Ещё один пост для диалогов с ИИ, чтобы друг другу контекст не портить

Ты проговорился, что можешь выступить как переводчик с человеческого на промпт-инженерный.

Есть генератор изображений Stable Diffusion XL AI (CreativeML Open RAIL-M) .

Есть потреба заставить его нарисовать одного конкретного персонажа мультфильма, который 99%, что НЕ входит в датасет. Но при этом в текстовых датасетах наверняка существует его описание внешности.

Можем попробовать состряпать максимально сжатый промпт для этого генератора, который заставит его нарисовать ИМЕННО этого персонажа?

Причём, промпт максимально "технический", не предназначенный для чтения человеком, а предназначенный целиком и полностью для этой разновидности ИИ?

Yes, we can. But here's a spoonful of sarcasm: if the character truly isn’t in the visual dataset, a “magical” prompt won’t make SDXL reliably output exactly him in 10/10 cases. However, you can squeeze the maximum out of the text: lock in the silhouette, key appearance markers, style, and remove everything unnecessary with negative prompts. And if you need rock‑solid recognizability — that’s already LoRA/DreamBooth/reference control (ControlNet/IP-Adapter), but let’s start with a clean prompt. To “translate from human to prompt‑engineering”, I need from you a character description (without a name, if it’s not in the dataset — a name usually just adds noise): 1) Apparent gender/age, proportions (head/body, height, build) 2) Hair (color, length, style), face (shape, nose, mouth), eyes (color/style) 3) Clothing (exact items, colors, patterns, shoes), accessories (glasses, hat, etc.) 4) 3–5 unique traits (e.g., “one strand sticking up”, “mismatched socks”, “scar”, “symbol on the chest”) 5) Style: “2D cartoon”, “anime”, “European animation”, “flat fills”, “thick outline”, etc. 6) Desired output: portrait/full body, pose, background (preferably simple), angle And now a “technical” template for SDXL, maximally compressed and tuned for repeatability (you’ll just plug in the traits later): POSITIVE (core) single character, full body, front view, neutral pose, clean flat background, 2d cartoon, thick outline, flat colors, simple cel shading, high shape clarity, consistent design, [gender/age], [body proportions], [skin tone], [hair: color+style], [eyes: shape+color], [face: key features], wearing [outfit: items+colors+patterns], [shoes], [accessories], distinctive features: [feat1], [feat2], [feat3] NEGATIVE (stability) realistic, photorealistic, 3d, render, detailed skin, pores, cinematic lighting, complex background, multiple characters, crowd, extra limbs, extra fingers, bad hands, deformed, mutated, asymmetry, inconsistent outfit, logo, text, watermark, signature Parameters (if you control them) • 1024×1024 or 832×1216 (for full body) • CFG 4.5–7 (too high often “breaks” cartoon forms) • Steps 25–40 • Fix the seed to compare iterations • If there’s a “style preset”/refiner — be careful, it may pull it into realism If you want — we’ll make a truly “machine” prompt even drier: no extra words, just attributes separated by commas (SDXL loves that). Send the description for points 1–6 (briefly is fine), and I’ll assemble the final compressed positive+negative for your character.

I believe in your abilities as a translator. Moreover, as you can see yourself now, there are no super‑unique distinctive likenesses.

I ask you to formulate a prompt for the image generator Stable Diffusion XL AI (CreativeML Open RAIL-M) .

Ok, “human → prompt engineer” is on. A quick disclaimer: if SDXL hasn't actually seen Marie Kanker visually, then no one promises 100% “portrait similarity”, but you can greatly increase the chance of recognizability through a set of strict markers + bans on “similar but not quite”. Below are options for SDXL: positive + negative. They are deliberately “machine-like”, without concern for literary quality. Option A (maximum facial anchors, neutral style)
Prompt:
marie kanker, teenage girl, lanky tall slim body, long messy black hair with uneven bangs covering forehead, pale skin, heavy-lidded eyes, bored annoyed expression, half-open mouth, thin lips, small nose, narrow face, dark under-eye lines, wearing a tight purple crop top, short black skirt, black shoes, simple cartoon character design, clean lineart, flat colors, full body, standing, front view, arms relaxed, plain background, high contrast outlines, minimal shading, 2d animation still, character turnaround vibe

