Illustrious and Anima Style Tags: What Actually Works

TL;DR

• Danbooru tags are the vocabulary: Both models train on Danbooru2023, so canonical tags like anime screenshot work; aliases like anime screencap may not register if the model saw only the renamed form.
• Artist tags need @ in Anima: The model was trained with @artist prefixes; omitting @ yields very weak effect.
• Test tag registration via controlled ablation: Fix seed, hold all other tokens constant, remove the candidate tag, and compare outputs along a single axis—Anima's tag dropout means a missing tag alone doesn't prove non-recognition.
• Build style shorthands to isolate training: Always tag style in training data; drop the shorthand when training character LoRAs to prevent contamination.

For the broader prompt-construction context, see the Illustrious and NoobAI anime prompt guide.

Where Style Tag Vocabulary Comes From

Illustrious XL and NoobAI XL are both fine-tuned SDXL derivatives trained primarily on Danbooru-derived datasets.

Illustrious XL is trained exclusively on Danbooru2023, while NoobAI XL also incorporates e621 tags, expanding its coverage of authoritative captions from two sources.

The Anima model card confirms this directly: "The model is trained on Danbooru-style tags, natural language captions, and combinations of tags and captions." This hybrid training explains why both approaches work: you can write a descriptive sentence like "A girl in a cafe" followed by comma-separated Danbooru tags like 1girl, red hair, school uniform.

Illustrious XL's official site describes the same architecture: "Blending advanced Natural Language Processing with structured Danbooru tags to deliver flawless prompt adherence." The model accepts hybrid prompts seamlessly, interpreting both English and tags as a unified signal.

For deeper coverage of Anima's training-time caption format, see the Anima LoRA training guide.

Danbooru Tag Rename: anime screencap → anime screenshot

Danbooru's wiki shows that anime screenshot is the current canonical tag. The historical tag anime_screencap has been aliased to it.

This rename matters for model recognition because checkpoints trained before the alias may only respond to the legacy token. Anima's model card lists the post-rename form anime screenshot as a meta tag and has no entry for anime screencap, suggesting it was trained after the rename (source 31037 notes "For Anima this tag is instead 'anime screenshot' due to a rename on Danbooru's end before that base model was trained.").

Illustrious XL's documentation does not explicitly confirm which spelling it was trained on. If you're targeting Illustrious specifically, test both forms—anime screenshot is safer as the canonical current tag.

The Artist Tag Prefix: Why @ Matters in Anima

Anima requires the @ prefix for artist tags. The model card explicitly states: "Prefix artist with @. E.g. '@big chungus'. You must put @ in front of the artist. The effect will be very weak if you don't."

This is not merely convention; it's a tokenization requirement. Anima was trained with @-prefixed artist tokens, so the model does not recognize the pattern without it. Using artist_name alone will produce minimal or no style effect compared to @artist_name.

The community has also developed a mixing syntax: @[artist1|artist2|artist3] with optional weights like @[artist1:4|artist2:3|artist3:2]. This shorthand appears in user discussions as the preferred way to blend multiple styles.

Prompt Weighting Syntax from AUTOMATIC1111

Both models inherit prompt-weighting from AUTOMATIC1111's Stable Diffusion web UI. The syntax uses () to increase attention and [] to decrease it:

• (keyword) increases attention to the enclosed words
• ((keyword)) increases attention further (nested parentheses amplify the effect)
• [keyword] decreases attention to the enclosed words
• [[keyword]] decreases attention further (nested brackets amplify the decrease)

The BREAK keyword also works as a hard chunk separator, useful for multi-concept prompts where you want independent processing.

Testing Whether a Tag Actually Registers

Anima's model card reveals a key insight: "The model was trained with random tag dropout. You don't need to include every single relevant tag for the image."

This means a tag's absence from your prompt doesn't prove it's not in the model's vocabulary. To test registration:

1. Fix the seed to eliminate stochastic variation
2. Hold all other tokens constant—only the candidate tag changes
3. Compare along a single axis—is the style shift visible?

For example, to test if flat color registers: generate with anime screenshot, flat color, 1girl at seed 12345, then regenerate with anime screenshot, 1girl at the same seed. Any visible difference in shading indicates the tag has an effect. See the captioning tools guide for seed management best practices when running ablation tests.

Building a Reusable Style Shorthand

The community tag catalogue recommends always including style tags in training data to prevent contamination between style and character concepts. This practice is critical when training LoRAs.

• For style LoRAs: Always tag the style (e.g., flat color, no lineart, pixel art) so the model learns to associate those tokens with the visual characteristics.
• For character LoRAs: Include style tags in training data, but drop them when prompting the LoRA so you don't accidentally blend styles. See the LoRA library management guide for organizational conventions when storing multiple style LoRAs.

A practical shorthand might look like:

Style: anime screenshot, official art, flat color, no lineart
Artist mixing: @[artist1|artist2:2]
Quality: masterpiece, best quality, score_7

This format keeps prompts organized and makes it easy to swap styles or artists without rewriting entire prompts. For face and hand cleanup when style tags cause degradation, see the fix-faces and Adetailer workflow guide.

Common Style Tags and Their Effects

Based on community testing and documentation, these tags have documented effects:

Quality Tags: score_9 through score_1

Anima uses abstract quality tags similar to PonyV7: score_9 through score_1, where higher numbers indicate better quality. These act as quality anchors in prompts.

NoobAI XL's blog recommends starting prompts with quality anchors, then adding character and artist tags for maximum control. This ordering matters because the model weights early tokens more heavily.

Errors and Fixes

FAQ

Does Danbooru renaming a tag break model recognition across checkpoints?

Yes, potentially. Models trained before a tag rename may only respond to the legacy spelling. Anima was trained after the anime screencap → anime screenshot rename, so it likely only recognizes the new form. For maximum compatibility, use the current canonical tag.

What's the difference between realistic, photo (medium), and photorealistic in anime checkpoints?

photo (medium) captures actual photographs or film frames. photorealistic is an illustration designed to look like a photograph. realistic is a semi-realistic illustration—think Mona Lisa or Norman Rockwell—with realistic proportions but clearly still art.

How do I verify a new style tag is actually understood by a model?

Use controlled ablation: fix the seed, hold all other tokens constant, remove the candidate tag, and compare outputs. Anima's tag dropout means a missing tag alone doesn't prove non-recognition.

Can I mix multiple artist styles in Anima without control nets?

Yes, using the community shorthand @[artist1|artist2|artist3] with optional weights like @[artist1:4|artist2:3]. This syntax appears to be the preferred method, though its official support varies by Anima version.

What's a reusable shorthand for character vs style LoRA tagging?

Always include style tags in training data. When prompting a character LoRA, drop the style shorthand to prevent contamination. A practical format: Style: anime screenshot, flat color, no lineart | Artist: @[artist1|artist2:2] | Quality: masterpiece, score_7.

Why does Anima drop some tags, and what should I do about it?

Random tag dropout during training means the model is robust to missing tags at inference. However, if a tag isn't producing expected effects, test it with controlled ablation. Most tags with clear visual effects (like flat color or anime screenshot) will register consistently.

Sources

• Danbooru wiki page for anime screenshot: https://danbooru.donmai.us/wiki_pages/anime_screenshot
• AUTOMATIC1111 Stable Diffusion web UI features documentation: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features
• Illustrious XL official site: https://illustriousxl.org/
• NoobAI XL official site: https://noobaixl.org/
• NoobAI XL blog guide: https://noobaixl.org/blog/noobai-xl-guide
• Anima model card (HuggingFace): https://huggingface.co/circlestone-labs/Anima