Token weights
Add emphasis to a specific term by wrapping it in parentheses with a multiplier:(term:1.3)
- Default weight is 1.0
- Above 1.0 increases influence —
(big breast:1.4) - Below 1.0 reduces influence —
(clothing:0.6) - Stay between 0.5 and 1.5 — outside that range the model produces artifacts
(long red hair:1.3), (pale skin:1.2), (black lingerie:0.8)
Good: a woman, (detailed face:1.3), soft studio light, photorealistic
Bad: a woman, (beautiful:2.0), (perfect:2.0), (amazing:2.0) — stacking high weights on vague sentiment terms produces oversaturated, distorted output.
Negative prompts
Open the Advanced toggle to access the negative prompt field. Negatives tell the model what to exclude from the output. Use them to fix recurring problems rather than as a general quality filter. Useful negatives by problem:| Problem | Negative terms |
|---|---|
| Deformed hands | deformed hands, extra fingers, missing fingers, fused fingers |
| Bad anatomy | bad anatomy, malformed body, extra limbs |
| Flat lighting | flat lighting, overexposed, washed out |
| Unwanted text | text, watermark, signature, logo |
| Blurry output | blurry, out of focus, low resolution |
| Clothing when unwanted | clothing, dressed, covered |
Token order
The model weights tokens that appear earlier in the prompt more heavily. Use this deliberately:- Put the subject and the most important visual elements first
- Put style, quality tags, and camera settings last
- If two elements conflict, whichever appears first usually wins
a woman, red hair, pale skin, black lingerie, sitting on a bed, soft window light, shallow depth of field, photorealistic, high detail
Example — quality tags first (wrong):
photorealistic, high detail, 8k, a woman, red hair — the model treats the quality tags as the primary subject and the actual subject as secondary.
Combining LoRAs with weighted prompts
When using a LoRA, add trigger words at the start of the prompt before other descriptors, then use weights to balance LoRA influence against the rest:<lora:64:0.6>, big breast, pink nipples, (a woman:1.0), dark hair, hotel room, soft light, photorealistic
If the LoRA overpowers the composition, reduce its weight. If it’s not visible, increase it — but stay below 0.9.
Common failure modes
Hands and fingers are deformed
Hands and fingers are deformed
Add
deformed hands, extra fingers, missing fingers, fused fingers to the negative prompt. For close-up hand shots, also add (hands:0.8) in the positive prompt to reduce the model’s focus on them.The output ignores part of the prompt
The output ignores part of the prompt
The prompt is too long or has conflicting terms. Cut it down to the ten most important elements and remove any redundant quality tags. Check token order — critical elements should be first.
The output looks oversaturated or distorted
The output looks oversaturated or distorted
Token weights are too high. Bring everything above 1.3 back down to 1.1–1.2 and regenerate.
Style and content conflict
Style and content conflict
The model and the style tags disagree. If you’re using Realism, drop anime or illustration style tags entirely — they confuse the model rather than blending styles.
Prompting Basics
Prompt structure and element order.
Model-specific tips
Per-model prompting quirks.