Ad-Ready: AI Advertising Image Generator

Generate professional advertising images from product URLs using a 4-phase AI pipeline on ComfyDeploy.

Source: github.com/PauldeLavallaz/ads_SV

Pipeline Architecture

The pipeline runs as a ComfyUI custom node deployed on ComfyDeploy. A single ProductToAds_Manual node executes 4 phases internally:

┌─────────────────────────────────────────────────────────────┐
│                  ProductToAds_Manual Node                     │
│                                                             │
│  PHASE 1: Product Scraping (Gemini Flash)                   │
│  ─────────────────────────────────────────                   │
│  Scrapes product URL → extracts title, description,         │
│  features, price, materials, image URLs                      │
│  Also scrapes HTML for high-res product images (≥1000px)    │
│                                                             │
│  PHASE 2: Campaign Brief Generation (Gemini Flash)          │
│  ────────────────────────────────────────────────            │
│  Brand Identity + Product Data + References →                │
│  10-point Campaign Brief (creative direction)                │
│                                                             │
│  PHASE 3: Blueprint Generation (Gemini Flash)               │
│  ──────────────────────────────────────────────              │
│  Master Prompt (funnel stage) + Brief + Keywords →           │
│  Production-Ready JSON Blueprint                             │
│                                                             │
│  PHASE 4: Image Generation (Nano Banana Pro / Imagen 3)     │
│  ──────────────────────────────────────────────────          │
│  Blueprint + all reference images → final ad image           │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Phase 2: Campaign Brief (The Creative Brain)

The Brief Generator is the most critical intermediate step. It acts as a "Senior Art Director" that translates raw data into actionable creative direction using a 10-point framework:

Strategic Objective — Why this campaign exists (awareness/positioning/launch)
Central Message — One idea perceivable without text
Visual Tone of Voice — Register: calm/energetic/intimate/monumental
Product Role — Hero vs co-protagonist vs implicit presence
Visual Language & Brand Coherence — Non-negotiable brand codes
Photographer & Equipment — Photography as concept, not execution
Extended Art Direction — Styling, casting, poses, hair/makeup, layout
Environment & Context — Where and why (conceptual, never decorative)
Texture, Material & Product Render — How surfaces are perceived
Final Image Signature — Finish, grain, temporal positioning

Without the brief, the Master Prompt must guess creative strategy. With it, the Master Prompt only executes.

The brief prompt template is included at {baseDir}/configs/Brief_Generator/brief_prompt.json.

Phase 3: Master Prompts (8 Funnel Stages)

Each funnel stage has a specialized Master Prompt that generates a production-ready JSON Blueprint. All share the same internal simulation:

ROUND -1: Brand Identity Forensics (stages 03+) — Unified Brand Style Manifest
ROUND 0: Fidelity Lock — Product geometry & talent identity are IMMUTABLE
ROUND 1: Stage Strategy — Strategic approach specific to funnel position
ROUND 2: Graphic Design — UI, typography, CTA engineering

The Blueprint JSON covers: scene production, talent lock, camera perspective, subject action/pose/wardrobe, lighting, product constraints, layout architecture, typography, CTA engineering, and brand asset placement.

Master prompt files are included at {baseDir}/configs/Product_to_Ads/.

Reference Analyzer

When reference images (pose, style, location) are provided, they're analyzed with forensic precision:

POSE_REF → Body position, limbs, weight, gaze, micro-gestures → replicated EXACTLY
PHOTO_STYLE_REF → Camera, lens, lighting, grading, grain → derived parameters
LOCATION_REF → Setting, materials, colors, mood → similar but creatively enhanced

The reference analysis prompt is at {baseDir}/configs/Reference_Analyzer/reference_analysis_prompt.txt.