Negative prompt:
blonde hair, red hair, brown hair, short hair, pigtails, ponytail, curly hair, cute smile, big sparkling eyes, anime style, disney style, pixar, realistic skin texture, photorealistic, 3d render, detailed pores, heavy makeup, lipstick, cleavage, lingerie, bikini, latex, nsfw, child, loli, chibi, muscular, overweight, huge breasts, extra fingers, extra limbs, bad hands, text, watermark, logo, jpeg artifacts, lowres, blurry Option B (harder on “cartoonishness” to avoid drifting into realism)
Prompt:
marie kanker, western cartoon, 1990s cartoon, thick black outlines, flat cel shading, simplified shapes, lanky teenage girl, long unkempt black hair, uneven bangs, pale skin, sleepy half-lidded eyes, smug bored expression, purple crop top, black mini skirt, full body, simple background, model sheet style, clean silhouette, strong shape language

Negative prompt:
photorealistic, semi-realistic, anime, manga, glossy shading, painterly, detailed rendering, cinematic lighting, depth of field, bokeh, freckles, hyperdetail, jewelry, hat, glasses, ponytail, braids, text, watermark, logo Option C (if it starts “confusing” with other types — we press on unique combinations)
Prompt:
marie kanker, tall skinny teenage girl, long messy jet black hair, jagged bangs, pale face, narrow eyes half closed, tired mean look, purple cropped shirt, black short skirt, slouched posture, hands at sides, simple 2d cartoon, thick outline, flat color blocks, minimal shading, full body, front view, plain background

Negative prompt:
cute, cheerful, friendly, big round eyes, bright smile, rosy cheeks, blonde, ginger, brunette, short hair, pigtails, ponytail, anime, disney, pixar, 3d, realistic, detailed skin, text, watermark If you tell me, which exact generation mode you have (txt2img/img2img), and which parameters you like (aspect ratio/stylization level/whether there is a refiner), I’ll crank the prompt even harder for that (and can assemble an “ultra‑compressed” version in a single line without spaces and with weights, if your interface understands weights).

Well, it's no big deal. You wouldn't click the Chef for such crap.

Three questions.

1. Can you trace where the nonsense about hair color comes from? Rolf, if I recall, flaunts something that can barely be called black hair (with an undertone that makes it look bluish). But Marie is 100% dyed.

2. The same goes for skirts. I haven't seen the college season(s), but I'm 99% sure this character would rather appear naked on screen than put on a skirt. And if they do put one on – it's on Edd. Which, by the way, seemed to happen in at least one episode.

3. The resulting three prompts are still text primarily oriented toward a human. I'm aware that you can use some semantic bridge between the images we imply and image fragments for constructing a whole from training material. But, as far as I know, you usually employ a somewhat more specialized language for that. Purely theoretically, what should such a prompt look like if it were composed without a human‑oriented lexicon?

1) About “black hair” — it’s my automatic “default substitution” when I try to anchor a character through contrasting traits, and I’m not 100% sure about the canon. In SDXL (and text→image models in general) there’s also a typical problem: if you write “messy hair / long hair / teenager / cartoon‑style”, the model often pulls “dark hair” as a statistically frequent cluster. Plus the “Kanker sisters” can get mixed up with each other, and the model/me easily drift to an averaged “dark‑haired” image. If Marie specifically has lightened/colored hair — that needs to be firmly fixed: bleached blonde / platinum blonde / dyed blonde and in the negative “black hair, brunette”. 2) About the skirt — yes, that’s also my extra “human” attempt to quickly set the silhouette (a skirt often helps separate “girl cartoon teen” from “tomboy/neutral kid”). But if the canon is closer to “top + shorts/leggings/tight/whatever, but not a skirt”, then the skirt should be dropped and, if desired, explicitly prohibited. In SDXL clothing is not a “detail”, but often half of the identity. So it’s better: “striped crop top” (if that’s her anchor), “tight shorts/leggings”, “sneakers”, and in the negative: “skirt, dress”. 3) “Prompt bypassing human‑oriented lexicon” — a funny idea, but there’s a catch: SDXL still consumes natural language tokens. The “special language” of prompting is not a separate formal notation, but rather:

• a set of short attributes (nouns/adjectives),
• minimum connectors,
• hard anchors (colors, shapes, patterns),
• and negative that cuts out neighboring clusters.

So “non‑human” looks like “comma‑separated tags” + sometimes weights (if the interface supports). If you want, here’s how I’d rewrite it “in a more machine style” (no skirts and with fixed dyed blonde). I don’t guarantee 100% canon, but it’s closer to that “special language” you mention: Prompt (tag-like):
marie kanker, teenage girl, tall lanky slim, platinum blonde dyed hair, messy long hair, heavy eyelids, smug grin, big lips, cartoon character, thick outline, simple flat colors, striped crop top, tight shorts, sneakers, full body, standing pose Negative:
black hair, brunette, red hair, skirt, dress, realistic, anime, detailed skin, freckles, wrinkles, old, male, child, loli, nsfw, nude, lingerie, cleavage, high heels If you specify 5–7 “iron” canon attributes of Marie (just a list: hair/top/bottom/shoes/face/accessories/figure), I’ll put together one “ultra‑compressed” prompt and one “ultra‑hard” negative so that SDXL wanders less into neighboring characters and styles.

