AI Image Generation

Automated product imagery at scale for e-commerce catalogues, fashion, furniture, and consumer goods brands that need contextual photography faster than physical shoot cycles allow. Model selection by output requirement: DALL-E 3 via OpenAI API for general lifestyle contexts with strong prompt adherence; Stable Diffusion XL or Flux via Replicate API (managed GPU inference, no infrastructure to maintain) when fine-tuning on specific product styles is required. Background removal on source product images using Remove.bg API or the open-source rembg library before compositing the product into generated lifestyle scenes, producing results that look photographed rather than generated. Colour and variant visualisation pipeline: given a hero product image, generate the same product in all colour variants by applying colour conditioning prompts, eliminating the need to photograph physical samples for each SKU. Scene variation generation: one product image multiplied across five or ten lifestyle contexts (kitchen countertop, outdoor table, gift unboxing) within the same batch job. PIM (Product Information Management) integration with Akeneo, Salsify, or Shopify product catalog: webhook triggered on new product creation kicks off the generation job automatically, stores output images to S3 with metadata linking back to the product SKU, and publishes to the catalogue via API. Generation cost at DALL-E 3 pricing runs $0.04--$0.12 per image; Replicate Stable Diffusion runs $0.0023--$0.0046 per image at 512×512, an order of magnitude lower for high-volume catalogues.

Batch generation of ad creative, social media visuals, email headers, and blog illustrations at the volume content teams need without proportionally scaling design headcount. Generation architecture for batch creative: a prompt template engine that combines fixed brand style descriptors (colour palette, lighting style, composition rules, negative prompts) with variable campaign elements (product name, offer text, seasonal theme), producing 50–200 creative variations from a single campaign brief. Format adaptation: one generation job produces the same creative in 1:1 (Instagram square), 4:5 (Instagram portrait), 16:9 (YouTube banner), and 1.91:1 (Open Graph) ratios using Cloudinary's image transformation API after generation, avoiding separate generation jobs for each format. Brand consistency enforced through a LoRA fine-tuned checkpoint trained on your existing approved creative assets: the model learns your illustration style, colour temperature, typography placement zones, and composition preferences, producing output that passes brand review without iteration. DAM integration with Bynder, Brandfolder, or Canto: generated images auto-tagged with campaign name, product category, format ratio, and generation date via the DAM's metadata API, making creative searchable and reusable immediately. CMS publishing integration: approved images pushed directly to Contentful, Sanity, or WordPress media libraries via API, eliminating the upload step from creative workflow. Time savings benchmarked at 3–5 hours per campaign for high-volume teams producing 20+ creative sets per month.

AI image generation embedded as a feature inside your product, design tools that generate background assets, personalisation flows that create custom imagery per user, content creation assistants with integrated generation, and avatar or profile image generation. The generation step is the smallest part; the surrounding infrastructure is what makes it production-safe at scale. Job queue architecture using BullMQ (Node.js) or Celery (Python) with Redis: generation requests queued asynchronously so a slow generation job (8–15 seconds for Stable Diffusion) does not block the user's session; the UI polls a job status endpoint or receives a WebSocket push notification when the image is ready. Pre-signed S3 URLs for image delivery: generated images stored in private S3, accessed via time-limited pre-signed URLs (15-minute expiry) so images are never publicly crawlable. Rate limiting per plan tier using a Redis token bucket: free tier users get 5 generations per day, paid tiers get configurable monthly quotas with overage detection and user-facing usage indicators. Usage attribution: every generation job stored with the authenticated user ID, timestamp, prompt, and output URL, enabling per-user generation history, regeneration from prior prompts, and billing data for metered plans. Content moderation at generation time (see Content moderation card below) prevents policy violations from reaching users. Generation UI components: prompt input with character limit and example prompt suggestions, generation progress indicator showing estimated wait time based on current queue depth, and an image gallery view showing the user's generation history with download and share controls.

