AI image upscaler SaaS built on open-source Stable Diffusion

Open-source image models, a paywall from day one, and a two-person team.

In plain English

Magnific takes a low-detail picture and rebuilds it so the result looks sharper and more detailed. People pay from the first day because the pictures are meant to look better than what free tools make. The picture work runs on rented computers using open-source image software, and sign-up, billing, and usage credits happen with no person in that step. The people who run Magnific still tune the settings, post before-and-after examples, and handle help and refunds. The AI does almost all of the daily work.

Automation90%

Startup costNot disclosed

Time to first $Launch day. They got serious in October 2023, launched Nov 28, 2023 with a paywall and no free tier, and were profitable from the start.

Monthly range*$850,000 Annual ÷ 12

$10.2M revenue in 2024 (2 people, ~90% gross margin after GPU costs), posted by Javi López on X, Apr 28, 2026.

DifficultyHard

SourceJavi López on X: Magnific AI in numbers (Apr 28, 2026) ↗

Operator: Javi López and Emilio Nicolás

Verified

Revenue independently checked: documents reviewed by a publication, or a public dashboard.

Not disclosed. A USD figure was never given. On the itnig podcast (May 2024) the founders said they formed the company with €3,000 of capital and spent under €10,000 in total, with zero outside investment.

The tool stack

ComfyUI ↗

Free (open source)

Build and test the upscaling pipeline locally on your own GPU

Stable Diffusion (open weights) ↗

Free open weights (check each model's license for commercial use)

Core image model; no closed API

Replicate ↗

Pay per second, e.g. Nvidia L40S $3.51/hr, A100 80GB $5.04/hr

Rented GPUs to serve the pipeline to customers

Stripe ↗

2.9% + 30¢ per successful domestic card charge

Subscriptions

Prices are the vendors' published list prices as researched in Oct 2026 and change often — always check the vendor's pricing page.

Step-by-step setup

Combine open Stable Diffusion models and community ComfyUI workflows into one opinionated pipeline that does a single job well
Test it locally until the results beat free tools
Deploy the pipeline on rented GPUs (Magnific used Replicate) behind a simple web app
Launch with a paid-only plan and token credits, with no free tier
Promote it with striking before/after examples on X

What's automated vs. the human 10%

The machine does

Every upscale runs through the GPU pipeline with no human in the loop
Signup, billing and credit metering

You still do

Tuning the pipeline and its settings
Launch content and community posts
Support and refunds

Overview

Magnific is an AI upscaler. It takes a low-resolution image and “re-imagines” it with sharper detail. Javi López and Emilio Nicolás built it on open models, not a closed API. On the itnig podcast they explained that it combines Stable Diffusion with community ComfyUI/A1111 workflows and runs on rented Replicate GPUs (“No llamamos a una API”, meaning “we don’t call an API”). El País reported the launch date of Nov 28, 2023, and 400,000 users in a month and a half. In May 2024, Freepik acquired it.

Why it’s “local”: you can build and run the whole pipeline on your own GPU with ComfyUI. Magnific rented cloud GPUs only to serve customers at scale.

How the money comes in

Customers pay a subscription for credits. There has never been a free tier. On itnig the founders said most users are on the $39/mo plan (2,500 tokens, about 300 upscales) and the average spend was $45 or less.

Realistic income

In an X post on Apr 28, 2026, Javi López reported revenue of $10.2M in 2024 with “just 2 people” and “a 90% gross margin after GPU costs”. He reported $34.42M for 2025, but that figure includes both websites after the acquisition. The revenue is founder-reported, while the acquisition was covered by independent press. This is one exceptional result from a founder with a big audience. It doesn’t predict yours.

Risks

Fast-moving models. Newer closed and open models can wipe out your quality edge in months.
Licenses. Check each open model’s commercial-use terms before you charge.
GPU costs. If you set credit prices wrong, heavy users can wipe out your margin.
Distribution. Paid marketing didn’t work for them: on itnig they said a €1,000 TikTok influencer post flopped. Most of their growth was organic.

How to start this week

Pick one narrow image job (for example product photo cleanup or old-photo restoration).
Build a ComfyUI workflow on your own GPU and compare its output against free tools.
Wrap it in a simple web app with Stripe and a paid-only credit pack.
Post before/after threads daily and ask for paid signups only.

*Income figures are illustrative ranges based on public pricing and typical outcomes reported by practitioners, not guarantees. Most new online businesses earn little or nothing at first. Not financial, legal or tax advice.

Featured in: AI Business Ideas for Graphic Designers: Eight Documented Stacks · AI Photo Editing Business Ideas: What Actually Makes Money · How to Make Money with Replicate: Eight Documented Paths · How to Make Money with Stable Diffusion: Eight Local Paths · Offline and On-Device AI Business Ideas (46 Real Examples)