How to Use FLUX for AI Image Generation and Editing
A practical guide to understanding that FLUX is models, not a finished product — and choosing the right access route and version before you pay for or download anything.
Beginner7 min read
What you'll be able to do by the end
- Understand FLUX isn't a finished product like Midjourney — it's a family of models you reach via API or download and run yourself, giving freedom and lower cost in exchange for more setup responsibility
- Distinguish versions: klein is open and free commercially (Apache 2.0), dev has open weights but requires a paid licence for commercial use, and pro/max are API-only
- Know it's among the best at rendering English text inside images, but other scripts are still weak with broken letters — generate clean and add text later in a design tool
- Do the real cost math: klein is nearly $0.014 per image via API, pro about $0.03, max about $0.07 — and price rises with resolution
- If your GPU lacks large VRAM, the simplest route is an intermediary platform where you pay per image without any local setup
Before you start
- A clear idea of the image type you need (product shot, banner with text, editing an existing photo) before choosing a route and version
- For local runs: a GPU with large video memory (16GB or more for the larger versions) and sufficient storage
FLUX is a family of image-generation and editing models from a German company called Black Forest Labs, founded by the same team behind the original Stable Diffusion. The thing to absorb from minute one: this isn’t a finished product like Midjourney that you open and type into — these are models you reach via API or download as weights and run yourself.
That difference buys you more freedom and lower cost for intermittent use, but hands you setup work and the obligation to read the licence before any commercial deployment — which is exactly where most people trying FLUX stumble.
This tutorial walks you from the route decision (intermediary platform, API, or local run) through picking the right version among klein, dev, pro, and max, to writing prompts that produce strong output — with a blunt warning about script limits inside images.
The steps
The steps below run from understanding FLUX’s nature as models, through choosing route and version, to prompt writing and handling in-image text correctly.
Steps
Step 1: Understand first that FLUX is models, not a ready tool
The fundamental difference between FLUX and tools like Midjourney: it isn't a product you open and type into. FLUX is a family of models from a German company called Black Forest Labs, founded by the same team behind the original Stable Diffusion. To use them, you connect through an API, or download the weights and run them on your own machine. That gives greater freedom and lower cost for intermittent use, but puts setup responsibility on you that no finished product would.
Note: Want to open something and use it immediately with zero setup? FLUX isn't the right choice — look for a finished product instead of raw models.
Step 2: Choose your access route: intermediary platform, direct BFL API, or local run
Three ways reach FLUX. First, an intermediary platform where you pay per image with no technical setup — fastest path if you lack technical experience or a strong GPU. Second, the direct API from Black Forest Labs if you're building a product or integrating into a project. Third, local running via tools like ComfyUI or Diffusers, if you have a big-VRAM GPU and want full independence from any provider. Choose by how much control you need versus how much setup time you're willing to spend.
Note: Unsure? Start with an intermediary platform — easiest way to sample quality before committing to an API or local setup.
Step 3: Pick the right version and mind the licensing trap
This is the point most people fumble. klein is fully open under Apache 2.0 — download, run, use commercially at no cost; excellent for experimentation and high volume. dev has open weights but commercial use requires a paid monthly licence ($999/month for commercial operation on your own hardware up to 100,000 images monthly) — suited to running a strong model on your own infrastructure. pro and max are reachable only via API under the provider's terms, and they're the default choice for most client work and final deliverables. Before any commercial project, read the licence of the version you're using yourself — the difference is not a minor detail.
Note: Use klein for experimentation and volume, pro for most daily work, and max only for final deliverables needing top quality.
Step 4: Write prompts in English and describe lighting and lens clearly
Write prompts in English because that's what the model understands best, and describe lighting (natural, golden hour, studio), lens type, and visual style explicitly — this noticeably lifts output quality. FLUX ranks among the strongest models at rendering English text inside images with high accuracy, making it excellent for banners, covers, and posters containing words. Need to edit an existing image? Upload it and describe the change in plain language — 'replace the background with a modern office' — and the model executes it without another tool.
Note: Realism is one of its strongest suits — skin, lighting, and reflections come closer to natural than almost any competitor; lean into it rather than trying to correct after generation.
Step 5: Don't render non-Latin script in-image — generate clean and add it later
The fact to know from the start: FLUX is very strong at English text in images, but other scripts remain weak with letters breaking or intersecting wrongly. The practical solution: generate the image with no text at all, then add non-Latin text with a design tool after generation. Finally, generate your final version at the highest resolution you'll actually use — price rises with resolution, so don't generate above what you'll need.
Note: Non-English prompts are also noticeably weaker than English ones — write scene descriptions in English even when the final overlay text won't be.
Common mistakes — and how to avoid them
MistakeExpecting FLUX to be a ready app you open and use immediately like Midjourney.
Do this insteadUnderstand it's models reached via API or self-run — if you want a no-setup experience, an intermediary platform is the nearest solution, not the raw model.
MistakeUsing dev commercially without reading the licence terms, assuming it's as open as klein.
Do this insteadOnly klein is open and commercially free under Apache 2.0. dev needs a paid monthly licence for commercial use — read every version's licence before any commercial project.
MistakeTrying to render non-Latin script directly in-image and expecting clean output.
Do this insteadGenerate the image without text, then add it with a design tool afterwards — such scripts are still weak and letters break.
MistakeBuying a dev licence or setting up local hardware just for casual experimentation.
Do this insteadUse an intermediary platform or klein for experiments and low volume; commit to local setups or paid licences only when scale genuinely justifies it.
❓ Frequently asked questions
What's the difference between FLUX and Midjourney?
Midjourney is a finished product you open and use directly. FLUX is models you reach via API or run yourself — more freedom and lower cost for intermittent use, but without Midjourney's large community and inspiration gallery, and with more setup on your side.
Which version should I choose: klein, dev, or pro?
klein for experimentation and high volume since it's open and commercially free under Apache 2.0. pro is the default for most work via API. dev if you want to run a strong model on your own infrastructure, though commercial use requires a paid monthly licence. max for final deliverables demanding top quality.
Can I write non-English text inside the image?
Not recommended. FLUX renders English in-image text accurately, but other scripts stay weak with letters breaking. Solution: generate the image without text and add it later with a design tool.
Do I need a powerful GPU to use it?
Only if you plan to run it locally — that means a GPU with 16GB+ of video memory for the larger versions plus basic familiarity with tools like ComfyUI or Diffusers. Without that, a per-image intermediary platform removes any need for local setup.