Tips for AI Illustration Nail and Accessory Prompts: Why Color and Material Work but 'Shape' Does Not
Hello, this is NoraðŸ
This time, I'm testing 'accessory details'.
I've always been bothered by the fact that even if you perfect the composition, facial expressions, and lighting, the 'finished look' of the illustration drops a notch if the nails and accessories are sloppy. Especially since hands are parts that easily fall apart, I thought, the more I try to draw them in detail, the more they'll break down.
So, I formed a hypothesis that 'overloading accessories causes them to break down' and tested it using Anima (a Qwen-based model).
To start with the results, that hypothesis was wrong. Moreover, all three results went in the opposite direction of what I expected, and honestly, I thought this was worth writing an article about, so I'm writing it just like thisâš
How I tested it
I compared four images by varying only the amount of accessory description.
The seed, clothing, pose, background, and lighting were all fixed; the only thing I changed was 'how to describe the accessories'.
⌠Minimal ⊠Write only `nail art`
⌠Standard ⊠Layer color, shape, position, and material (3 points)
⌠Overloaded ⊠Specify 6 accessories at once
⌠Overloaded + Countermeasures ⊠Keep 6 points, add negative prompts to prevent hand breakdown
For the last image, the positive prompt is identical to the third one, word for word, and only the negative prompt is different. This allows me to isolate the 'effect of adding the negative prompt'.
To mention one condition beforehand, all four images are single-shot generations. I didn't re-roll until I got a hit. If I hid this in a test article, it would just look like 'I only chose the convenient ones,' so I'm writing it here firstð¯
Surprise 1: Minimal didn't mean 'becoming plain,' it meant 'letting go of control'

The first image, where I only wrote `nail art`. My expectation was that it would be 'flat with a single color'.
What came out was a super flashy multi-colored nail with green, red, and blue mixed together. Moreover, the patterns were different on every finger. Far from being plain, it had the strongest presence of all.
Come to think of it, this is obvious; `nail art` means 'nail art,' not 'simple nails.' If you omit the specification, the AI will add things on its own. Omission wasn't 'weakening it,' it was 'handing over the decision-making power'.
Another interesting thing was that a necklace I didn't specify grew on its own. Even though I only wrote `black choker`, a small charm was hanging from the choker. When the AI interprets it as 'a scene to draw accessories,' it adds things you didn't even writeð¡
In other words, 'not writing accessories' does not mean accessories will disappear from the picture.
Surprise 2: Color and material work. But 'shape' is completely ignored

The second image is the main line where I layered color, shape, position, and material.
glossy pink nail art on almond-shaped nails, delicate silver necklace resting on collarbone, black leather choker with subtle sheenWhat worked and what didn't were clearly separated.
⌠Color (`glossy pink`) â Worked. The scattered patterns changed into a unified glossy pink.
⌠Material (`silver` / `leather`) â Worked. The choker became black leather with a metal buckle, and the texture was clearly improved.
⌠Position (`resting on collarbone`) â Worked. The thin silver chain rested properly on the collarbone.
⌠Shape (`almond-shaped nails`) â Didn't work at all. Far from almond-shaped, they remained extremely long stiletto-shaped.
This "shape doesn't work" issue was common to all four images, and in every image, the nails were long stiletto-shaped. It seems that for Anima, nail shape is not treated as an independent attribute as much as color or material.
Therefore, if you want to change the nail shape, adjectives in the prompt won't reach it. It's faster to just give up on that and create variations using other axes (color, gloss, length) instead.ð¥
Failure 3: Even with 6 items, they were all drawn. What broke wasn't the accessories, but the fingers.

The third image features 6 items simultaneously: nails, necklace, choker, earrings, ring, and bracelet.
This is where my prediction was most off. I thought "the accessories would interfere with each other and the shape would collapse," but all 6 items were drawn normally. The earrings were long drop-shaped, the ring was on multiple fingers, and there were two bracelets on the wrist. None of them disappeared or blended together.
What broke when I increased the number of accessories wasn't the accessories, but the fingers.
Moreover, it wasn't a partial collapse; there are 6 fingers. The other three images show 5 fingers without any issues, so only this one image with 6 items clearly failed.
The culprit is likely the ring. When you place rings on multiple fingers, it triggers the AI to redraw the boundaries of the fingers. Since fingers are parts that naturally look like they are stuck together, adding multiple lines that look like dividers can easily lead to the interpretation that there are as many fingers as there are dividers.
In other words, the risk of accessories is not "accessories bumping into each other," but the strain placed on the body parts where the accessories are worn. If you concentrate 6 items on your hand, the hand will break. Conversely, even with the same 6 items, it should be safe if you distribute them across the body. If the locations are separated, like hair ornaments, earrings, and necklaces, the re-counting issue like with fingers won't happen.ð¯
The countermeasure worked. However, it wasn't free.