⚠️ CRITICAL: Required Inputs Checklist

Before running ANY ad generation, ensure ALL of these are provided:

Input	Required?	How to Get It
`--product-url`	✅ ALWAYS	User provides the product page URL
`--product-image`	✅ ALWAYS	Download from the product page, or user provides
`--logo`	✅ ALWAYS	Download from brand website or search online. MUST be an image file
`--reference`	✅ RECOMMENDED	An existing ad whose style we want to clone. Search online or use previously generated images
`--brand-profile`	✅ NEVER EMPTY	Pick from catalog or run brand-analyzer first. NEVER leave as "No Brand" if a brand is known
`--prompt-profile`	✅ ALWAYS	Choose based on campaign objective
`--aspect-ratio`	Default: 4:5	Change if needed for platform
`--model`	✅ RECOMMENDED	Model/talent face. Ads with talent perform much better

🚨 NEVER Skip These Steps:

Product image — Download the main product photo from the product URL. The scraper is fragile; always provide a product image explicitly.
Brand logo — Download the logo from the brand's official website or search for "{brand name} logo" online. Must be a clean logo image (PNG preferred).
Brand profile — If the brand doesn't exist in the catalog, run brand-analyzer skill FIRST to generate one. Never submit with "No Brand" when a brand is known.
Reference image — Search for an existing ad or visual with a style that matches what we're generating. This dramatically improves output quality.

Auto-Preparation Workflow

When the user asks to generate an ad:

1. User provides: product URL + brand name + objective

2. CHECK brand profile exists:
   → ls ~/clawd/ad-ready/configs/Brands/ | grep -i "{brand}"
   → If not found: run brand-analyzer skill first

3. DOWNLOAD product image:
   → Visit the product URL or fetch the page
   → Find and download the main product image
   → Save to /tmp/ad-ready-product.jpg

4. DOWNLOAD brand logo:
   → Search "{brand name} logo PNG" or fetch from brand website
   → Download clean logo image
   → Save to /tmp/ad-ready-logo.png

5. FIND reference image:
   → Search for "{brand name} advertisement" or similar
   → Or use a previously generated ad that has the right style
   → Save to /tmp/ad-ready-reference.jpg

6. SELECT prompt profile based on objective:
   → Awareness: brand discovery, first impressions
   → Interest: engagement, curiosity
   → Consideration: comparison, features
   → Evaluation: deep dive, trust, proof
   → Conversion: purchase intent, CTAs (most common)
   → Retention: post-purchase confidence
   → Loyalty: emotional bond, lifestyle
   → Advocacy: social amplification, community

7. RUN the generation with ALL inputs filled

Usage

Full command (recommended):

COMFY_DEPLOY_API_KEY="$KEY" uv run {baseDir}/scripts/generate.py \
  --product-url "https://shop.example.com/product" \
  --product-image "/tmp/product-photo.jpg" \
  --logo "/tmp/brand-logo.png" \
  --reference "/tmp/reference-ad.jpg" \
  --model "models-catalog/catalog/images/model_15.jpg" \
  --brand-profile "Nike" \
  --prompt-profile "Master_prompt_05_Conversion" \
  --aspect-ratio "4:5" \
  --output "ad-output.png"

Auto-fetch mode (downloads product image and logo automatically):

COMFY_DEPLOY_API_KEY="$KEY" uv run {baseDir}/scripts/generate.py \
  --product-url "https://shop.example.com/product" \
  --brand-profile "Nike" \
  --prompt-profile "Master_prompt_05_Conversion" \
  --auto-fetch \
  --output "ad-output.png"

List available brands:

uv run {baseDir}/scripts/generate.py --list-brands

API Details

Endpoint: https://api.comfydeploy.com/api/run/deployment/queue Deployment ID: e37318e6-ef21-4aab-bc90-8fb29624cd15

ComfyDeploy Input Variables

Variable	Type	Description
`product_url`	string	Product page URL to scrape
`producto`	image URL	Product image (uploaded to ComfyDeploy)
`model`	image URL	Model/talent face reference
`referencia`	image URL	Style reference ad image (used for both pose + location)
`marca`	image URL	Brand logo image
`brand_profile`	enum	Brand name from catalog (70+ brands)
`prompt_profile`	enum	Funnel stage master prompt
`aspect_ratio`	enum	Output format (1:1, 4:5, 5:4, 9:16, etc.)

