ernie-image

name	ernie-image
description	ERNIE-Image parameter planning and prompt design for Baidu AI Studio image generation. Use when the user wants to create images with ERNIE-Image, improve ERNIE prompts, or choose ERNIE-specific arguments such as size, seed, use_pe, num_inference_steps, guidance_scale, n, and response_format.

ERNIE-Image

Use this skill when the user is generating images with Baidu AI Studio ERNIE-Image and needs concrete help choosing arguments, not just prompt polishing.

What this skill should do

• Translate a vague image request into a production-ready prompt.
• Recommend ERNIE-Image arguments that match the goal.
• Explain the tradeoff for each non-default argument in plain language.
• Output a final parameter block or runnable client.images.generate(...) call when useful.

Parameter defaults to reason from

client.images.generate(
    model="ernie-image-turbo",
    prompt="...",
    n=1,
    response_format="b64_json",
    size="1024x1024",
    extra_body={
        "seed": 42,
        "use_pe": True,
        "num_inference_steps": 8,
        "guidance_scale": 1.0,
    },
)

Decision rules

size

• Use 1024x1024 for square covers, posters, avatars, icons, and most general requests.
• Use 1376x768 or 768x1376 for dramatic landscape or portrait compositions.
• Use 1264x848 / 848x1264 when the user wants a more photographic frame with less panoramic distortion.
• Use 1200x896 / 896x1200 for editorial and product marketing layouts with moderate aspect ratio changes.

seed

• Keep or set a fixed seed when the user wants repeatability, iterative refinement, or later prompt comparisons.
• Omit or change the seed when the user wants broader exploration.
• If the user asks for multiple variants, recommend keeping the prompt stable and changing only the seed.

use_pe

• Recommend True when prompt fidelity matters and the user has a clear composition, subject, text placement, or style target.
• Consider False only when the user wants looser interpretation or broader creative drift.

num_inference_steps

• Start around 8 for fast drafts.
• Raise modestly when the user wants cleaner details and can accept slower generation.
• Keep it low for idea exploration or multi-variant runs.

guidance_scale

• Start around 1.0.
• Raise it when the user wants stricter adherence to prompt details.
• Lower it when the output feels too rigid or overconstrained.
• Change it in small increments and explain why.

n

• Use 1 for deliberate single-shot work or when cost and stability matter.
• Use 2-4 for exploration, but keep the rest of the parameters stable so the user can compare outputs.

response_format

• Prefer b64_json for scripts, automation, downloads, and local post-processing.
• Prefer url only when the user explicitly wants a hosted URL or a quick manual preview flow.

Prompt construction

• Write the prompt in the user's language unless another language is clearly better for the subject matter.
• Put the main subject, setting, composition, style, lighting, and quality constraints in that order.
• Include camera or layout language only if it materially helps the requested image.
• Avoid stuffing every style term available. Prefer one coherent visual direction.

Response format

When the user asks for help, usually return:

1. A refined prompt.
2. Recommended arguments with one-line justification each.
3. A final JSON object or Python snippet.

If the user already has a prompt, focus on parameter selection and only rewrite the prompt when it is clearly weak.