製図の心得 #4 過剰適合 Overfitting 過度擬合 (日本語、中文 and EN) (Pixiv Fanbox)

simple picture1

今日はオーバーフィッティングについて話しましょう。好きな人もいれば、嫌いな人もいます。まずは「オーバーフィッティング」という言葉の定義を見てみましょう。Wikipediaのページ（日本語、中国語、英語）を見ると、この現象はAIによる画像生成に現れ、AI呪文術士のモデルのコントロールが低下したことを意味します。一般的に、現在の二次元モデルの主流は「Mix」シリーズで、さまざまなモデルを断片化、再構築、階層的に融合させています。モデル自体が言語やテキストの理解に関して初期のトレーニングレベルを超えているため、「汚染」が生じます。CivitaiのAnythingV3の元の作者が、このモデルや現象についての理解を示しているので、確認してみてください。

Today we will talk about overfitting, which some people like and some people dislike. First, let's look at the definition of "overfitting" on the Wikipedia page (available in Japanese, Chinese, and English). We find that this phenomenon appears in AI-generated images, which means that the AI wizard's control over the model is decreasing. This is usually because the current mainstream models in the 2D world are various "Mix" series, which have been constantly broken up, reassembled, and layered. The model's understanding of language and text has already exceeded the level it was originally trained on, resulting in "pollution". You can take a look at the original author of AnythingV3 on the Civitai website to better understand this model and phenomenon.

今天來講過度擬合，有些人喜歡，有些人討厭，先看一下"過度擬合"的定義，維基百科的頁面 (日本語、中文 and EN)，發現這個現象在AI生成圖像中出現，意味著AI咒文術士對模型的控制力下降，通常是因為目前二次元的模型主流是各種 "Mix"系列，經過不同模型不斷的破碎重組和分層融合，模型本身對語言文字本身的理解早已超過了原本剛開始訓練的程度，產生了"汙染"，可以去看看 Civitai 網站中 AnythingV3 的原作者，對此模型和現象的理解。

simple picture2

しかし、これは全て悪いことではありません。"Mix"モデルには、予想よりも望ましくない変化が生じる場合がありますが、想像力に欠けるAIイラスト作家は、このモデルの"脳内補完"能力に頼って"驚異"を生み出し、それをSNS (Twitter、Instagram、2ch、4ch、Reddit) に投稿することがあります。今日は、これが何を引き起こすかを少し探求してみましょう。

However, this is not necessarily a bad thing, it just means that various "Mix" models themselves have produced unexpected changes. Some AI illustration creators who lack imagination rely on this model's "fantasy" ability to create "wonders" and then post them on SNS (Twitter, Instagram, 2ch, 4ch, reddit). Today we will explore what will happen in this case.

但這不全然是壞處，只是各種 "Mix" 模型本身產生了不如預期的變化，有些缺乏想像力的AI插圖作者會倚靠這種模型的 "腦補" 能力來產生 "奇觀"，再張貼到 SNS (Twitter, Instagram, 2ch, 4ch, reddit)，我們今天來稍微探索一下會發生什麼事。

prompt: 🛵

prompt: 😻

prompt: 🎧

prompt: 1girl / masterpiece

DreamStudio ( stability.ai ) prompt: 1girl

ただ、2023年5月7日現在、あるモデルがある程度の過剰適合（AIが細部を補完すること）を持たない場合、ダウンロードされ使用されないことがあります。これは非常に興味深い現象で、魔法使いが魔法の杖（モデルの比喩）を振るい、杖が火球を発射すると同時に、魔法使いが杖（モデル）に花火のようなエフェクトを追加することを要求し、火球にはLEDライトエフェクトを発する必要があることを意味します。そうでないと、それは合格した杖ではないとなります。これは考えさせられる現象でもあります。

But in the current time of May 7, 2023, a model without a certain level of overfitting (AI brain-supplied details) will not be downloaded and used. This is truly an interesting phenomenon, just like a spellcaster wielding a wand (model metaphor), the wand casting a fireball spell, while the spellcaster requires the wand (model) to also add firework-like special effects and have the fireball emit LED light effects, otherwise it would not be a qualified wand. This is also a thought-provoking phenomenon.

