Stable Diffusion AI online

Yes. It is an open-source model, free to download and to run on your own computer. Every tool on this site is free as well: the online demos, inpainting, upscaling and the prompt generator need no account, and text-to-image on our own GPU only asks for a free account. No credits, no subscription.

No account is needed for the online demos, inpainting, image upscaling and the prompt generator: open the page and start. A free account, created with an e-mail address, unlocks text-to-image on our own GPU with SD 1.5 and SDXL 1.0, with full control over steps, CFG scale, seed and image size.

Open the free demo, type a description of the image you want, then click Generate. The image appears after a few seconds and you can download it. To improve the result, add details about style and lighting, write a negative prompt, or change the seed to explore variations. Nothing has to be installed.

Generate an image →

No. Stable Diffusion is developed by Stability AI, whose official website is stability.ai. stablediffusion.fr is an independent site, online since 2022, that offers free tools to use the model and other open models in the browser, with guides, prompt examples and artist references. It is not affiliated with Stability AI.

About this site →

A prompt is a text description that guides the AI. It can include the subject, artistic style, lighting, mood, and medium — for example: 'a medieval knight in a dark forest, oil painting, dramatic lighting, cinematic'. The more specific your prompt, the closer the result to your vision. Keep it under 350 characters for best results.

A negative prompt tells the AI what to leave out of the image.

Read the full guide →

A negative prompt tells the AI what NOT to show. Common examples: 'blurry, low quality, watermark, text, deformed hands'. Use these to remove typical AI mistakes.

Negative prompts boost image quality fast. Add weights — like '(blurry:1.4)' — to make them stronger.

Read the full guide →

The seed is a number that initializes the random process behind image generation. Every image starts from noise, and the seed determines what that starting noise looks like. Leave it at -1 to get a random seed — and a different image — every time.

Fix the seed to reproduce the exact same image. This is useful for experimenting: keep the seed constant while changing only the prompt or settings to compare variations precisely. It is also how you recreate a favourite result later.

Read the full guide →

LoRA stands for Low-Rank Adaptation. It is a small add-on file (typically 50–150 MB) that fine-tunes a base model for a specific style, character, or subject — without retraining the entire model.

Load a LoRA to generate images in a specific artistic style, reproduce a particular character, or match a reference concept. You can combine several LoRAs in one prompt using weights: <lora:my-style:0.8>. Thousands of free LoRAs are available on Civitai and Hugging Face.

Read the full guide →

Running it locally requires a recent NVIDIA GPU with at least 6 GB of VRAM. AMD GPUs work with ROCm drivers on Linux. Apple Silicon (M1/M2/M3) is also supported via the MPS backend. If you have no compatible GPU, our free online tool runs directly in your browser — no installation needed.

Read the full guide →

You need at least 4 GB of VRAM to run the model. 6 GB is the recommended minimum for SDXL models. With 8 GB or more, you can generate larger images and use more features.

If you have no GPU, you can still run it on CPU — but it is much slower. You can also use our free online demo with no GPU required.

Img2img lets you use an existing image as a starting point. You provide a source image and a prompt. The model then transforms the image based on your instructions.

Use it to change the style of a photo, fix a generated image, or refine a rough sketch. The denoising strength controls how much the AI changes the original — low keeps it close, high transforms it fully.

It depends on the model. The base models use the CreativeML Open RAIL-M license. This allows commercial use, but with some restrictions — you cannot use images to harm or deceive people.

Always check the license of the specific checkpoint you use. Some community models restrict commercial use. If you sell or publish AI-generated images, keep a record of the model and its license.

ControlNet is an extension for the model that gives you precise control over image composition. You provide a reference image — a pose, an edge map, or a depth map — and the AI follows that structure.

Use ControlNet to keep a person's exact pose, copy a scene layout, or trace a sketch precisely. It works alongside your prompt, so you control both the content and the structure of the image.

A checkpoint is a saved AI model file. It contains everything the model learned during training. You load a checkpoint to set the style and subject of your images.

Different checkpoints give very different results. Some are trained for realistic photos, others for anime or illustration. You can download free checkpoints from Hugging Face or Civitai.

ComfyUI is a free, node-based interface for the model. You build image workflows by connecting blocks called nodes. Each node does one step — loading a model, adding a prompt, or saving the image.

ComfyUI gives you full control over each step. It is faster and lighter than Automatic1111. Power users prefer it for complex workflows. You can share workflows as JSON files.

Flux is a family of text-to-image models by Black Forest Labs, founded by former Stability AI researchers. The newest open-weights release is FLUX 3 Image Klein, published on 15 October 2026: 5.6 billion parameters, native 4K generation, editing from up to ten reference images, and an int8 build that runs on an 8 GB graphics card.

FLUX 3 Image Klein is released under the FLUX Kommunity License: free commercial use for individuals and companies under 5 million dollars in annual revenue; larger organisations need a commercial licence from Black Forest Labs. FLUX.2, released in November 2025, is still current in five variants: three hosted ones, [max], [pro] and [flex], and two open-weights ones, [dev] and [klein]. This site's demo runs FLUX.2 [dev], and the art-styles page was rendered with FLUX.2 [klein].

See the art styles rendered with FLUX.2 →

Qwen-Image 2.1 is an open-weights text-to-image and image-editing model from Alibaba's Qwen team, released in September 2026. It follows whole sentences rather than keyword lists, writes legible text inside images and renders native 2K, 2048 × 2048 for a square, without an upscaler. For editing, it takes a sentence and up to ten reference images.

You can try it free on this site's Qwen-Image 2.1 demo, with no account. It runs at home in ComfyUI; the int8 weights fit a 16 GB graphics card. The weights are under the Qwen Research License: research and evaluation only, and any commercial use needs a separate licence from Qwen. The first Qwen-Image, released in August 2025, was Apache 2.0.

Read the Qwen-Image 2.1 guide: prompt form, measured settings and native 2K sizes →

The latest release is Stable Diffusion 3.5 , released in October 2024. It produces sharper images with significantly better text rendering, richer colors, and improved fine details. It comes in three sizes: Large (8B parameters), Large Turbo for fast generation, and Medium — all runnable locally with moderate VRAM.

Read the full guide →