Verified 18 August 2026. Every link below was opened and checked. Every licence was read from the source file, not the badge on the page.
You don't need to be technical to use any of this. Four of these have an official free website where you can make images in about 30 seconds, with nothing to install.
Start there. Only install something once you know you like the output.
Try everything first, install nothing
Official free demos. Open the link, type what you want, get an image. Works on a phone.
| What you want | Open this |
|---|---|
| Photorealistic images, fast | Z-Image Turbo |
| Images with readable text in them | Qwen-Image-2512 |
| Edit a photo by describing the change | Qwen-Image-Edit-2511 |
| Both, in one model | FLUX.2 klein 4B |
These run on free shared servers. Logged out you get a couple of minutes of generation time a day; a free account raises that. Enough to decide whether you want it.
On a Mac and want it running locally? Get Draw Things from the Mac App Store. It's free, runs offline on your own machine, and needs no terminal, no Python, nothing. It supports Z-Image Turbo, Qwen-Image 2512 and the FLUX.2 klein family. For a Mac owner this is by far the easiest route.
On Windows? Install ComfyUI Desktop. See the last section for why it's the right pick for a beginner.
Read this before you download any model
The most expensive mistake in AI images right now is downloading the bigger version of a model and assuming bigger is better.
FLUX.2 klein comes in a 4B and a 9B. Only the 4B is free for commercial use.
| Version | Can you sell the images? |
|---|---|
| FLUX.2 klein 4B | Yes |
| FLUX.2 klein 4B Base | Yes |
| FLUX.2 klein 9B | No |
| FLUX.2 klein 9B KV | No |
| FLUX.2 klein 9B Base | No |
| FLUX.2 dev | No |
Same product name, opposite licence. If you download the 9B because it sounds better, you cannot legally sell anything you make with it. Download the 4B.
1. Z-Image Turbo — the one to start with
What it does. Makes photorealistic pictures from a sentence, very fast.
Try it free: huggingface.co/spaces/Tongyi-MAI/Z-Image-Turbo
To run it yourself. In ComfyUI Desktop, open the Templates sidebar and pick "Z-Image-Turbo: Text to Image". On a Mac, use Draw Things instead.
What computer you need. A graphics card with 16GB comfortably runs it. This is the most Mac-friendly model on the list — it has proper Apple Silicon support built in.
What usually goes wrong. On older graphics cards (RTX 20-series and similar), you get pure black images. It's a known bug with no fix yet. If that happens, use the web demo instead.
Can I sell what I make? Yes. Apache 2.0, no conditions attached.
One thing to know. Development has been quiet since February 2026, and the Edit version that was announced never shipped. The model works fine, it just isn't moving.
2. FLUX.2 klein 4B — fastest, and the safest licence
What it does. Makes images and edits existing ones, in about a second on a gaming PC.
Try it free: huggingface.co/spaces/black-forest-labs/FLUX.2-klein-4B
To run it yourself. ComfyUI Desktop templates: "Flux.2 [Klein] 4B: Text to Image" and "Flux.2 [Klein] 4B: Image Edit". Easiest option of all is InvokeAI, which has a "FLUX.2 Klein starter pack" that downloads everything as one bundle. Mac users: Draw Things supports it.
What computer you need. The version ComfyUI ships needs about 8.4GB of graphics memory. The full-quality version needs about 13GB. You'll see "8GB" quoted a lot — that's the compressed version, which is the one you want anyway. Works on M-series Macs, roughly 17 seconds per image.
How long it takes. Around a second on a recent NVIDIA card.
What usually goes wrong. A "node not found" error when you open the template. Update ComfyUI and it goes away.
Can I sell what I make? Yes, but only with the 4B. See the table above. This is the single most important thing in this guide.
3. Qwen-Image-2512 — the one that can spell
What it does. Makes images with real, readable, correctly spelled text inside them. Posters, ads, signage, social graphics. Most image AI still produces gibberish letters; this one doesn't.
Try it free: huggingface.co/spaces/Qwen/Qwen-Image-2512
To run it yourself. ComfyUI Desktop template "Qwen-Image-2512". Be warned before you start.
What computer you need. This is the heavy one. You want a 24GB NVIDIA graphics card, and the downloads come to around 30GB. There's a compressed version that fits in about 13GB if you're tight. On a Mac it technically runs but it's painful — a top-end M4 Max with 128GB of memory takes about 3 minutes for a single small image. Mac users should use the free web demo.
How long it takes. An hour or more to set up including downloads. Then roughly 70 to 90 seconds per image on a good card.
What usually goes wrong. The template offers a fast "4-step Lightning" option that produces visible grid artefacts. It's a known unresolved bug. Use the 50-step version instead.
Can I sell what I make? Yes. Apache 2.0, no conditions.
One thing to know. You may see "Qwen-Image 2.0" mentioned. It has no downloadable version and appears to be API-only. 2512 is the newest one you can actually run.
4. Qwen-Image-Edit-2511 — change a photo by describing it
What it does. Edit an existing image in plain English. "Make the sofa leather." "Remove the person on the left." "Change the background to a beach." It takes up to three reference images.
