gpt-image

GPT-Image Generation Skill

Generate photorealistic and artistic images using Azure OpenAI GPT-image-2 for presentations.

When to Use This Skill

Use Case	Example	Why GPT-image-2
Cover / Title slides	"AI赋能数字化转型" concept art	No existing tool can create photorealistic concept illustrations
Industry scenario imagery	Retail store, factory floor, hospital, smart city	Scene-based visuals for solution demos
Abstract concept visuals	"Zero Trust Security" metaphor, "Cloud Native" illustration	Visual storytelling beyond boxes and arrows
Product UI mockups	Dashboard preview, mobile app concept	More realistic than excalidraw hand-drawn
Hero backgrounds	Gradient mesh with tech elements for slide backgrounds	CSS gradients are flat; AI creates depth
Custom icons / logos	Branded concept icons not in Azure icon library	Fills gaps in standard icon sets
Multilingual infographics	Images with embedded Chinese/English text	Native CJK text rendering support

When NOT to Use This Skill

Scenario	Use Instead
Cloud architecture diagrams (Azure/AWS/GCP)	azure-diagrams — has 700+ official icons
Hand-drawn / whiteboard style sketches	excalidraw-diagram — purpose-built for that aesthetic
Data charts (bar, pie, line)	pptx built-in PptxGenJS charts
Simple icons and gradient bars	SVG + Sharp rasterization

Execution Method

Always execute inline — do not create separate .py files:

python3 << 'EOF'
import os, base64, requests

endpoint = os.environ["AZURE_OPENAI_ENDPOINT"]
api_key = os.environ["AZURE_OPENAI_API_KEY"]
deployment = os.environ.get("GPT_IMAGE_DEPLOYMENT", "gpt-image-2")
api_version = "2025-04-01-preview"

url = f"{endpoint}openai/deployments/{deployment}/images/generations?api-version={api_version}"

response = requests.post(url, headers={
    "Api-Key": api_key,
    "Content-Type": "application/json"
}, json={
    "prompt": "YOUR PROMPT HERE",
    "n": 1,
    "size": "1536x1024",
    "quality": "high",
    "output_format": "png"
}).json()

if "error" in response:
    print(f"ERROR: {response['error']}")
else:
    img_data = base64.b64decode(response["data"][0]["b64_json"])
    output_path = "outputs/{project}/diagrams/image-name.png"
    os.makedirs(os.path.dirname(output_path), exist_ok=True)
    with open(output_path, "wb") as f:
        f.write(img_data)
    print(f"Image saved: {output_path} ({len(img_data)} bytes)")

EOF

Prerequisites

Configuration (Required)

推荐方式：编辑 config.json（在 skill 根目录 gpt-image/config.json）：

{
  "azure_openai_endpoint": "https://your-resource.openai.azure.com/",
  "azure_openai_api_key": "your-api-key",
  "deployment_name": "gpt-image-2",
  "api_version": "2025-04-01-preview"
}

配置优先级：CLI 参数 > config.json > 环境变量

备选方式：环境变量（如果 config.json 未填写，会自动回退到环境变量）：

export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
export AZURE_OPENAI_API_KEY="your-key"
export GPT_IMAGE_DEPLOYMENT="gpt-image-2"

Python Dependencies

pip install requests pillow

Verify Setup

# 查看当前配置来源
python3 scripts/generate_image.py config

# 完整验证（含 API 连通性测试）
python3 scripts/generate_image.py verify

API Parameters

Parameter	Values	Default	Notes
size	See size table below	1024x1024	Both edges must be multiples of 16
quality	low, medium, high	high	low = fast/cheap, high = best quality
n	1-10	1	Number of images per request
output_format	png, jpeg, webp	png	Use png for transparency support
background	auto, transparent	auto	Requires png format
output_compression	0-100	—	JPEG only

Recommended Sizes for PPT

Use Case	Size	Aspect Ratio	Notes
Full-width slide image	1536x1024	3:2	Fits 16:9 slide well
Right-panel diagram	1024x1024	1:1	For 左文右图 layout
Wide banner / hero	1920x1080	16:9	Full slide background
Tall sidebar	1024x1536	2:3	Vertical illustration
4K high-res	3840x2160	16:9	For print or zoom

GPT-image-2 supports arbitrary resolutions: both edges must be multiples of 16px, long edge up to 3840px, pixel count 655,360–8,294,400.

Prompt Engineering for PPT Images

Core Principles

1. Describe the visual scene, not the concept — "A modern glass-walled data center with blue LED server racks and holographic dashboards floating in the air" is better than "cloud computing"
2. Specify the style — "corporate illustration style", "isometric 3D", "flat design with subtle gradients", "photorealistic"
3. Include composition guidance — "wide angle", "bird's eye view", "centered subject with blurred background"
4. Match the PPT color palette — mention colors explicitly: "predominantly white and light gray background with orange (#E8913A) accent elements"
5. Request clean backgrounds — for slide embedding: "clean white background" or "transparent background"

Prompt Templates by Use Case

Cover / Title Slide Art

A [style] illustration of [concept], featuring [specific visual elements].
The scene uses a clean, modern aesthetic with a predominantly white/light
background. Key accent colors: warm orange and cool blue tones.
Corporate presentation quality, high detail, professional.

