Interactive Visual Task Guide • Production Edition

From Catalog to Brand-Aligned Images

This comprehensive manual walks you through all 16 core tasks of the PromptiQ platform — from ingesting fashion collections and tuning textual gradient descent loops to banking Brand Memory exemplars and training custom LoRA brand models.

⚡ 16 Visual Tasks🎯 85–100 Calibrated Target⏱️ 1.2s Real-Time Polling🧠 Visual RAG Flywheel

Task Index & Quick Jump Navigation

16 Tasks • 5 Sections

Section I — Pipeline Setup & Optimization

Section II — Brand Architecture & Calibration

Section III — Brand Studio (LoRA Training & Serving)

Section IV — Asset Management & Curation

Section V — Collaboration, Admin & Reference

The 4-Stage PromptiQ Architecture

Understand how PromptiQ transforms raw product photos into a self-learning brand intelligence system.

01Input Layer

Asset Ingestion & DNA

Images are uploaded to storage and converted into 512-dim visual embeddings with jina-clip-v2 alongside structured Brand DNA.

02Self-Improving Loop

TGD Optimization Loop

Generates candidate images, scores against product + brand centroid, and refines prompts via textual gradient descent.

03Compounding Moat

Brand Memory Bank

Prompts scoring ≥85 are banked as exemplars. Visual RAG injects top-3 matching proven vocabularies into future runs.

04Production Serving

Generation & Studio

Produce campaign concepts across Fast/Balanced/Quality presets or serve dedicated weights trained via Brand Studio LoRA.

Section I

Workspace Setup & Prompt Optimization Pipeline

01

Navigating the Workspace & Dashboard Overview

Primary App Shell • Navigation • Organization Metrics

Upon logging in to PromptiQ, the Dashboard serves as your mission control center. It surfaces aggregate brand statistics, active catalog counters, prompt optimization status, and brand model readiness.

promptiq.cloud/dashboard
PromptiQ Dashboard Overview Interface
Figure:PromptiQ primary dashboard view showing workspace metric cards, left navigation sidebar, top organization switcher, and live catalog activity.
Sidebar NavigationMetric Summary CardsTheme Switcher

Step-by-Step Instructions:

  1. 1.Access Primary Navigation: Use the persistent left sidebar to switch between Dashboard, Catalogs, Generate, Concepts, Collections, Prompts, and Settings.
  2. 2.Monitor Workspace KPIs: Review the 4 primary metric cards tracking Total Catalogs, Optimized Prompts, Generated Concepts, and Brand Model status.
  3. 3.Customize Preferences: Open the top header profile dropdown to toggle between Light, Dark, or System visual themes.
02

Creating a Catalog & Exploring Sample Data

Catalog Management • Costa Sole SS25 Pre-Loaded Data

A Catalog groups related product images (e.g., a seasonal drop, product line, or capsule collection). PromptiQ ships pre-seeded with the Costa Sole SS25 Resort Collection sample catalog so you can inspect an 8-look production catalog immediately without uploading files.

promptiq.cloud/catalogs
PromptiQ Catalogs Page and New Catalog Creation Modal
Figure:Catalogs listing page displaying existing seasonal collections and the 'New Catalog' dialog for defining collection titles and descriptions.
Costa Sole SS25 SampleNew Catalog ModalCatalog Cards

Step-by-Step Instructions:

  1. 1.Navigate to Catalogs: Click Catalogs in the left sidebar.
  2. 2.Create a New Catalog: Click New Catalog in the top-right corner. Enter a descriptive title (e.g., Autumn Winter 2026 Outerwear) and optional collection notes.
  3. 3.Explore Pre-Seeded Sample Data: Click the Costa Sole SS25 catalog to explore 8 fully-calibrated looks (Silk Slip, Linen Seated, Archway, Knit Coord, etc.) with pre-optimized prompts.
Catalog Sizing Recommendation
For optimal brand centroid convergence, group 10 to 50 focused product assets per catalog. Oversized catalogs with mixed aesthetic lines dilute the brand centroid embedding.
03