Funnel Stages — Strategic Detail

01 — Awareness

Goal: Scroll-stop, curiosity, brand introduction Reject: Generic "product on table" concepts Strategy: Dynamic camera angles, world-building environments, high-concept creativity CTA: Soft or optional Visual Hierarchy: Talent → Product → Optional CTA

02 — Interest

Goal: Sustained attention, introduce value proposition Reject: Abstract visuals that hide the product Strategy: One clear visual idea, believable micro-world hinting at use-case CTA: Learn More, Discover, See Details Visual Hierarchy: Talent → Product → Headline → CTA

03 — Consideration

Goal: Informed evaluation, reduce uncertainty Reject: Pure mood storytelling, vague emotional content Strategy: Communicate WHAT product does, ONE primary differentiator, ONE proof cue CTA: Compare, See Details, Explore Visual Hierarchy: Talent → Product → Key Benefit → Proof Cue → CTA New: Adds Brand Identity Manifest to Blueprint JSON

04 — Evaluation

Goal: Validate purchase decision, proof & trust Reject: Pure mood, unsupportable claims, visual clutter Strategy: One trust anchor (quality/legitimacy/authority), one proof cue (reviews/certification) CTA: See Reviews, Verified Quality, Learn More Visual Hierarchy: Trust Anchor → Proof Cue → Product → Talent → CTA

05 — Conversion

Goal: Trigger decisive action, remove friction Reject: New hesitation-inducing info, complex compositions Strategy: One hero (product), one action, optional micro-reassurance CTA: Buy Now, Get Yours, Complete Order (PRIMARY visual element) Visual Hierarchy: Product → CTA → Optional Reassurance → Brand → Talent

06 — Retention

Goal: Post-purchase confidence, reduce churn Reject: Hard-sell, urgency, price talk Strategy: "You made the right choice" + "Here is the next step" CTA: Start, Set Up, Learn, Track (guidance, not purchase) Visual Hierarchy: Confirmation → Next Step → Product → Talent

07 — Loyalty

Goal: Strengthen emotional bond over time Reject: Sales layouts, instructional tone, aggressive CTAs Strategy: "This brand is part of who you are" — habitual engagement CTA: Optional: Explore, Be Part Of, Continue Visual Hierarchy: Brand World/Mood → Talent (identity mirror) → Product → Brand

08 — Advocacy

Goal: Turn customers into voluntary brand ambassadors Reject: Sales language, instructional tone, forced testimonials Strategy: Signal belonging, create share-worthy imagery, enable organic sharing CTA: Optional or absent: Join the Movement, Part of Us Visual Hierarchy: Mood → Talent (identity proxy) → Product (symbol) → Brand

Creating New Ad Types

To create a new funnel stage or specialized ad type:

Copy the closest existing Master Prompt from {baseDir}/configs/Product_to_Ads/
Redefine ROUND 1 with the new strategic objective
Adjust ROUND 2 UI hierarchy accordingly
Shift talent/product narrative roles
Modify CTA philosophy and copy voice
Keep the JSON output structure identical for pipeline compatibility
Maintain the Fidelity Lock (ROUND 0) — product and talent are always immutable
Save as Master_prompt_XX_NewStage.json — the node auto-discovers new profiles

Key Evolution Pattern Across Stages:

Aspect	Early (01-02)	Mid (03-05)	Late (06-08)
Talent role	Attention anchor	Credibility anchor	Identity mirror
Product role	Secondary hero	Evaluation hero	Familiar symbol
CTA	Soft/exploratory	Proof-led → Decisive	Guidance → Optional
Copy voice	Intriguing	Clarity, proof, action	Supportive → Proud
Visual density	High-concept	Structured, scannable	Editorial, spacious
Environment	World-building	Context-rich	Lifestyle, intimate

Image Input Types

Binding Images (strict fidelity — immutable)

talent: Face/body locked, no deviation in facial structure, ethnicity, proportions
product_1-4: Shape, label text, material, proportions preserved 1:1
brand_logo: UI/button style derived from logo geometry

Soft References (creative guidance)

pose_ref: Body position replicated EXACTLY (spine, limbs, weight, gaze, micro-gestures)
photo_style_ref: Camera/lighting/grading/grain derived (can be too literal)
location_ref: Environment inspired but creatively enhanced

In the live deployment, the same reference image feeds both pose_ref and location_ref.