Industry Scenario

A photorealistic view of [industry setting], showing [specific activity].
Modern, well-lit environment. Shot from [angle]. Professional quality,
suitable for a business presentation. No text overlays.

Abstract Concept

An abstract [style] visualization representing [concept]. Use geometric
shapes, flowing lines, and a color palette of [colors]. Minimalist
composition with strong visual hierarchy. Clean background suitable
for slide overlay text.

Background / Hero Image

A subtle, abstract background pattern for a presentation slide.
Soft gradients blending [color1] to [color2] with [geometric/organic]
elements. Low contrast, suitable for overlaying dark text. Resolution:
1920x1080, corporate and modern feel.

Text in Images (Multilingual)

GPT-image-2 supports rendering text in Chinese, English, Japanese, Korean, Hindi, Bengali:

Create an infographic with the title "数字化转型路线图" in bold Chinese text
at the top. Below it, show 4 stages with labels in Chinese...

Tip: Keep embedded text short and large. Complex multi-paragraph text may render poorly.

Integration with CSA-PPT Workflow

File Output Convention

All generated images go to {workspace_path}/diagrams/ with descriptive names:

outputs/{project}/diagrams/
├── cover-concept.png          ← Title slide illustration
├── retail-scenario.png        ← Industry scenario image
├── zero-trust-visual.png      ← Abstract concept art
├── hero-background.png        ← Slide background
└── manifest.md                ← Updated with new entries

Manifest Entry Format

When adding to diagrams/manifest.md:

| cover-concept.png | Concept Art | gpt-image-2 | 1536x1024 | AI digital transformation concept for title slide |

Quality Verification

After generating each image:

1. Check file exists and is non-empty: ls -la {path}
2. Verify dimensions: python3 -c "from PIL import Image; img=Image.open('{path}'); print(img.size)"
3. Visual check: Use the Read tool to view the image (Claude Code supports image viewing)

Retry Strategy

If generation fails:

1. content_policy_violation: Rephrase the prompt to be less specific about people/faces
2. Rate limit (429): Wait 10 seconds, retry with quality="medium"
3. Timeout: Retry with smaller size or quality="low"
4. Persistent failure: Log to progress.md as [ERROR], suggest using SVG+Sharp as fallback for simple visuals

Image Editing (Advanced)

GPT-image-2 also supports editing existing images (inpainting):

python3 << 'EOF'
import os, base64, requests

endpoint = os.environ["AZURE_OPENAI_ENDPOINT"]
api_key = os.environ["AZURE_OPENAI_API_KEY"]
deployment = os.environ.get("GPT_IMAGE_DEPLOYMENT", "gpt-image-2")
api_version = "2025-04-01-preview"

url = f"{endpoint}openai/deployments/{deployment}/images/edits?api-version={api_version}"

with open("source_image.png", "rb") as img_file:
    response = requests.post(url,
        headers={"Api-Key": api_key},
        data={
            "prompt": "Replace the background with a modern office setting",
            "n": 1,
            "size": "1024x1024",
            "quality": "high"
        },
        files={
            "image": ("source.png", img_file, "image/png"),
            # Optional mask for targeted editing:
            # "mask": ("mask.png", open("mask.png", "rb"), "image/png"),
        }
    ).json()

if "error" in response:
    print(f"ERROR: {response['error']}")
else:
    img_data = base64.b64decode(response["data"][0]["b64_json"])
    with open("edited_image.png", "wb") as f:
        f.write(img_data)
    print("Edited image saved")

EOF

Use cases for editing:

• Background replacement: Generate a concept image, then replace its background to match slide theme
• Style transfer: Edit an existing diagram to add artistic elements
• Localization: Replace English text in an image with Chinese text

Error Handling

Error	Cause	Resolution
DeploymentNotFound	Wrong deployment name	Check GPT_IMAGE_DEPLOYMENT env var
401 Unauthorized	Invalid API key	Verify AZURE_OPENAI_API_KEY
429 Too Many Requests	Rate limit	Wait + retry with lower quality
content_policy_violation	Prompt blocked by content filter	Rephrase prompt, avoid specific faces/people
InvalidPayload	Bad parameters	Check size is multiple of 16, valid format

Guidelines

• Always use quality="high" for final presentation images. Use "medium" only for drafts/iteration.
• Prefer 1536x1024 as the default size for most PPT use cases (good balance of quality and speed).
• Transparent backgrounds (background="transparent") are useful for overlaying on colored slide backgrounds.
• Keep prompts detailed — vague prompts produce generic results. Include style, composition, colors, mood.
• Match the PPT color palette — reference the CLAUDE.md color scheme (orange #E8913A, blue #4472C4, etc.) in prompts.
• No text-heavy images — GPT-image-2 handles short labels well but struggles with paragraphs. Use slide text overlays instead.
• One image per prompt — generate n=1 for quality control. Batch generation (n>1) gives varied results.