Uploading & Managing Product Assets

Asset Ingestion • Supabase Storage • 512-dim CLIP Vectors

Each uploaded photo becomes an asset that feeds PromptiQ's vision model. When an asset is uploaded, it is stored in Supabase and automatically transformed into a 512-dimensional vector fingerprint via jina-clip-v2.

promptiq.cloud/catalogs/costa-sole-ss25
Catalog Asset Upload Dropzone and Grid View
Figure:Catalog detail view displaying the drag-and-drop asset ingestion dropzone and the grid of uploaded product photography.
Drag & Drop DropzoneBatch IngestionCLIP Extraction

Step-by-Step Instructions:

  1. 1.Open Catalog Detail: Select your target catalog from the Catalogs list.
  2. 2.Upload Images: Drag and drop JPG or PNG files into the upload dropzone (or click to browse). Support up to 100 images per catalog.
  3. 3.Inspect Processing Status: Each asset appears in the grid with a thumbnail, filename, and initial status (ready for prompt optimization).
Image Requirements
For optimal CLIP embedding fidelity, use high-resolution studio shots (minimum 512 × 512px) on neutral or architectural backdrops. Avoid heavy graphical overlays or text watermarks on source photos.
04

Configuring Structured Brand DNA

6-Dimension Profile • Style Directives • Strict Avoid-List

Brand DNA provides the explicit editorial rules and creative boundaries that govern every prompt draft, critique, and refinement pass. Unlike generic prompts, PromptiQ injects this structured schema into the LLM critique loop to enforce brand guidelines programmatically.

promptiq.cloud/settings?tab=brand-dna
Structured Brand DNA Configuration Form
Figure:Structured Brand DNA profile showing the 6 core dimensions: visual style, brand voice, signature elements, photography style, and the critique avoid-list.
Visual Style & PaletteBrand VoiceAvoid-List Guardrail
1. Visual Style & Color Palette

Explicit hex/pantone references, material textures, and mood (e.g., 'warm neutrals, ivory, sand, single lilac accent').

2. Brand Voice

Editorial phrasing tone (e.g., 'understated, confident, high-end editorial, precise garment cut descriptions').

3. Signature Elements

Recurring design codes: bias-cut drapery, natural linen slub, architectural lapels, minimalist seams.

4. Photography Style

Lighting and composition rules (e.g., 'soft Mediterranean golden-hour sunlight, medium format 80mm lens feel').

5. Critique Avoid-List

Strictly forbidden elements (e.g., 'harsh direct flash, synthetic sheen, neon highlights, busy logos'). Stripped during TGD critique.

6. Concrete Vocabulary

Use precise fashion nouns ('double-breasted crepe blazer') rather than generic adjectives ('pretty clothes').

Why the Avoid-List is Crucial
During textual gradient descent, the LLM critique agent compares candidate images against your avoid-list. Any detected violation triggers immediate prompt rewrites to eliminate off-brand styling.
05

Setting Up Prompt Optimization (TGD Loop)

Textual Gradient Descent • Parameter Setup • Inngest Orchestration

Prompt optimization turns raw photos into high-fidelity generative prompts using Textual Gradient Descent (TGD). An AI loop drafts, generates, critiques, and refines prompts iteratively until your target score is achieved.

promptiq.cloud/catalogs/[id]/optimize
Prompt Optimization Setup Modal
Figure:Optimization setup dialog allowing selection of iteration passes (3, 5, 7, 10) and Target Brand Fidelity threshold (0.70–1.00).
Iterations: 3–10Target Fidelity: 0.85Inngest Durable Job

Step-by-Step Instructions:

