A self-hosted AI image generator runs an open-weight model on your own computer, with no subscription, no per-image fees and no images leaving your machine. In late 2026 the strongest open models are Black Forest Labs’ FLUX.2 family and Ideogram 4.0, alongside the older Stable Diffusion ecosystem. You run them in an app such as ComfyUI. The main requirement is a graphics card: about 8 GB of VRAM for the smallest FLUX.2 model, and 16 to 24 GB for the larger ones.
This guide covers which model fits your hardware, which app to use, licensing traps and a basic setup.
Why run an image generator offline
- Privacy. Product photos, client work and personal images stay on your machine.
- No limits. Generate as many images as your hardware allows, without credits or queues.
- Control. Choose seeds, samplers, steps and resolution, train on your own style and build repeatable workflows.
- Availability. Nobody can change the model, its filters or its price under you.
The trade-offs are setup time, the cost of a good GPU and your own responsibility for how you use the output.
Open image models compared
| Model | Developer | Sizes | Minimum VRAM (quantized) | License for commercial use |
|---|---|---|---|---|
| FLUX.2 Klein 4B | Black Forest Labs | 4B | About 8 GB | Apache 2.0, commercial use allowed |
| FLUX.2 Klein 9B | Black Forest Labs | 9B | About 16 GB | Apache 2.0, commercial use allowed |
| FLUX.2 dev | Black Forest Labs | 32B | About 18 to 24 GB at 4-bit | FLUX Non-Commercial License; commercial use needs a separate license or the paid API |
| Ideogram 4.0 | Ideogram | Not stated | Check model card | Commercial license with tiers; see Ideogram’s licensing page |
| Stable Diffusion family | Stability AI and community | Various | 6 to 12 GB for common versions | Varies by version; read each license |
Sources: Botmonster FLUX.2 guide, updated May 2026, ComfyUI on FLUX.2, Ideogram 4.0. VRAM figures are community guidance for quantized builds; full-precision FLUX.2 dev needs about 80 GB. Verify the license on each model’s official page before commercial use.
The license column matters more than people expect. FLUX.2 dev, the most capable open FLUX model, is not licensed for commercial output without a separate agreement. If you sell what you make, the Apache-licensed Klein models are the simpler choice.
Which model fits your GPU
| Your GPU VRAM | Good choice | Notes |
|---|---|---|
| 8 GB | FLUX.2 Klein 4B | Fast; Klein 4B can generate in about 4 steps |
| 12 GB (RTX 3060, 4070) | Klein 4B, or a low-bit build of a larger model | Expect slower generation |
| 16 GB (RTX 4080) | Klein 9B, or FLUX.2 dev at a small quantization | Good balance |
| 24 GB (RTX 3090, 4090) | FLUX.2 dev at 4-bit or FP8 mixed | Best open-model quality on one consumer card |
| 32 GB (RTX 5090) | FLUX.2 dev with more headroom | Faster, larger images |
| Apple Silicon Mac | Smaller models through Mac-native apps | Slower than a comparable NVIDIA GPU |
Guidance from the Botmonster guide. Plan for about 32 GB of system RAM (64 GB if the model spills over from the GPU) and around 60 GB of free disk space for FLUX.2 dev and its text encoder.
No suitable GPU? Rent one by the hour from a cloud GPU provider and run the same software there. It’s cheaper than buying for occasional use.
Apps to run open image models
ComfyUI
ComfyUI is the most widely supported app for open image models. You build workflows by connecting nodes (load model, encode prompt, sample, save), which looks complex at first but makes results repeatable and shareable. It had day-one support for FLUX.2, and most new models ship with ComfyUI templates. Install it directly or through a manager such as Stability Matrix.
Forge and other web UIs
Forge, a fork of the long-popular Automatic1111 interface, gives you a simpler form-based UI: prompt box, settings and a generate button. It’s easier for beginners and supports many Stable Diffusion and FLUX models.
Simplified apps
Apps such as Fooocus and InvokeAI hide most settings and focus on prompting and editing, and Mac-native apps exist for Apple Silicon. Check that the app supports the model you want before you install it.
A basic FLUX.2 setup in ComfyUI
This outline follows the community guide linked above; check the model’s Hugging Face page for current file names.
- Install an up-to-date NVIDIA driver, Python 3.10 or 3.11 and Git.
- Install ComfyUI, directly or through Stability Matrix.
- Download the model files from the official FLUX.2 repository on Hugging Face: the diffusion model, the text encoder and the VAE, and put each in its ComfyUI models folder.
- Open ComfyUI, load the FLUX.2 template from the templates menu and select the files in the loader nodes.
- Start with the recommended settings for that model (FLUX.2 dev: about 20 to 30 steps; Klein 4B: about 4 steps), type a prompt and run.
For writing better prompts, see how to write text-to-image prompts.
Features worth learning next
- Reference images. FLUX.2 accepts up to 10 reference images to keep a character, product or style consistent across outputs.
- LoRAs. Small add-on files that teach a model a style, character or product without retraining it.
- Inpainting. Regenerate only part of an image, such as a hand or a background.
- Upscaling. Generate at a moderate size, then upscale to save VRAM and time.
Responsible use
Running a model locally removes the platform’s filters, so the responsibility is yours. Don’t create images of real people without consent, follow the law on explicit and harmful content, and label AI images where platforms require it. Download models only from official pages or reputable hubs, since model files from unknown sources can carry malicious code in some formats; prefer the .safetensors format.
FAQ
Can I run an AI image generator offline?
Yes. After you download an open model such as FLUX.2 and an app such as ComfyUI, image generation runs entirely on your computer without an internet connection.
What is the best self-hosted AI image generator?
For quality on one consumer GPU, FLUX.2 dev in ComfyUI, if you don’t need commercial use. For commercial work, the Apache-licensed FLUX.2 Klein models, or Ideogram 4.0 under its commercial license.
How much VRAM do I need?
About 8 GB for FLUX.2 Klein 4B, 16 GB for Klein 9B and 18 to 24 GB for FLUX.2 dev in a quantized build.
Is self-hosting cheaper than Midjourney?
If you already own a capable GPU, yes: there are no per-image fees. Buying a GPU only for image generation costs more than several years of a basic subscription, so for light use a hosted tool is cheaper. See the best AI image generators for hosted options.
Conclusion
Self-hosting gives you private, unlimited image generation with FLUX.2, Ideogram 4.0 or Stable Diffusion models, as long as you have the GPU for it and check each model’s license. Start with FLUX.2 Klein on 8 to 16 GB or FLUX.2 dev on 24 GB, in ComfyUI. To run open language models as well, see open-source ChatGPT alternatives.