Brand Profiles

Catalog (70+ brands):

ls ~/clawd/ad-ready/configs/Brands/*.json | sed 's/.*\///' | sed 's/\.json//'

Creating new brand profiles:

Use the brand-analyzer skill:

GEMINI_API_KEY="$KEY" uv run ~/.clawdbot/skills/brand-analyzer/scripts/analyze.py \
  --brand "Brand Name" --auto-save

The Brand Analyzer uses a 3-phase methodology:

Phase 1: Official research via Google Search (canonical data: name, founding, positioning, vision, mission, tagline)
Phase 1.1: Independent campaign research (10+ distinct campaigns via Google Images/Pinterest)
Phase 2-3: Visual analysis → JSON profile following the standard template

Output covers: brand_info, brand_values, target_audience, tone_of_voice, visual_identity, photography, campaign_guidelines, brand_behavior, channel_expression, compliance.

Aspect Ratios

Ratio	Use Case
`4:5`	Default. Instagram feed, Facebook
`9:16`	Stories, Reels, TikTok
`1:1`	Square posts
`16:9`	YouTube, landscape banners
`5:4`	Alternative landscape
`2:3`	Pinterest
`3:4`	Portrait

Config Files Reference

The skill includes reference copies of all pipeline configuration files:

{baseDir}/configs/
├── Brief_Generator/
│   └── brief_prompt.json              # 10-point campaign brief framework
├── Product_to_Ads/
│   ├── Master_prompt_01_Awareness.json
│   ├── Master_prompt_02_Interest.json
│   ├── Master_prompt_03_Consideration.json
│   ├── Master_prompt_04_Evaluation.json
│   ├── Master_prompt_05_Conversion.json
│   ├── Master_prompt_06_Retention.json
│   ├── Master_prompt_07_Loyalty.json
│   └── Master_prompt_08_Advocacy.json
└── Reference_Analyzer/
    └── reference_analysis_prompt.txt   # Pose/style/location analysis prompt

These configs are the canonical reference for the pipeline's behavior. The actual live configs are stored in the ComfyUI deployment at ads_SV/configs/.

Known Limitations

Product image scraping is fragile — always provide product images manually
photo_style_ref can be too literal — style reference may be replicated too closely
Some websites block scraping — provide product data manually when scraping fails
Single reference = pose + location — live deployment uses one image for both
Gemini hallucinations — occasional issues in complex reasoning steps
No brief editing — brief is generated automatically; manual override not yet supported

Ad-Ready vs Morpheus

Feature	Ad-Ready	Morpheus
Input	Product URL (auto-scrapes)	Manual product image
Brand intelligence	70+ brand profiles	None
Funnel targeting	8 funnel stages	None
Brief generation	Auto (10-point creative direction)	None
Creative direction	Objective-driven (brief → blueprint)	Pack-based (camera, lens, lighting)
Best for	Product advertising campaigns	Fashion/lifestyle editorial photography
Control level	High-level (strategy-first)	Granular (every visual parameter)

API Key

Uses ComfyDeploy API key. Set via COMFY_DEPLOY_API_KEY environment variable.

Source Repository

GitHub: PauldeLavallaz/ads_SV
Architecture: ComfyUI custom node package with 3 nodes:
- ProductToAds_Manual — Full manual control, single format
- ProductToAds_Auto — Auto-downloads images, generates 4 formats
- BrandIdentityAnalyzer — Analyzes brands via Gemini + Google Search