NovelAI Diffusion Prompt Guide

Paste (silver_hair:1.4) into NovelAI and nothing happens — that is AUTOMATIC1111 syntax, and NAI quietly ignores it. NovelAI Diffusion V3 is one of the strongest anime-and-illustration models, and like its imageboard-trained peers it speaks in danbooru-style tags, but it applies emphasis through its own bracket system: curly braces strengthen a tag ({tag}, {{tag}} stronger still) and square brackets weaken it, each layer adjusting weight by roughly five percent. Quality boosters (best quality, amazing quality, very aesthetic, absurdres) go at the front; negatives live in the Undesired Content field with preset fill-ins. This guide assembles a correct positive prompt, shows the exact weighting syntax for any tag and weight you enter, and recommends sampler and step settings — conventions documented in the official NovelAI docs.

NAI’s weighting syntax in full

The key distinction between NAI and AUTOMATIC1111 / Forge is the bracket system:

NAI does not support (tag:1.3) syntax — that is an AUTOMATIC1111 convention and is ignored or treated as literal text in NAI.

A practical example: if a character’s hair color is not appearing strongly enough, try {{silver_hair}}. If the background is competing with the subject, try [background] to push it back.

Quality tags and prompt structure

For V3, structure a positive prompt like this:

1. Quality boosters (always first): best quality, amazing quality, very aesthetic, absurdres
2. Subject and character: 1girl, solo, long silver hair, blue eyes, school uniform
3. Action or pose: standing, looking at viewer, smile
4. Setting and environment: outdoors, cherry blossoms, spring
5. Lighting and mood: soft lighting, afternoon, warm tones
6. Weighted emphases: {{detailed face}}, {natural lighting}

The quality boosters at the front are what activate V3’s cleaner outputs — without them you get a noticeably less polished result even with the same character tags.

Undesired Content levels

NAI offers preset Undesired Content levels in the UI that fill in common negative tags automatically:

• Light — removes the worst anatomy artifacts and low-quality markers
• Heavy — adds more specific negatives for common failure modes (bad hands, deformed limbs, duplicate features)

The Heavy preset is the practical default for character images. You can add your own tags on top of the preset, such as text, watermark, signature if those appear in your outputs.

Sampler and steps guide

For most V3 outputs, 28 steps is a reliable sweet spot. Pushing to 40-50 steps rarely improves V3’s output quality significantly and costs more Anlas per generation.

Tips and notes

• Stack braces, don’t guess numbers. Want a stronger color? {{blue_eyes}}, not a numeric weight — NAI ignores AUTOMATIC1111-style (tag:1.4).
• Use the presets, then refine. Turn on Quality Tags and a sensible Undesired Content level, then layer your own tags on top.
• 28 steps is plenty. V3 converges quickly; spending Anlas on 50+ steps rarely pays off.
• Euler a for variety, DPM++ for consistency. Switch samplers when you want more or less variation between seeds.
• Underscores vs. spaces both work. Danbooru tags are canonically underscored (long_hair), but NAI tolerates spaced forms; pick one style and stay consistent so your weighted tags match your base tags.

Troubleshooting common failures

• A tag is being ignored. Check its spelling against danbooru’s canonical form first; a tag the model never saw in training cannot be emphasized into existence. {{{misspelled_tag}}} does nothing.
• Over-emphasis artifacts. Past three or four brace layers, colors bleed and anatomy warps. If {{{{tag}}}} still is not enough, the problem is tag choice, not weight — find a more specific tag.
• Style drift across a batch. Fix the seed and change one variable at a time; NAI’s samplers differ enough that comparing prompts across samplers tells you nothing.
• Negatives leaking into the image. Undesired Content is a negative conditioning, not a ban list; heavily weighted negatives can distort composition. Keep UC lean and targeted.

If you are on a newer NAI model

NovelAI has shipped newer model generations since V3. The house conventions this guide teaches — brace/bracket weighting, front-loaded quality boosters, the Undesired Content field — are NAI-wide UI behavior, but each model generation has its own recommended booster set and sampler sweet spots. When you switch models, keep the syntax and re-check the boosters against the current official documentation rather than carrying V3’s booster string forward blindly.

Sources

• NovelAI official documentation — weighting syntax, quality tags, and Undesired Content presets
• NovelAI — the hosted service where these settings apply

The prompt builder runs entirely in your browser; no prompts are sent to NovelAI or anywhere else.