Training Stable Diffusion XL or Flux models on your existing approved brand assets so the model generates in your visual language by default, without lengthy style descriptors on every prompt. Fine-tuning approach selected by use case: LoRA (Low-Rank Adaptation) using the Hugging Face PEFT library for style training on 20–30 brand images; LoRA produces a compact adapter file (10–50 MB) that loads on top of the base model, reducing training cost compared to full fine-tune. DreamBooth fine-tuning for product or character consistency: trains the model to recognise a specific product (your shoe, your mascot, your device) by identifier token (sks product) so every generation featuring that token produces the correct product with accurate geometry and branding detail, not a plausible approximation. Training infrastructure: NVIDIA A100 80GB or RTX 4090 via cloud GPU (Lambda Labs, Vast.ai, or AWS p3 instances); a LoRA style fine-tune completes in 2–4 hours at a cost of $20--$80; a DreamBooth product fine-tune completes in 1–3 hours. Training pipeline automated with Kohya-ss trainer configured for the target resolution (1024×1024 for SDXL), noise offset, and min-SNR loss weighting for stable convergence. Validation after training: the fine-tuned LoRA evaluated against a test prompt set covering your typical generation use cases, with output rated by your brand team before production deployment. The delivered artifact: a model checkpoint and LoRA adapter hosted on your inference infrastructure or a private Hugging Face repository, with a prompt reference guide showing the trigger tokens and negative prompts that reliably produce on-brand output.

Production image generation requires content moderation at two independent layers, input and output, because a permissive generation system is a liability for any product with end-user access. Input layer: every user prompt screened by the OpenAI Moderation API (available at no additional cost, returns category scores for hate, violence, sexual content, self-harm, and harassment) before the generation job is queued. Prompts scoring above a configurable threshold per category are rejected with a policy violation message before GPU time is consumed. For self-hosted models without API moderation, a fine-tuned text classifier (DistilBERT or a rule-based blocklist augmented with semantic similarity) handles the same gate. Output layer: generated images screened by a NSFW image classifier before delivery. Open-source options: Falconsai/NSFW-image-detection (Hugging Face) for general NSFW classification; NudeNet for nudity-specific detection; Google Cloud Vision SafeSearch API for a managed option with categories including adult, violence, racy, and medical content. Confidence thresholds set per use case: a children's platform requires higher sensitivity than a medical imaging tool; thresholds documented and reviewed quarterly. Human review queue for images flagged at medium confidence (between the auto-block and auto-pass thresholds): a lightweight admin interface showing the flagged image, the prompt that generated it, the classifier confidence score, and one-click approve or reject, resolving edge cases without false-positive blocking at scale. Audit log: every moderation decision (prompt screened, image screened, outcome, score, reviewer if human) stored with timestamp and user ID, the record regulators and platform policies require. DALL-E 3's built-in moderation handled automatically; the moderation infrastructure above is designed for self-hosted Stable Diffusion and Flux deployments.

End-to-end image pipeline connecting the generation step to your existing storage, delivery, and content management infrastructure, because a generation API call that outputs a raw image URL is not a production asset until it is stored durably, delivered efficiently, and discoverable in your systems. Storage layer: generated images uploaded to S3 (private bucket with IAM policy restricting public access) or to Cloudinary via Cloudinary Upload API, which auto-generates responsive variants (thumbnail, web, full-resolution) and applies automatic background removal or quality optimisation transforms at upload time. Metadata tagging at upload: image stored with structured metadata (prompt used, generation model, campaign tag, product SKU, user ID, generation timestamp) enabling search and filtering in DAM and CMS systems. CDN delivery: CloudFront distribution in front of S3 for low-latency global delivery with cache-control headers set for generated images (immutable content served with long TTL; prompt variations served with shorter TTL). ImageKit as an alternative to Cloudinary for teams requiring real-time URL-based image transformations (resize, format conversion, quality adjustment) without pre-generating variants at upload time. CMS integration: Contentful, Sanity, or WordPress media library updated via API when a generation job completes, approved images appear in the content team's media browser without a manual upload step. Webhook architecture: a webhook fired on generation completion notifies downstream systems (CMS, DAM, email platform) with the image URL, metadata, and job reference ID, enabling async workflows where generation is triggered by one system and consumed by another without polling. End-to-end latency from generation trigger to CDN-available URL: 15–30 seconds for DALL-E 3 (API-managed), 10–20 seconds for a well-provisioned Stable Diffusion inference endpoint on A100 GPU.