Anime-tuned SDXL checkpoints do not respond to prompts the way general-purpose ones do. Write a fluent English sentence and you get a mediocre result; write a comma-separated list of booru tags in the right order and the same model produces something clean and on-model. This is about that difference - which prompt grammar these checkpoints expect, why, and how to build one that works.
Why base SDXL struggles with anime
Base SDXL is a generalist. Its training set is dominated by photographs and general web imagery, so its idea of anime is an approximation assembled from whatever illustrated material happened to be in there. Ask for it and you get anime-adjacent output: proportions slightly off, line weight inconsistent, eyes rendered with a soft photographic falloff instead of the flat cel shading the style depends on.
A specialised fine-tune fixes this by retraining on a large illustration corpus. Animagine XL, from Cagliostro Research Lab, is one of the better-known examples, and versions of it are widely mirrored on Hugging Face and Civitai. What matters more than the specific checkpoint is what that retraining does to the text encoder's expectations.
Tags, not sentences
These models are trained on images labelled with booru-style tags - the vocabulary used by large anime image boards. Every training caption is a comma-separated list of short, standardised tokens, not prose. So the model has essentially never seen a sentence describing an image, and the fluent English you would write for a photographic checkpoint is out-of-distribution input.
The practical consequence is that phrases the model has never been trained on get partially ignored, while an exact tag it has seen tens of thousands of times lands precisely. 1girl, long hair, school uniform, classroom, sunlight will outperform a carefully written sentence describing the same scene, every time.
The grammar is worth internalising because it is stable across the whole family of anime fine-tunes, not specific to one checkpoint. A common ordering looks like this:
| Slot | Purpose | Examples |
|---|---|---|
| Subject count | How many figures, and gender |
1girl, 2boys, solo
|
| Character | Named character, if any | character tag as spelled on the boards |
| Source | Series the character comes from | series tag |
| Appearance | Hair, eyes, clothing, expression |
silver hair, red eyes, smiling
|
| Composition | Framing, pose, angle |
upper body, from below, sitting
|
| Scene | Location, lighting, background |
rooftop, sunset, cityscape
|
| Quality | Aesthetic steering tokens |
masterpiece, best quality
|
Order carries weight in a CLIP-conditioned model - tokens near the front influence the result more - so put what matters most first and let scene detail trail.
Character tags are lookups, not descriptions
If you want a specific character from a known series, the model very likely has a tag for them, and using it is dramatically more effective than describing their appearance. But the tag has to be spelled the way the training data spells it, which usually means the exact booru form. A near-miss silently degrades to a generic character with the right general vibe.
This is also the reason these checkpoints feel uncannily good at some characters and useless at others. It is a function of how many tagged examples existed, nothing more.
Quality and era tags
Most anime fine-tunes are trained with aesthetic-score tags baked into the captions, which is why masterpiece, best quality at the end of a prompt does something measurable rather than acting as superstition the way it does on many photographic checkpoints. It is a real token the model associates with a higher-rated slice of its training data.
Later checkpoints in the Animagine line extended this with era tags that select for the visual conventions of different periods of anime illustration - flatter palettes and heavier lines at one end, current rendering conventions at the other. When a checkpoint supports them, they are the cheapest way to shift the whole aesthetic without touching anything else in the prompt. Check the model card, because the exact vocabulary differs between releases.
Negative prompts earn their keep here
On anime checkpoints the negative prompt does more visible work than on most photographic models, largely because the failure modes are so consistent. A serviceable baseline:
lowres, worst quality, low quality, jpeg artifacts, bad anatomy, extra digits, missing fingers, watermark, signature, username, text, blurry
Add nsfw if the checkpoint is capable of it and you do not want it. Beyond that, resist the urge to keep growing the list. Very long negatives dilute each token's effect and start suppressing things you wanted.
Settings that matter
These are SDXL fine-tunes, so the usual SDXL rules apply. Generate at roughly one megapixel and use the standard aspect buckets rather than arbitrary dimensions - 1024 x 1024 square, or something like 832 x 1216 for a portrait. Push much past that natively and you get duplicated heads and doubled bodies, the classic sign of exceeding the training resolution.
A moderate CFG in the middle of the range and somewhere around 25 to 30 steps is a sensible starting point; anime checkpoints tend to burn out at high CFG faster than photographic ones do, because flat colour has nowhere to hide extra contrast. Always check the model card first - fine-tunes ship with recommendations and they are usually right.
Two prompts, two grammars
The contrast is easiest to see side by side. Here is a prompt written for a general-purpose SDXL checkpoint, in natural language, leaning on photographic vocabulary:
long exposure photo of a Paris street, Eiffel Tower, streaks of light, surreal, dreamy, ghosting effect, highly detailed
And here is the same intent expressed for an anime fine-tune, as tags:
1girl, solo, from behind, city street at night, eiffel tower, neon lights, light trails, wide shot, detailed background, masterpiece, best quality
Run each on the wrong model type and both underperform. That is the whole lesson in two lines.
Practical takeaways
Anime fine-tunes speak tags because they were captioned in tags, so translate your idea into short comma-separated tokens and put the important ones first. Use exact character and series tags when you know them - they are lookups into the training data, not descriptions. Quality tokens are functional on these checkpoints rather than decorative, and era tags, where a release supports them, restyle everything at once. Keep the negative prompt short and targeted, generate at standard SDXL resolutions, and read the model card before assuming settings carry over from another checkpoint.
Top comments (0)