Everyone compares the pictures. Almost nobody compares the terms
Pick any two image generators and the sample galleries will look roughly as good as each other. That is not where they differ any more. They differ on three things that decide whether you can actually finish a job: what one image costs you, how much control you get when the first result is wrong, and what you are permitted to do with the file afterwards.
The third one is the one that ends careers-in-miniature — a designer who delivers 200 assets built on weights they were not licensed to use commercially.
If you have not read how these models actually work, read that first. This course assumes it and never repeats it.
The hosted tools, and what each is actually for
Midjourney still produces the most immediately attractive frame with the least effort. The cost of that is an opinion baked into every image — a house look you have to fight. It is subscription-only, with no free tier, and on the entry plan your generations are visible in the public gallery; private generation sits on the higher tiers. If you work under NDA, check that before your first render, not after.
Adobe Firefly is the conservative choice. It is trained on Adobe Stock and licensed or public-domain material, and Adobe offers indemnity against training-data claims on its business plans. It is less exciting and more defensible. It has a free monthly credit allowance, and Generative Fill inside Photoshop is the same engine.
Google's Gemini and Imagen line is strong at following a long instruction and at editing an image you supply. Google stamps its generated images with SynthID, an invisible watermark. There is meaningful free use through the Gemini app and AI Studio.
Ideogram exists because text inside images was broken for years. It renders legible words far better than the general-purpose models. It is still not typesetting — more on that in the finishing lesson.
Recraft aims at design output rather than pictures: it will give you actual vector SVG, which matters if the thing you are making has to scale to a billboard. See vector and print for why that distinction is not cosmetic.
Higgsfield is a different species — a recipe-driven platform rather than a raw model. The next lesson is about that whole category.
Running the weights yourself
The open-weights line runs Stable Diffusion → SDXL → SD 3.x, alongside FLUX from Black Forest Labs and Qwen-Image from Alibaba. Running them yourself costs nothing per image, gives you every control that exists, and works offline.
The requirement is a GPU. SDXL is comfortable on 8 GB of VRAM. FLUX's larger variants want much more, though quantised versions run on 8-12 GB slowly. If you do not own a card: Google Colab's free tier, Kaggle's free weekly GPU hours, and Hugging Face Spaces all run these models at no cost. Running models yourself and Hugging Face cover the mechanics.
Free front ends, all genuinely free: ComfyUI (a node graph, ugly, total control), Fooocus (Midjourney-like defaults, almost no settings), InvokeAI, and Krita with the AI Diffusion plugin, which is the best free inpainting experience there is.
The licence trap, stated plainly
A model's weights carry a licence, and it is a separate thing from what you may do with the pictures.
- FLUX.1 [schnell] is Apache 2.0. Do what you like.
- FLUX.1 [dev] is a non-commercial licence. Thousands of people have downloaded it, run it locally, and billed clients for the output. That is a breach of the model licence, regardless of any question about who owns the image.
- Qwen-Image was released Apache 2.0.
- Stability's recent releases use a community licence with a revenue threshold: free below it, paid above.
- Hosted tools replace all this with their terms of service, and free tiers frequently restrict commercial use even when paid tiers do not.
Those were the positions in 2025 and they move. Check the model card on the day you start the job. Lesson ten deals with output rights, which is a different question again.
Choosing, in about thirty seconds
- A single striking image, no budget for fiddling — Midjourney or the Gemini app.
- Corporate work where somebody will ask about training data — Firefly, or Getty's and Shutterstock's own generators.
- Words inside the image — Ideogram, then set real type over it anyway.
- Hundreds of near-identical variants — a preset platform.
- Precise control, a repeating character, or no money at all — open weights, locally or on free GPU hours.
Today: open the model card or terms page of whatever you are currently using and find the sentence about commercial use. Most people have never read it.
Before you move on