  1. 1.Initiate Optimization: Click Optimize All from the catalog toolbar (or select individual checkboxes and click Optimize Selected).
  2. 2.Set Iteration Passes: Select 3, 5, 7, or 10 passes (Default: 5). 5 passes provides an optimal balance between convergence speed and compute efficiency.
  3. 3.Configure Target Brand Fidelity: Set your target score between 0.70 and 1.00 (Default: 0.85, representing 85 on the calibrated scale).
  4. 4.Launch Background Job: Click Start Optimization. The job is handed off to Inngest durable workflows, routing you to the live progress page.
06

Monitoring Live Optimization & Score Climbing

1.2s Real-Time Polling • Climbing Similarity Gauge • Plateau Early-Stop

The live progress page (/optimize/[jobId]) connects to the server and polls every 1.2 seconds. You can watch the Similarity Score gauge climb in real time as the critique agent tunes prompt language.

promptiq.cloud/optimize/[jobId]
Live Optimization Progress with Climbing Gauge
Figure:Live optimization monitor showing real-time iteration passes, candidate preview renders, and the climbing similarity score gauge.
1.2s PollingClimbing Score GaugeEarly Plateau Stop

Loop Termination Conditions:

Target Cleared

Stops immediately when candidate score hits or exceeds target (≥85).

Plateau Early-Stop

Stops if score does not improve by >1.0pt over 2 consecutive passes, saving compute cost.

Max Passes Reached

Finalizes after completing the selected number of iteration passes (e.g. 5 passes).

Section II

Brand Architecture, Fidelity Calibration & Memory Flywheel

07

Interpreting the Calibrated Brand Fidelity Score

Calibrated 0–100 Scale • 0.7 Product / 0.3 Centroid Blend • Score Bands

PromptiQ measures brand alignment with a Calibrated Brand Fidelity Score (0–100). This is not raw mathematical cosine similarity; it is calibrated so that true stylistic drift lands near zero and authentic brand coherence clears 85+.

promptiq.cloud/catalogs/costa-sole-ss25
Calibrated Brand Fidelity Score Badges in Catalog Grid
Figure:Catalog detail grid showing calibrated Brand Fidelity badges (e.g., 90/100 Green On-Brand, 86/100, 88/100) across Costa Sole SS25 assets.
85–100 On-Brand70–84 Close0.7/0.3 Centroid Blend

The 3 Calibrated Scoring Bands:

85 – 100ON-BRAND (TARGET CLEARED)
Green Badge

Consistent garment silhouette, authentic color palette, correct fabric texture, and compliance with Brand DNA. Banked automatically to Brand Memory.

70 – 84CLOSE (NEEDS REFINEMENT)
Amber Badge

Recognizable garment direction but exhibits minor color drift, background noise, or missed drapery nuance. Recommended to re-run with +2 passes.

< 60OFF-BRAND DRIFT
Red Badge

Significant aesthetic divergence. Check the source photo for background clutter or update Brand DNA avoid-list before re-optimizing.

Mathematical Centroid Blend Formula

PromptiQ calculates each candidate score using a weighted blend:

Brand Fidelity = (0.7 × Product_Similarity) + (0.3 × Catalog_Centroid_Cohesion)

This guarantees generated looks match the specific target product while staying cohesive with the entire collection's visual identity.

Benchmark Calibration Notice
Legacy guidance citing “70–85% as excellent” reflected uncalibrated raw cosine math. Under PromptiQ's calibrated system, 85+ is the true on-brand benchmark.
08

Leveraging Brand Memory & Exemplar Flywheel

Visual RAG • Automatic Exemplar Banking • Compounding Accuracy

PromptiQ becomes smarter the more you use it. Every prompt that clears the ≥85 On-Brand threshold is automatically saved to your organization's brand_exemplars vector bank.

promptiq.cloud/brand-memory
Brand Memory Exemplar Retrieval Diagram
Figure:Brand Memory Visual RAG architecture showing cosine similarity retrieval of top-3 approved exemplars feeding proven vocabulary into future runs.
≥85 Auto-BankingTop-3 Vector RAGCompounding Quality

How the Memory Flywheel Operates:

  1. 1.Automatic Banking: When any asset reaches a Brand Fidelity Score ≥85, its prompt, CLIP vector, and metadata are banked into your private brand memory.
  2. 2.Visual RAG Retrieval: When you optimize a new asset, PromptiQ searches your memory bank using cosine similarity and retrieves the top-3 most visually related winning prompts.
  3. 3.Vocabulary Seeding: The LLM seeds its initial draft with proven vocabulary (e.g. your specific lighting terms, fabric descriptors), achieving faster convergence in fewer iterations.
Re-Optimization Compounding Benefit
If you re-optimize an older asset after building an exemplar bank, it will often score higher than on its initial run because it leverages accumulated brand memory.
09

Generating Images from Optimized Prompts & Presets

From Catalog Tab • Quality Presets • Fast / Balanced / Quality

With optimized prompts stored in your catalog, head to the Generate page to create high-resolution campaign imagery using tailored quality presets.

promptiq.cloud/generate
PromptiQ Generate Page with From Catalog Tab and Presets
Figure:Generate page displaying the 'From Catalog' prompt loader, prompt editor, aspect ratio selector, and Quality Presets.
From Catalog TabQuality PresetsInstant Render

Selecting the Right Quality Preset:

Fast~10 sec
Seedream-3.0

Rapid concept ideation and storyboard drafting. Best for testing multiple creative variants quickly.

Balanced~25 sec
Gemini 2.5 Flash

Superior garment comprehension and drape nuance. Recommended for lookbook and e-commerce shots.

Quality~60 sec
HiDream-I1-Full

Maximum textural fidelity and photorealism. Designed for hero campaign banners and print collateral.

Section III

Brand Studio — Custom LoRA Fine-Tuning & In-App Serving

10

Training a Custom Brand Model in Brand Studio

Settings → Brand Studio • Cost Acknowledgement • Quality Gate

On the Brand Studio tier, you can go beyond prompt tuning and fine-tune dedicated image-generation weights (a custom FLUX LoRA). Your brand model renders in your house style at the weight level.

promptiq.cloud/settings?tab=brand-studio
Brand Studio Training Modal and Cost Acknowledgement Checkbox
Figure:Brand Studio model card showing active status and the training confirmation modal with required cost-acknowledgement authorization.
FLUX LoRA Fine-TuningCost AuthorizationQuality Gate

Step-by-Step Training Walkthrough:

  1. 1.Open Brand Studio: Navigate to Settings → Brand Studio tab.
  2. 2.Inspect Model Status: Review your current card state (No Model, Training, Active, or Failed).
  3. 3.Launch Training Modal: Click Train Brand Model (or Retrain). Review the dataset size (50–200 approved images recommended) and estimated GPU cost.
  4. 4.Authorize Spend: Check the required Cost Acknowledgement checkbox (“I understand this is a paid GPU training run and authorize the spend”).
  5. 5.Background Processing & Quality Gate: Training runs on Modal GPUs with weights saved to Cloudflare R2. A self-verifying quality gate tests the model against the prompt-only baseline; it is only promoted to production if it strictly beats the baseline score.
Retrain Cadence Guardrails
To prevent accidental compute overruns, automatic retrains are rate-limited to once every 14 days per model, and manual retrains are limited to once per 24-hour window.
11

Generating with Brand Model & Side-by-Side Comparison

✨ Use Brand Model Toggle • Engine Badges • 2-Up Side-by-Side View

Once your Brand Model is active, the Generate page unlocks direct weight-level inference and an interactive 2-up comparison tool to evaluate baseline output versus your trained model.

promptiq.cloud/generate?compare=true
Side-by-Side 2-Up Brand Model Comparison View
Figure:Side-by-side labeled 2-up comparison showing baseline prompt generation versus custom FLUX LoRA brand model generation.
✨ Brand Model Toggle2-Up ComparisonEngine Provenance

Generation & Comparison Steps:

