ERNIE-Image Generation#

Generate images with Baidu AI Studio's ERNIE-Image models through the OpenAI-compatible API. Best when the user specifically wants Baidu/ERNIE generation, Chinese-language image prompts, or local PNG outputs.

Safety Boundary#

• Use this skill only for explicit image-generation requests.
• Prompts and parameters are sent to Baidu AI Studio. Do not include secrets, credentials, private personal data, or confidential business data in prompts.
• Follow Baidu AI Studio terms and applicable platform rules. Decline unsafe, deceptive, exploitative, or rights-violating image requests.
• Do not print, store, or ask the user to paste AI_STUDIO_API_KEY into chat. The script reads it from the environment only.
• Generated files are written locally. Confirm the output directory when the user has not specified one.

Prerequisites#

• Python 3.11+ with uv installed
• AI_STUDIO_API_KEY environment variable set to your Baidu AI Studio access token
• Get a token at: https://aistudio.baidu.com/account/accessToken

Prefer ERNIE-Image-Turbo, 1024x1024, n=1, and b64_json unless the user asks for different parameters.

Models#

ModelBest forERNIE-Image-TurboFast drafts, iteration, batch previewsERNIE-ImageSlower, higher-quality final outputs

Chinese prompts usually work especially well with these models.

Generation Workflow#

Step 1 -- Compose the prompt#

Write a descriptive prompt (max 1024 characters, ~150 words). Chinese and English both work well. Be specific about subject, style, composition, and mood.

Good: "A golden retriever puppy sitting in a sunflower field at sunset, warm golden light, shallow depth of field, professional photography" Bad: "dog"

Step 2 -- Choose parameters#

ParameterValuesDefaultmodelERNIE-Image-Turbo, ERNIE-Image``ERNIE-Image-Turbosize1024x1024, 768x1376, 1376x768, 896x1200, 1200x896, 848x1264, 1264x848``1024x1024n1-41seedany integerrandomsteps4-20provider defaultguidance1.0-7.5provider defaultuse-peflagoff

Select size based on content: portraits and posters use vertical ( 768x1376, 848x1264, 896x1200), landscapes and covers use horizontal ( 1376x768, 1264x848, 1200x896), general use 1024x1024.

Step 3 -- Run the generation script#

Execute the bundled script with uv run:

bash

uv run {baseDir}/scripts/generate.py "<PROMPT>" --model ERNIE-Image-Turbo --size 1024x1024

For batch generation:

bash

uv run {baseDir}/scripts/generate.py "<PROMPT>" --n 4 --output ./output_dir

For higher quality with more inference steps and stronger guidance:

bash

uv run {baseDir}/scripts/generate.py "<PROMPT>" --model ERNIE-Image --steps 16 --guidance 3.5

For reproducible results:

bash

uv run {baseDir}/scripts/generate.py "<PROMPT>" --seed 42

Step 4 -- Output#

The script saves images as PNG files to the output directory and prints:

text

Saved: ernie_20260430_110100.png (1.7 MB)
MEDIA:/absolute/path/to/ernie_20260430_110100.png

The MEDIA: line enables automatic image attachment in compatible environments.

The script avoids overwriting existing files by adding a numeric suffix when needed.

For JSON output, add --json to get structured results:

json

{
  "success": true,
  "model": "ERNIE-Image-Turbo",
  "files": [{"path": "/abs/path/ernie_20260430_110100.png", "size_bytes": 1715660}],
  "prompt": "...",
  "parameters": {"size": "1024x1024", "seed": 42}
}

Quick Triggers#

When the user says any of these, treat the text after the trigger as the prompt and generate immediately with defaults:

• Chinese: "生成图片：xxx" / "文生图：xxx" / "百度生图：xxx" / "ERNIE生图：xxx" / "AI画图：xxx" / "AI绘图：xxx" / "出图：xxx" / "画画：xxx"
• English: "generate image: xxx" / "ernie image: xxx" / "baidu image: xxx" / "create image: xxx" / "text-to-image: xxx" / "AI art: xxx" / "make a picture: xxx" / "draw: xxx"

Defaults: model= ERNIE-Image-Turbo, size= 1024x1024, n=1, b64_json format.

Notes#

• Images are saved locally as PNG files with MEDIA:/absolute/path/to/file.png for auto-attach.
• Chinese prompts work particularly well with ERNIE models.
• Prompt enhancement ( --use-pe) lets the model expand simple prompts into richer descriptions before generation. Enable for short prompts, disable for precise control.
• Use --quiet to suppress progress output; only MEDIA: lines (or JSON with --json) are printed. Recommended when calling from agent workflows to reduce context noise.
• For full API reference, prompt guidance, and troubleshooting, read {baseDir}/references/api-guide.md.