中文文档 | English gallery | 中文案例库
Advertising prompt agent for Image2: turn business briefs and reference images into debranded, production-ready image prompts.
中文定位:面向 Image2 广告作图的开源 Prompt Agent,把客户白话需求、文案、行业和参考图转成可测试、可追踪、可集成的优化提示词。
This is not another prompt dump. Prompt libraries are useful for browsing examples; Image2 Ads Studio is a working prompt pipeline for advertising production. It parses a customer brief, retrieves advertising templates and visual recipes, rewrites references through an LLM prompt brain, validates deterministic execution details, and returns an optimized prompt that can be manually tested in Image2.
This repository is the Community Edition. It opens the agent framework, local Web workbench, all current 320 business templates, 240 visual recipes, LLM prompt brain interface, deterministic prompt checks, reference image policy, and a public case gallery.
Most Image2 prompt repositories answer: "What prompt can I copy?"
Image2 Ads Studio answers: "Given this business brief, copywriting, industry, aspect ratio, and reference image, what prompt should the agent produce?"
| Prompt libraries | Image2 Ads Studio |
|---|---|
| Curated examples for inspiration | Runnable advertising prompt agent |
| Start from existing prompts | Start from customer brief and reference images |
| General creative coverage | Advertising vertical: storefronts, posters, ecommerce, product ads, signage, brand walls, commercial photography |
| Prompt text is the product | Parser, retriever, compiler, LLM rewrite, deterministic checks, and JSON records are the product |
| Usually copy/paste driven | Produces parsed intent, matched templates, visual recipes, design plan, optimized prompt, and source policy |
| References may stay close to source | Public gallery prompts are regenerated, debranded, attributed, and made reusable |
| Hard to connect to business systems | Designed for later ERP fields, template governance, adapters, and production workflows |
The public gallery contains 100 image-to-prompt cases from project-generated concepts and multiple in-repo source libraries. Each case includes one preview image, English and Chinese optimized prompts, source attribution, and an in-repo composite source made from matched advertising templates plus visual recipes.
View the full gallery: English cases / 中文案例
Source and license notes are tracked in gallery attribution. The gallery is not a copied prompt collection: upstream prompts/images are used as structure references, then this project regenerates, debrands, and normalizes the final prompt through the same agent flow used by the Web workbench.
General image-generation prompts are unstable for advertising production. A local shop owner may say "make a milk tea storefront signboard", but production-ready output needs structure: task type, industry, copywriting, aspect ratio, material constraints, reference image policy, and clear composition.
Image2 Ads Studio turns that loose brief into a structured advertising plan and an optimized prompt that can be manually tested in Image2 or another image-generation tool.
- Advertising vertical coverage: storefront signboards, posters, roll-up banners, event backdrops, wayfinding signage, brand walls, local promotions, ecommerce main images, product ads, and commercial photography.
- Open template library: 320 business templates and 240 visual recipes.
- LLM prompt brain: uses a Responses-compatible LLM endpoint to refine the rule prompt.
- Reference image policy: user uploads can be read by the LLM and marked for later Image2 editing; upstream references are only used for visual understanding.
- Deterministic prompt validation: removes vague expressions and requires composition, lighting, typography, material, and image-reference policy.
- Local Web UI: prompt workbench with brief input, optimized prompt output, preview placeholder, and retrieval signals.
pnpm install
pnpm --filter ad-image-agent-core test
pnpm --filter ad-image-agent build
export OPENAI_API_KEY="your_api_key"
pnpm --filter ad-image-agent serve:llmOpen:
http://127.0.0.1:5174
The app does not call an image-generation API in this edition. Copy the optimized prompt into your image-generation tool and upload user reference images when the app marks them for Image2 participation.
flowchart LR
A[User Brief] --> B[Intent Parser]
B --> C[Template Retriever]
C --> D[Visual Recipe Retriever]
D --> E[LLM Prompt Brain]
E --> F[Optimized Prompt]
F --> G[Image2 Manual Test]
G --> H[Case Library]
H -. Template / Recipe Update .-> C
H -. Template / Recipe Update .-> D
| Capability | Community Edition | Commercial Edition |
|---|---|---|
| Local Web UI | Included | Included |
| Core prompt agent framework | Included | Included |
| Business templates | 320 open templates | Custom/private templates and governance |
| Visual recipes | 240 open recipes | Custom/private recipes and scored evaluation |
| LLM prompt brain | Interface + local server | Production setup |
| Image generation | Manual adapter | Real adapters and storage |
| ERP integration | Overview docs | Connectors, field mapping, deployment |
| Case gallery | Public samples | Private scored case library |
Commercial work can add ERP connectors, real image-generation adapters, deployment support, custom private templates, and production case evaluation.
apps/ad-image-agent Local Web workbench
packages/ad-image-agent-core Parser, retrievers, compiler, adapters, tests
examples Sample briefs and generated records
docs Architecture, adapter, ERP, authoring, roadmap
scripts Community validation and asset counting
pnpm --filter ad-image-agent-core test
pnpm --filter ad-image-agent-core build
pnpm --filter ad-image-agent typecheck
pnpm --filter ad-image-agent buildApache-2.0. See LICENSE.