  1. 1.Toggle Brand Model: On the Generate page, switch on “✨ Use my Brand Model”.
  2. 2.Inspect Provenance Badge: Generated images display an engine provenance tag (e.g. replicate:flux-dev-lora) confirming custom weights produced the pixels.
  3. 3.Run 2-Up Comparison: Click “Compare with & without Brand Model” to run the prompt through both pipelines simultaneously and inspect side-by-side results.
Section IV

Asset Management, Curation & Prompt Library

12

Managing Concepts (Lightbox, Notes, & Favorites)

Concepts Gallery • Fullscreen Lightbox • Star Favorites • Note History

Every generation is automatically cataloged in the Concepts gallery. The interactive Lightbox lets you inspect full generation metadata, edit creative notes, star favorites, and trigger sharing.

promptiq.cloud/concepts
Concepts Gallery and Fullscreen Lightbox Modal
Figure:Concepts workspace showing grid filtering, favorite starring, and the full-screen Lightbox with prompt text, model metadata, and feedback notes.
Lightbox InspectionInteractive NotesStar Favorites

Lightbox Capabilities:

  • Full Technical MetadataInspect exact prompt text, generation engine, quality preset, resolution, and timestamp.
  • Creative Notes CRUDAdd, edit, or delete directorial notes directly on the image record for team collaboration.
  • Favorite StarringClick the star icon to bookmark hero looks and filter the gallery with the “Favorites only” toggle.
  • Instant Re-generationPre-fill the prompt editor directly from the concept to spin up variations.
13

Organizing Concepts into Curated Collections

Multi-Select Action Bar • Collection Boards • Mood Board Curation

Collections let you curate concepts into thematic boards for marketing campaigns, editorial lookbooks, or stakeholder sign-off packages.

promptiq.cloud/collections
Multi-Select Concepts and Add to Collection Bar
Figure:Concepts gallery showing multi-select hover checkboxes, the floating bulk action bar, and the 'Add to Collection' modal.
Multi-Select CheckboxesFloating Action BarCurated Boards

How to Build a Curated Collection:

  1. 1.Multi-Select Concepts: In the Concepts gallery, hover over concept cards and check the selection boxes.
  2. 2.Add to Collection: Click Add to Collection in the floating bottom action bar.
  3. 3.Create or Assign: Choose an existing collection or click Create New Collection (enter name and description).
  4. 4.Manage Board: Open Collections in the sidebar to review, reorder, or export your curated set.
14

Managing the Prompt Library

Prompts Workspace • Category Filters • Tag Search • Usage Tracking

The Prompt Library is your organization's repository of standalone creative starting points. It tracks team usage counts, category tags, and allows one-click generation pre-fills.

promptiq.cloud/prompts
Prompt Library Workspace with Search and Tags
Figure:Prompt Library page showing category tags, real-time text search, usage counter badges, and the direct 'Generate' action button.
Category FilteringReal-Time SearchUsage Counter

Prompt Library Best Practices:

  1. 1.Save Reusable Directives: Click New Prompt to store high-performing baseline structures with categories (e.g., Editorial, Lookbook, Macro Detail).
  2. 2.Filter & Search: Search by title, keyword, or tag to instantly locate proven styling formulas across your team.
  3. 3.One-Click Generation: Click Generate on any card to load the prompt directly into the Generate studio and increment its team usage counter.
Section V

Collaboration, Workspace Settings & Public Sharing

15

Public Sharing & Stakeholder Reviews

Zero-Login Sharing • /share/[id] • Read-Only Banner • View Analytics

Share concepts with external creative directors, clients, or agency partners without requiring them to create an account or sign in.

promptiq.cloud/share/c9f28a7e
Public Read-Only Concept Share View
Figure:Public standalone share page (/share/[token]) showing high-res asset render, prompt details, model provenance, and the read-only banner.
No Login RequiredShared View BannerView Tracking

Sharing Workflow:

