Why Stable Diffusion prefers tags
Most Stable Diffusion checkpoints were trained on images paired with short captions and tag lists, so they respond best to compact, comma-separated tags rather than flowing sentences. Our generator orders the tags the way experienced users do: main subject first, then subject details, setting, composition, lighting, colour palette, style and finally quality tags.
Order matters because of how prompts are encoded. The text encoder reads prompts in chunks of about 75 tokens, and earlier tags tend to have more influence. Keeping the prompt focused, typically 25 to 45 tags, gives each tag room to count instead of competing with dozens of near-duplicates.
Weights: (tag:1.2) and when to use them
Interfaces such as ComfyUI, Forge and AUTOMATIC1111 support attention weights written as (tag:1.2). A value above 1 strengthens a tag, below 1 weakens it. We add a moderate weight to the one or two tags that define your image, usually the main subject, and leave everything else at the default.
Resist the urge to push weights above about 1.4. Very high weights often cause artefacts, oversaturated colours or repeated elements. If something still doesn’t show up, moving the tag earlier in the prompt usually works better than raising its weight further.
Negative prompts that fit the image
The second line of every result is a negative prompt: things the model should steer away from. Instead of a generic wall of fifty tags, we write eight to fifteen that suit the picture, such as extra fingers for portraits, watermark and text for stock-style photos, or blurry and lowres for detailed scenes.
Paste it into your interface’s negative prompt field, not the main prompt. With newer models such as SD 3.5, which understand natural language better, shorter negative prompts are usually enough.
SDXL, SD 3.5 and fine-tuned checkpoints
Tag-style prompts work across SD 1.5, SDXL and most community checkpoints. SD 3.5 also accepts longer natural-language descriptions, so if your results feel stiff there, try the General format instead. Anime and illustration checkpoints often expect their own tag vocabulary, so keep the style tags your checkpoint’s documentation recommends.
A quick example
For a photo of a lighthouse at sunset, a typical result starts with (white lighthouse:1.3) and rocky coastline, moves through golden hour light, long exposure sea, wide-angle view and matte painting style, and ends with quality tags such as highly detailed and sharp focus. The negative line might read blurry, lowres, watermark, text, oversaturated. Each tag names one visible fact, which makes it easy to swap one out and see exactly what changes.
Guides from our blog
- Why your AI images look generic, and seven prompt habits that fix it
AI image prompts that say how a picture should look, not just what is in it: seven habits, with before-and-after examples, for sharper results.
- How to reverse-engineer any image into a prompt, and when it works
A practical reverse prompting workflow: describe an image, format it for your generator, and refine one element at a time. What works, and what never will.