但在 2023/05/07 的現在，一個模型沒有相當程度的過度擬合 (AI腦補細節)，就不會被下載使用，這真的是很有趣的現象，就像是咒文術士揮舞著魔杖(模型的比喻)，魔杖施展出火球術的同時，咒文術士要求魔杖(模型)同時要補充出如同煙火般的特效，火球還要發出如同LED光線特效，不然就不是一個合格的魔杖。這也是一個引人反思的現象。

simple picture3

simple picture1 prompt: ((masterpiece, best quality)), (1girl), solo focus, (black hair:1.1),(full body:1.1),best quality, ultra high res, nsfw,large breasts,(abs:0.7), (wide hips:0.7), narrow waist, cleavage, bursting breasts, narrow waist, shiny skin, (wet),(cleavage), sweatdrop, steam, blush, shy, straight-on , lying on side , babydoll , Caramel eyes, Fire Truck Red side ponytail hair front ponytail hair tubes , hand on own thigh , cushion leaf leaf

Negative prompt: EasyNegative, paintings, sketches, (worst quality:1.4), (low quality:1.4), (normal quality:1.4), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, glans,extra fingers,fewer fingers, paintings, sketches, loli, animal ears,

Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 4082830841, Size: 864x1304, Model hash: 0b8d2055c5, Model: darkSushiMixMix_brighter, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Mask blur: 4, Hashes: {"vae": "f921fb3f29", "model": "0b8d2055c5"}, Hashes: {"vae": "f921fb3f29", "model": "0b8d2055c5"}

simple picture2 prompt: (masterpiece:1.3), best quality, perfect anatomy, lustrous skin, (realistic:1.2), (((ultra-detailed:1.2))), (((8k:1.1))), (masterpiece:1.3), high quality, novel illustration, detailed eyes, eye reflection, (1 girl high school student loli girl:1.1), (loli face:1.1), 18 years old girl, black hair, ponytail, (finely detailed eyes and detailed face:1.3), cyberpunk, dazzling blue eyes, (((detailed classroom, window, hills, clouds, sky:1.2))), POSE VARIATION, covered nipples, (HUGE BREASTS:1.1), SMIRK, (DETAILED PUPILS:1.1), (TONED ABS:0.7), (THICK THIGH:0.8), (INTRICATE:1.1), SHARP, looking at viewer.

Negative prompt: EasyNegative, paintings, sketches, (worst quality:1.4), (low quality:1.4), (normal quality:1.4), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, glans,extra fingers,fewer fingers, paintings, sketches, loli, animal ears, watermark, english text,

Steps: 30, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3746727609, Size: 512x768, Model hash: 81829bfe8e, Model: colorful_cornflowerXFeatOffset_xON, Denoising strength: 0.55, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires upscaler: Latent (nearest-exact), Hashes: {"vae": "f921fb3f29", "model": "81829bfe8e"}, Hashes: {"vae": "f921fb3f29", "model": "81829bfe8e"}

simple picture3 prompt: beautiful, masterpiece, best quality, extremely detailed face, perfect lighting, [Dark and mysterious, eerie and ominous, back view, goddess, walking towards the horizon, large breast, narrow waist, wide hip, black shroud, flowing tresses, ominous, powerful, moonlit forest, foggy] night, no humans, original, outdoors, photoshop medium, scenery, sky, solo, spoilers, standing, tree, white hair, zerocastle, zoom layer, yellow eyes, yoshioka yoshiko, zombie land saga, zipper pull tabard, zieru, zieru's zodiac, zourius zuri warship, zurikishi kancolle, zurikishi

Negative prompt: EasyNegative, bad-hands-5, badhandv4, (low quality:1.4), (worst quality:1.4), (monochrome:0.8), lowres, (extra digits), watermark, username, signature, text, error, fewer digits, cropped, jpeg artifacts, blurry, poorly drawn, trembling,

Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2064944203, Size: 512x768, Model hash: 7d94169e92, Model: 0now_kwcTest_kwcv10, Denoising strength: 0.55, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires upscaler: Latent (nearest-exact), Hashes: {"vae": "f921fb3f29", "embed:bad-hands-5": "aa7651be15", "model": "7d94169e92"}, Hashes: {"vae": "f921fb3f29", "embed:bad-hands-5": "aa7651be15", "model": "7d94169e92"}