Stable Diffusion and SDXL prompt generator

What the output looks like

Stable Diffusion interfaces such as AUTOMATIC1111, Forge, ComfyUI or InvokeAI have two text boxes, and the builder writes for both. The positive prompt is a list of comma-separated tags; the negative prompt follows on its own line. Below the result you will also see the recommended SDXL size.

(a samurai standing in a bamboo forest:1.2), cinematic film still, volumetric light rays, low-angle shot, rule of thirds composition, mysterious atmosphere, green and gold color palette, morning mist, highly detailed, sharp focus, best quality

Negative prompt: modern buildings, lowres, blurry, jpeg artifacts, bad anatomy, extra fingers, deformed hands, watermark, signature, text

Recommended SDXL size for 16:9: 1344×768

The tags highly detailed, sharp focus, best quality are added to every prompt as general quality boosters. Your own avoid terms come first in the negative prompt, followed by a default set covering low resolution, artifacts, anatomy problems and watermarks.

Understanding weights like (term:1.2)

In most Stable Diffusion interfaces, wrapping words in parentheses with a number changes how strongly they influence the image. (red scarf:1.2) means about 20% more emphasis; (red scarf:0.8) means less. The builder weights only the subject, because the most common problem is the subject getting lost among style tags. You can add your own weights after copying:

• Stay between 0.8 and 1.4. Higher values often create distorted textures or repeated objects.
• Weight one or two key ideas, not every tag; if everything is emphasized, nothing is.
• If an element appears too often, lower its weight instead of adding it to the negative prompt.

Note that weighting syntax can differ between interfaces and some online services ignore it, so check how your tool handles parentheses.

Recommended SDXL resolutions

SDXL was trained around one megapixel, so it produces the cleanest results at sizes close to 1024×1024 in total area. Very different sizes can cause duplicated subjects or stretched bodies. The builder shows these values for each ratio:

Older SD 1.5 models were trained at 512 pixels and need smaller sizes; follow the notes of the checkpoint you use. Sampler, steps, CFG scale and seed are set in your interface, not in the prompt.

Fixing common SD results

• Subject missing or tiny: raise its weight to 1.3 or move camera tags like “wide-angle” out.
• Hands and faces broken: keep the anatomy terms in the negative prompt and try a close-up or portrait framing.
• Washed-out or noisy images: pick a lighting tag and remove contradictory mood words.
• Two subjects blending into one: give each its own short phrase with a position, such as “a cat on the left, a dog on the right”, or generate them separately.
• Style ignored: move the style tag right after the subject and remove competing style words.

For a negative prompt tailored to your problem, use the negative prompt generator. To start from an existing picture, image to prompt can write an SD-formatted line, and the image prompt generator shows the same idea for other models.