Anima negative prompt vs Illustrious: trim the long bad-hands list first

When generating with Anima, the temptation is to paste the long negative prompt I had been using with Illustrious or SDXL.
Something like bad hands, extra fingers, watermark, text, worst quality.
I write it that way too out of habit, but if you look at how the Anima authors recommend it, the right starting point is much shorter.

The Anima official model card recommends masterpiece, best quality, score_7, safe, on the positive side, and worst quality, low quality, score_1, score_2, score_3, artist name on the negative side.
It also notes that Anima was trained on Danbooru-style tags, natural language, and a mix of both, with tags written in lowercase and separated by spaces, only the score tags using underscores.
Anima will still parse an SDXL-style tag spell as-is, but the tag system is worth rebuilding once according to Anima’s own rules.

In the Anima-Base v1.0 on M1 Max writeup I re-read the README’s prompt conventions and tested putting quality / meta / year / safety tags at the front of the positive side.
The negative side here follows the same direction: rather than dumping a giant block of banned words from the start, the official short starting point plus targeted additions is easier to track.

Anima’s official minimal negative

Start with this.

worst quality, low quality, score_1, score_2, score_3, artist name

Quality-killer tags, low score tags, and the artist-name leak filter. Nothing else.
The bad anatomy and bad hands that Illustrious users tend to drop in are not in Anima’s default.
The Anima docs say the base model “will produce unwanted content when given short or under-detailed prompts,” and the recommended workaround is the safety tag plus a sufficiently detailed prompt.
The design assumes that before you stuff everything into the negative, you write a richer positive.

GIGAZINE’s Anima-Base walkthrough also recommends the same short negative for quality control.
So as of May 2026 (Anima-Base v1.0), the official-aligned starting point converges on “short quality negative + safe”.

WAI-Anima adds a few more

The WAI-Anima v1 distribution page bumps the recommended negative slightly.

worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, lowres, censor

blurry, jpeg artifacts, lowres, and censor are added.
In the WAI-Anima writeup I noted that this family has rougher tag control than WAI-Illustrious and tends to wobble on composition and clothing.
For WAI-Anima, adding just image-degradation and censorship terms like this is reasonable.

On the other hand, throwing in long_body, missing_limb, disconnected_limbs, liquid fingers all at once makes it hard to tell which entry the result actually reacted to.
Since Anima also reads natural language, writing pose, clothing, and number of characters concretely on the positive side often stabilizes the output more than piling up bad anatomy-style negatives “to be safe”.

Split the long Illustrious negative by role

On Illustrious or WAI-Illustrious you commonly see something like this as the negative.

nsfw, text, watermark, bad hands, extra digit, fewer digits, worst quality,
ugly, bad anatomy, jpeg artifacts, low resolution, out of focus,
missing arms, extra arms, extra legs, extra fingers, missing fingers,
signature, username, stamp, date

In the WAI-Illustrious v17 writeup I used the distribution-page negative bad quality, worst quality, worst detail, sketch, censor, only adding nsfw when I wanted to suppress NSFW output.
Illustrious-family models learn rating tags as general / sensitive / nsfw / explicit, so using nsfw on the negative side as a filter is straightforward.

Anima also documents the safety tags safe, sensitive, nsfw, explicit.
But the official recommendation is to put safe at the front of the positive side, and keep the negative side a short quality list.
To avoid NSFW, the cleaner order is: safe on the positive side first, then add things like nsfw, explicit, nipples, pubic hair on the negative side.

Don’t expect the negative alone to fix hands and limbs

bad hands, extra fingers, fewer digits, missing fingers is fine to keep.
But if you let the negative side carry all the hand/limb correction by itself, Anima’s composition spec stays too rough.

If you want a full-body shot with visible hands, write this on the positive side as well.

full body, standing, both hands visible, open hands, five fingers, front view

If hands can be hidden, push it the other way.

upper body, hands out of frame

Hammering bad hands on the negative side won’t help much when the model has no idea how the hands are supposed to appear in frame — it will just compose around the problem in a different way.
Anima picks up natural-language descriptions too, so adding a short “what you want to see” alongside the “what you want to avoid” makes the situation easier to separate.

`2D, flat color, oil painting` should change with intent

The user-supplied example oil painting, 2D, flat color collides with itself if you put all three in the same negative field.
Anima’s official docs mention LAION-POP and DeviantArt as non-anime datasets in the training mix, and even document deviantart and ye-pop dataset-tag syntax.
Putting oil painting in the negative to avoid an oil-paint or thick-paint look is fine, but if you want flat anime shading, removing flat color also cuts the direction you wanted.

Words that were in SDXL prompts “just because” — drop them in Anima first, and only add them back when the output actually breaks in that direction.
In particular 2D is too close to anime art itself; if you want to avoid 3D-ness, 3d, cgi, render on the negative is a more specific call.

Composition negatives are easy to keep on Anima

multiple views, split view, grid view, two shot, cropped, out of frame is worth keeping on Anima too.
Anima is strong at multi-character composition and natural-language positioning, but with short prompts it sometimes spreads the composition out anyway.
In the Anima inpaint on Mac ComfyUI writeup I noted that the original pose and composition can leak into the result, and there were cases where the negative prompt alone couldn’t remove that.

For a single standing character, the negative side stays like this.

multiple views, split view, grid view, two shot, cropped, out of frame

On the positive side, write through to solo, full body, white background, front view.
Rather than killing the split layout from the negative side, get the model to read it as “a single full-body standing portrait” from the start.

My current starting point

For SFW character images on Anima-Base v1.0, this is what I run first.

worst quality, low quality, score_1, score_2, score_3, artist name,
blurry, jpeg artifacts, lowres, text, watermark, signature, username,
bad hands, extra fingers, fewer digits, missing fingers,
multiple views, split view, grid view, cropped, out of frame,
nsfw, explicit, nipples, pubic hair

Positive-side prefix:

masterpiece, best quality, score_7, safe,

The long negative I carried over from Illustrious is easier to handle when split by category: quality, text, hands, composition, safety.
Evaluation words and body-shape words like ugly, bad anatomy, long_body, liquid fingers, fat, huge breasts — I don’t include them from the start.
Only when the output actually breaks in that direction do I add the matching term.

The starting point for Anima isn’t a huge negative block. It’s the short official negative, the positive side with safe up front, and a concrete natural-language description — together.
If you bring an Illustrious-style spell over, trim it before you paste it.