AI Image Generator takes a text prompt and produces one or more images using diffusion models including FLUX and Stable Diffusion variants. You choose the model, set the output size (128–8192 pixels), and optionally upload up to 16 reference images to guide the style or composition. Generated images are available for immediate download and support PNG, JPG, and WebP output.
How prompt wording affects results
The more specific the prompt, the more reliably the model finds a consistent generation direction. "A cat" and "an orange tabby cat sitting on a windowsill, soft window light from the left, background blurred, cinematic color grading" produce very different results. Covering four dimensions — subject, style, lighting, and composition — typically satisfies most visual needs.
The negative prompt field lets you exclude unwanted elements: blurry, low quality, watermark, deformed hands. Not all models respect the negative prompt — when unsupported, the field is silently ignored.
Steps and CFG Scale
Steps control denoising iterations. 20–30 steps is fast and good for confirming direction; 40–50 steps adds detail for a final render. Beyond 50 steps, returns diminish quickly.
CFG Scale controls how strictly the model follows the prompt. 7–12 is the common range: lower values give the model more creative latitude and produce more variety, but results may drift from the description; values above 14 tend to produce oversaturated or color-shifted images. FLUX-family models have different CFG sensitivity than SD-based models and rarely benefit from values above 7.
Fast iteration phase
- Steps 20–25
- CFG Scale 7
- 1 image, default size
- Confirm composition and style before adjusting further
Final render phase
- Steps 35–50
- CFG Scale 7–10 (tune as needed)
- Generate 2–4 images to choose from
- Use your intended output size
What reference images do
Uploading reference images (up to 16) lets the model extract style, composition, or subject characteristics from them to guide the output. How reference images are used varies by model: some treat them as style guidance, others for subject adaptation, and some (such as models requiring reference input) cannot run without them. When reference images are marked "required," omitting them blocks generation.
Output size constraints
Output supports PNG, JPG, and WebP at 128–8192 pixels. Some models require dimensions to be multiples of 64, and submissions are aligned automatically. For unconventional aspect ratios — very wide banner formats, for instance — some models may produce unstable compositions.