A $4/month n8n server as the AI backend for 100+ users across 20+ brands
Summary
Socialwiz’s entire AI backend — image generation, editing, resizing, video ads and brand extraction — runs behind a $4/month n8n server. It serves 100+ users across 20+ brands today, on infrastructure built for orchestrating AI calls rather than for raw compute.
Problem
A multi-tenant ad-creative product needs to call out to several different AI models — routed by quality tier, chained into image-to-image and video pipelines, scored by vision models — from one backend, without every new model or feature meaning a redeploy of hand-written glue code. Building and maintaining that orchestration layer as bespoke backend code would have been the slower, more expensive path for a small team shipping features monthly.
Architecture
The front end is a Webflow site with Memberstack handling accounts. It submits jobs to an n8n instance that does the actual orchestration: picking a model, calling it, handling the response, and writing results back. Airtable holds the underlying schema behind that pipeline. n8n was chosen over a custom backend for three reasons: workflows are faster to build in it than backend code is to write and deploy for every new feature, it’s built for exactly this kind of model-orchestration work, and it runs on a single $4/month server instead of a dedicated application stack.
The product has gone through three rebuilds. A 2023 prototype (DALL-E 2, Midjourney, EdenAI) proved AI ad creative was viable at all. A November 2024 MVP — the first paid version — added the full front end, an eight-entity Airtable schema, live editing and subscription payments, built in one month. The current version re-architected around gpt-image-1, adding custom ad sizes, async backend processing, an n8n agent layer for SEM and social copywriting, and an Ad Analyzer that combines Gemini’s video understanding with OpenAI vision and a judge agent to score creative.
2026 folded five more features into that same n8n backend: Edit Image (upload a photo, write a prompt, pick a quality tier); a rebuilt Ad Creative flow that generates from a user’s own assets and context instead of a fixed template; brand identity extraction, where Cloudflare Browser Rendering screenshots and crawls a submitted website and stores the extracted identity for reuse across every later generation; AI-assisted resize, producing ten output sizes from one generation without cropping or breaking the composition; and an ad animation pipeline that turns a static ad image into a video ad. A separate four-month R&D effort is working on turning product stills into UGC-style video ads.
The Hard Parts
Edit Image routes by quality tier: the tier a user picks maps to a different underlying model — gpt-image-1 or nano banana 2 — trading cost against quality without exposing that choice to the user. Getting that routing right, invisibly, behind one quality selector took more care than the feature looks like it should need.
The animation pipeline turns a static ad into video by feeding GPT-4o vision into a Seedance pipeline, but text melts under generation — any text baked into the source image doesn’t survive being regenerated as motion. The fix strips text out before generation and composites the original back onto the finished video afterward, so it stays pixel-perfect instead of being redrawn.
Results
100+ users across 20+ brands run their ad creative through Socialwiz today, all served by the same $4/month n8n server.
Stack
n8n orchestrates every AI call behind the product. gpt-image-1 and nano banana 2 handle image generation and editing; Gemini and OpenAI vision back the Ad Analyzer; GPT-4o vision and Seedance drive the animation pipeline. Cloudflare handles image and video storage and website crawling. Webflow and Memberstack run the front end and accounts, with Airtable behind the data model.