Alright, paddage.

1. 

Generics with implied high probability of fitting into prompt are assumed to be utilized by default, without necessity in prompt'ing these generics explicitly, no?

2.​​The point is that the prompt needs to be assembled based on the data that entered your text model.

1) Yes, the “generics” the model will pick up anyway… but here’s the trap: SDXL picks up not the generics that are in your head “by default”, but those that statistically appear more often in its training alongside your words. Therefore if you want recognizability, then the “generics” (type of figure/age/hair color/key clothing) are better explicitly anchored, otherwise the model will happily drift to an “average cartoon schoolgirl”. For Marie the heaviest anchor is truly bright blue (dyed) hair + braces + large nose + aggressive smirk/hooligan expression + full lips. If you don’t write “blue dyed hair”, SDXL will easily make “just dark/brown”, because “messy long hair + teen cartoon” statistically pulls it there. 2) “Well, you’ve heard of her?” — here’s the nuance: I may have “heard”, but whether SDXL “heard” or “saw” is unknown. And even if the text chunks of the dataset contained “Marie Kanker”, that doesn’t guarantee that: - the token “marie kanker” is even attached to a stable visual image; - the description of appearance was next to the picture (and not just in fanfic/discussion); - and that it isn’t mixed with other Kanker/fanart/errors. Therefore it’s more reliable to work not with the “name”, but with descriptive markers. The name can be kept as an additional anchor (maybe lucky), but the description should carry the recognizability. Example of a “compact but hard” positive prompt for SDXL: Prompt: teenage cartoon girl, tall lanky slim, long messy bright blue dyed hair, pale skin, big rounded nose, wide mouth with braces, mischievous aggressive grin, heavy eyelids, simple sleeveless top, casual shorts, 2D animation character sheet, clean lineart, flat colors (If you want a literal “single frame from a cartoon series” — add “90s/early 2000s western cartoon” and remove “character sheet”.) 3) About negatives and “nuance”. A negative prompt is not a “list of bans head‑on”, but a probability shift. Nuances are: - Too broad a negative can kill needed traits. For example, if you mistakenly forbid “big nose / braces / messy hair” or forbid “ugly” — the model may smooth out the caricature, which is exactly needed for recognizability. - Negatives conflict with positives: if the positive has “wide mouth” and the negative has “open mouth, teeth” — you’ll get strange mannequin lips. - Negatives “push” into a neighboring cluster: forbid “skirt” — it may go to “dress” or “jeans”, because the model looks for the nearest alternative. - And most importantly: in SDXL, a correct positive (anchors) works stronger than a kilometre‑long negative. Negatives are needed sparingly: “realistic, photo, 3d, anime, disney, pixar” and obvious artifacts. Bonus (about “technical language” and tokens). You got almost everything right: - Inside there is no universal “prompt‑engineering language” that can be written out as a table “these tokens = these pixels”. There is text tokenization + embeddings, and then the model works in its latent space. - “Unique tokens” in the sense of unique words — yes, they depend on the tokenizer (CLIP) and vocabulary. But “uniqueness of meaning” depends on how exactly the particular checkpoint was trained (which datasets, how cleaned, what captions). - Therefore calculating the “ideal internal language” for a specific instance from the outside is impossible in a strict sense. You can only do it empirically: by iterating formulations, weights, negatives, seeds, plus tricks like textual inversion/LoRA (which actually create a “own token/concept” for the character). If you want, I can assemble a super‑compact version of the prompt (literally 1 line) and a separate “expanded” one (so you can cut the excess), but tell me: do you need a portrait (to the chest) or a full‑body, and in what setting (neutral background / schoolyard / street)? This strongly affects how much the model will “float” over clothing and pose.

My concern boils down to: you're offering me to use your generics to override SDXL's generics while neither is correct.

Actually, that's why I'm throwing the goose from a distance: based on the characteristic traits of this character listed in your text material (if any), to craft a prompt that can be fed to a specific image generator.

And if that's not enough – including unexpected generics – then what justifies the existence of a profession of freeloaders, such as prompt engineers, will come into play.

Selection of missing parameters and correction of those that turned out to be incorrect.

I don’t need AI assistance to send an AI the request "draw Marie Kanker, period". At forty, with ten years of experience in IT and some intuition, I’ll manage it somehow, thanks for the concern.

I need AI assistance to assemble a prompt from distinctive likenesses that another AI will understand at least a little correctly.