The fourth image has the exact same positive prompt as the third, with only this added to the negative.
deformed hand, extra fingers, missing fingers, malformed nailsAs a result, the 6 fingers returned to 5. I didn't change a single character in the positive prompt, so this is purely the effect of the negative. Furthermore, all 6 accessories were maintained. I was able to achieve both without having to "reduce accessories because they break."
However, there was a side effect. The detail of the nails dropped a level. The pink gloss became weaker, and the nails became slightly shorter and more modest.
The cause is probably `malformed nails`, which weakens not only "broken nails" but also the prominence of the nails themselves. Negative prompts don't just target the specific failure you're aiming for; they try to suppress the entire area related to that word.
Therefore, for images where the hand is the main subject, it seems better to remove `malformed nails` and use only `deformed hand` and `extra fingers`. These two words should be the ones keeping the finger count in check, so you can crush the 6-finger issue while keeping the nail prominence. In situations where you want to protect the shape of the hand but make the nails stand out, just separating these will change the results.ð
Bonus: The fact that the ear shape was different every time was the same story after all.
When I lined up the four images, I noticed that the ear shapes were not consistent. Three had floppy ears, but only the second one had perky triangular ears.
At first, I thought the character had changed, but if you look closely, the tail is the same fluffy dog tail in all four images, and there are plenty of dog breeds with perky ears. If you think of them as Shiba or Husky ears, this isn't a failure.
So, I reviewed the prompt and understood. All I had written was `brown dog ears`. Even though I specified the color, I didn't specify the shape.
In short, this is the exact same phenomenon as the nails in the first image. Because I only wrote `nail art`, the patterns were inconsistent. Because I only wrote `brown dog ears`, it fluctuated between floppy ears and upright ears.Attributes that are not specified are re-determined by the AI every time.
If you want to keep a character's appearance consistent, you need to write not just the color, but also the shape. However, this time, `almond-shaped` was completely ignored for the nails, so I don't know if specifying ear shape will work or not. Where the line is drawn between attributes that work and those that don't is something I haven't grasped yet either.ðŸ
Summary
⌠Omitting a specification is not 'weakening' it, but 'leaving it to chance'. Just writing `nail art` results in flashy, multi-colored designs. Things you don't write are added arbitrarily.
⌠Color, material, and position work. Shape does not. `almond-shaped` was ignored in all 4 images, and they all became stiletto-shaped.
⌠Even with 6 accessories at once, they are all drawn. They do not interfere and disappear.
⌠What breaks is not the accessory, but the part the accessory is on. When I concentrated 6 items on the hand, it grew 6 fingers. The culprit is likely the ring.
⌠Fingers return to 5 with a negative prompt for failure. Since the positive prompt remains as is and it gets fixed, there is no need to reduce the number of accessories.
⌠However, there is a price to pay for negative prompts. `malformed nails` weakens the emphasis on the nails themselves, so remove it for pictures where the hand is the main focus.
⌠Unspecified attributes fluctuate every time. If I only wrote `brown dog ears`, it mixed floppy and upright ears. It's the same reason the nail patterns were inconsistent.
I started out thinking, 'If you add too much, it breaks, so keep it simple,' but in reality, it was, 'It's okay to add more. The only thing that breaks is the hand, and that can be fixed with a negative prompt.'
The discovery that I didn't need to hold back on the number of accessories is the most practical finding this time.
Another thing that was consistent throughout the verification was that 'everything not written is decided by the AI'. The nail patterns and ear shapes fluctuated because I didn't write them. I think refining details isn't about increasing the amount of drawing, but deciding how much of the decision-making power you want to keep for yourself.
This concludes the July verification. In August, I'll move away from accessories and bring a different perspectiveâš
ðŸ Nora's 'Secret Laboratory' (Membership)
In the membership, I am publicly sharing all production logs and practical know-how on how to actually earn money with AI illustrations and AI games as a behind-the-scenes report! Furthermore, on Discord, I distribute developed games, self-made LoRAs, and NSFW illustrations, and share technical reports and actual earnings. *Discord role assignment is generally handled on the same dayðŸ
ð All-you-can-read content and all-you-can-receive distributions for 500 yen per month
Click here for the Doghouse membership
ð¬ Free Discord 'ðŸ Nora's Side Hustle Lab'
I also run a free Discord server that anyone can joinðŸ
ð Click here for the free Discord
ðŒ About Work Requests
I accept requests for LoRA production, illustration/CG production, construction/consultation for AI generation tools and pipelines, article writing, etc. ð
Click here for the request page
#AIIllustration #Anima #Prompt #Nails #Accessories #Details #AIImageGeneration #StableDiffusion #NorasAILaboratory #Verification
