ecom-image2

Overview

Generate e-commerce images using GPT-Image-2 via Codex CLI. Match user intent to structured JSON prompt templates, assemble concise prompts, and invoke image generation.

Workflow

Step 1: Intent Recognition

From the user's request, extract:

• Scene type: hero image, lifestyle, flat lay, macro detail, poster/banner, social media, UGC, model showcase, before/after, packaging, infographic, creative concept, size spec, multi-product, livestream, virtual try-on, exploded view, ghost mannequin, multi-angle grid, magazine editorial, seasonal campaign, luxury atmospherics, device mockup, storefront, sports campaign
• Product info: category (beauty/electronics/food/fashion/home/jewelry/sports), description, material, key selling points
• Style preference: luxury, fresh, tech, minimal, or other variant
• Reference image: whether user provided a product photo path

If the user provides a product photo path, note it for --image parameter.

Step 2: Template Matching

Read the matching template from references/templates/. Match by scanning keywords and trigger_phrases in each template:

No match → default to 01-hero-image.json.

Only read the matched template file (progressive disclosure). Do not load all templates.

Step 3: Prompt Assembly

From the matched JSON template:

1. Take prompt_template as the base structure
2. Replace {variables} with user-provided info
3. If user specified a style variant → apply variants.<name>.overrides
4. If product category known → apply category_tips.<category>
5. Simplify: keep only core fields with values, remove empty/null fields
6. Output a concise JSON object (not the full template metadata)

Key principle: keep prompts simple. Only include essential information. Image2 performs best with concise, focused prompts rather than overly complex ones.

Example assembled prompt for a beauty hero image:

{
  "type": "product photography",
  "subject": "frosted glass serum bottle with matte white cap",
  "background": "clean white background",
  "lighting": "soft diffused studio lighting",
  "composition": "centered, front view",
  "quality": "8K, commercial e-commerce photography",
  "category_note": "emphasize texture and glow"
}

Step 4: Image Generation

Run the generation script:

bash scripts/imagegen.sh --prompt-file <(echo '<assembled_json>') --mode auto

Or call codex exec directly:

Without reference image:

codex exec --ephemeral --skip-git-repo-check --sandbox read-only --color never - <<< "Use imagegen to create an image with this request:
<assembled_json>

Requirements:
- Generate the image directly
- Do not provide explanation
- Return only the image result"

With reference image:

codex exec --ephemeral --skip-git-repo-check --sandbox read-only --color never \
  --image /path/to/ref.png \
  - <<< "Use imagegen to create an image with this request:
<assembled_json>

Reference image(s) are attached. Use them as visual identity/style references.
Requirements:
- Generate the image directly
- Do not provide explanation
- Return only the image result"

HTTP service mode: If curl -sf http://127.0.0.1:4312/health succeeds, submit via HTTP instead:

curl -sf -X POST http://127.0.0.1:4312/v1/images/generations \
  -H 'content-type: application/json' \
  -d '{"prompt":"<assembled_json>","images":["/path/to/ref.png"],"timeout_sec":180}'

Step 5: Result Cleanup

After generation, images are saved to ~/.codex/generated_images/<session_id>/. Must clean up:

1. Copy generated image to the user's working directory (or specified output path) and rename with descriptive name
2. Delete the original codex session folder to avoid duplicate storage:

rm -rf ~/.codex/generated_images/<session_id>

3. Report the final image path to the user

Step 6: Suggestions

If applicable, suggest:

• Try a different style variant (list available variants from template)
• Adjust product category for more tailored results
• Add a reference image for better product consistency
• Try a different scene type

Anti-AI Tips (for UGC / Livestream / Social Media scenes)

When generating UGC, livestream, or social media content, these rules are critical:

• Specify exact phone model: iPhone 14 Pro, iPhone 15 Pro
• Add visible imperfections: pores, slight noise, warm color cast, imperfect framing
• Use candid language: NOT professional photography, NOT AI-generated look
• Show real environment: slightly messy, real objects, water stains, used towels
• Reference film tone: Kodak Portra 400 color feel
• Explicitly state: NOT retouched, NOT smoothed
• Avoid AI-signature words: no perfect, flawless, stunning, hyper-realistic

Prompt Writing Guidelines

• Keep it simple: only core information, no excessive constraints
• Natural language preferred: Image2 understands descriptive sentences better than keyword lists
• Specify material: describe textures explicitly (frosted glass, brushed metal, matte finish)
• Lighting matters: always include lighting direction and quality
• Use references: passing a product photo via --image significantly improves consistency