Try it free: huggingface.co/spaces/Qwen/Qwen-Image-Edit-2511
To run it yourself. ComfyUI Desktop template "Qwen-Image-Edit-2511". Needs ComfyUI version 0.6.0 or newer.
What computer you need. Same heavy requirements as above: 24GB NVIDIA card, plus 64GB of system memory if you want to use its LoRA features. Mac users: skip this one. It fails outright on some Macs with an error the older version didn't have. Use the web demo, or use the previous version (2509) which does work on Mac.
What usually goes wrong. Running out of memory when LoRAs are enabled, which is awkward since that's the headline feature. Turn them off if it crashes.
Can I sell what I make? Yes. Apache 2.0.
5. HiDream-O1 — generation and editing in one
What it does. One model that both creates and edits images, including dropping a specific person or product into a new scene.
Try it free: huggingface.co/spaces/HiDream-ai/HiDream-O1-Image — but be patient. The official demo runs on slow shared hardware and is often queued or unresponsive.
To run it yourself. ComfyUI Desktop templates "HiDream O1 Full" and "HiDream O1 Dev". Generating and editing use the same template — there's a "Switch to Image Edit" toggle, and it takes up to 10 reference images.
What computer you need. Around 12 to 16GB of graphics memory using the compressed version, or 17 to 20GB for full quality. Mac: not supported. There's an open bug where it produces completely white images on Apple hardware. Mac users should skip it.
What usually goes wrong. Visible seams and grid lines across the image. This is common enough that ComfyUI ships a patch node turned on by default to reduce it. It's also noticeably weaker on photorealistic people, who come out plastic-looking. It's much stronger on stylised work, text and reference-based editing.
Can I sell what I make? Yes. MIT, the most permissive licence on this list.
One thing to know. Development has been quiet since June 2026 and issues are going unanswered. It works, it's just coasting.
6. ComfyUI — the program that runs all of them
What it does. The free program you install on your computer to run every model above.
Get it here: comfy.org/download — choose ComfyUI Desktop, the proper installer.
Why Desktop specifically, and not the other versions: with Desktop, the Template Library downloads the model files for you automatically. You click a template, click Download, and it handles the rest. With the other versions you have to download multi-gigabyte files by hand and put them in exactly the right folders. For a beginner that difference is everything.
To find the templates: click the Templates icon in the sidebar, or go to Workflow → Browse Workflow Templates.
What computer you need. Windows 10 or newer, or macOS 13+ on an M-series Mac (older Intel Macs aren't supported). 8GB RAM minimum, 16GB recommended, about 5GB of disk for the program itself before any models.
Can I use it in my business? Yes, free and unrestricted. The licensing is technically split — the core program is GPL-3.0 and the desktop app is dual-licensed — but that only matters if you want to redistribute the app itself or run it as a hosted service for other people. Making and selling images triggers none of it.
If the node interface intimidates you. SwarmUI is a free MIT-licensed alternative with a normal form-based interface instead of a node graph, and it supports all five models above. The catch is installing it: on Windows it's a file you download and run, which is fine, but on Mac it needs the terminal. Consider it a step-two upgrade for Windows users, not a starting point.
Where to start
Never done this before? Open the four web demos at the top. Nothing to install, works on your phone, and you'll know within five minutes whether any of this is for you.
On a Mac? Draw Things from the App Store. Free, local, no terminal. Stick to Z-Image Turbo and FLUX.2 klein 4B; the Qwen and HiDream models don't work well on Apple hardware.
On a Windows gaming PC? ComfyUI Desktop, then start with the Z-Image Turbo template.
Need text inside your images? Qwen-Image-2512, but use the web demo unless you have a 24GB card.
Only care about being able to sell the output? FLUX.2 klein 4B or HiDream-O1. Both permissive, no conditions. Just never the 9B.
The part most lists get wrong
All six tools here are Apache 2.0, MIT or GPL. You can use every one commercially and sell the images.
That is not true of the tools you'll see recommended most often:
- FLUX.1 dev, FLUX.2 dev, and FLUX.2 klein 9B — all non-commercial. The 4B klein is the only one in the family you can sell from.
- Ideogram 4 — the biggest trap of 2026. Its inference code is Apache 2.0, which is what every roundup quotes, but the model itself is non-commercial. Best-in-class text rendering that you legally cannot use for business.
- Stable Diffusion 3.5 — free only while your company earns under $1M a year, from any source. Cross that and your licence terminates. (Older SDXL and SD1.5 have no such cap and stay genuinely free.)
- Google's Gemini API — every single image model on its pricing page reads "Free Tier: Not available." Free image generation only survives in the consumer Gemini app, with a visible watermark on by default.
- Leonardo, Ideogram's free tier, Craiyon, Bing Image Creator — all have real restrictions on commercial use, credits, or downloads that the listicles skip.
The pattern is always the same: the code and the model are licensed separately. Sites quote the code licence because that's the one with the friendly badge on the repo page.
Check the model's licence, not the repo badge.
Same homework, done for the other two media: 6 Free AI Video Tools and 6 Free AI Audio Tools.