scenarios.md: Advanced usage scenarios (avatar generation, watermark removal, object erasure, face replacement)