  1. 1.Generate Share Link: In the Lightbox, click Share. A cryptographically unique share URL is generated.
  2. 2.Distribute Link: Send the link (e.g. https://www.promptiq.cloud/share/[token]) to external reviewers.
  3. 3.Recipient Experience: Viewers see a clean presentation with high-res imagery, generation parameters, and an amber Shared View (Read-Only) banner. Each unique access increments view analytics.
16

Workspace Settings, Team Management, & API Usage

Settings Tabs • General • Team • Brand Studio • Billing • API Usage

The Settings workspace provides comprehensive administration across 5 dedicated tabs, allowing you to configure defaults, invite colleagues, manage LoRA models, and monitor GPU cost burn.

promptiq.cloud/settings
PromptiQ Settings Workspace Tabs
Figure:Settings workspace overview displaying all 5 administration tabs: General defaults, Team invitations, Brand Studio, Billing, and API Usage metrics.
5 Dedicated TabsTeam InvitationsAPI Cost Burn
General Tab
  • • UI Theme preference (Light, Dark, System)
  • • Default Target Brand Fidelity (0.70–1.00)
  • • Default Max Optimization Iterations (3–10)
Team Tab
  • • Invite team members via email
  • • Assign roles (Designer or Viewer)
  • • Track pending invitations & active seats
Brand Studio Tab
  • • Manage custom LoRA model lifecycle
  • • Trigger paid training & view weights
  • • Enforces 14-day auto / 1-day manual caps
Billing Tab
  • • Subscription plan tier and seat counts
  • • Upgrade options to unlock Brand Studio
  • • Invoices and payment methods
API Usage Tab
  • • Total image generations and LLM token consumption
  • • Provider breakdown (Together AI, Google Gemini, Replicate)
  • • Real-time monthly GPU spend tracking

Frequently Asked Questions

Common questions regarding scoring benchmarks, optimization mechanics, and model training.

Q:How long does prompt optimization take?

A single asset at 5 iterations takes approximately 1–3 minutes. A full catalog of 10 assets runs in parallel and completes in 5–12 minutes. The progress monitor updates in real time with 1.2-second polling.

Q:What Brand Fidelity Score should our creative team aim for?

Aim for 85 or higher — that is the calibrated on-brand target where consistent color palette, garment silhouette, and styling converge. 70–84 indicates close alignment needing minor prompt refinement. Below 60 indicates stylistic drift.

Q:Why did our optimization stop before the maximum iterations?

The loop terminates early if it clears the target score (≥85) OR if the score plateaus (<1.0pt gain over 2 consecutive passes). This early-stop feature prevents unnecessary GPU costs when further passes won't improve fidelity.

Q:How does Brand Memory improve our results over time?

Every winning prompt scoring ≥85 is banked into brand_exemplars. On future runs, Visual RAG retrieves the top-3 most similar exemplars to seed proven fashion vocabulary into the initial prompt draft.

Q:What is the difference between Prompt Optimization and Brand Studio?

Prompt optimization tunes language prompts against your Brand Fidelity target. Brand Studio (premium) fine-tunes dedicated LoRA model weights on your catalog so images render in your house style natively at the weight level.

Q:Can team members edit prompts before generating images?

Yes. In the Generate → 'From Catalog' tab, clicking any optimized prompt pre-fills the prompt editor where you can freely adjust styling, angles, and details before generating.

Q:What image formats are supported for catalog uploads?

JPG and PNG formats are supported up to 100 images per catalog. Product photography with clean, well-lit backgrounds yields the highest CLIP embedding accuracy.

Q:Can we share concepts with external clients who don't have accounts?

Yes. Use the Lightbox 'Share' button to create a public URL (e.g. promptiq.cloud/share/[token]). External reviewers see a high-res read-only view with zero login required.

Ready to Build Your Brand-Aligned AI Pipeline?

Log in to your PromptiQ workspace to start ingesting catalogs, optimizing prompts, and exploring custom brand models.

Go to DashboardCreate Organization Account