距離上一個POST都隔咗大半年，期間生圖AI的變化快過火箭幾半倍…
凡其是Google的Gemini(NanoBananaPro)更是變態，生成的圖像可以直接拿出來見人了！
果然想完美生圖也是要搵大MODEL才可以，和SDXL(Illustratious/NoobAI)跟本不可比！
特別是在可以在圖裡面出文字（凡其是英語以外的文字），那樣做Infograhpic之類有字的圖變得非常有利。
例如這個：
参照画像のキャラクターを使って、アニソンDJの初心者が分かりやすい、DJマルチプレーヤーの操作方法について、インフォグラフィックで作成してください。 キャラクターは真剣にDJマルチプレーヤーを操作しています。或者這個：
Use inforgraphic and timeline to explain Los Angeles Dodgers ownership dispute in 2009-2012除此之外就是可以生成多個角色然後依提示詞給予不同的表情和動作。角色可以用生成或匯入參考圖片的做法，所以對SDXL順手的話可以先用SDXL生成後再匯入參考圖片。
例如這個：
A Japanese Anime Illustration:
Characters:
- Attachment A (Black hair Black serafuku) as Character A.
- Attachment B (Silver Hair Maid) as Character B.
- Attachment C (Black hair White sailor dress) as Character C.
- Attachment D (Orange hair white techwear coat) as Character D.
Situation:
- Character A rise her hand up, holding smartphone gimbal, doing group selfies.
- All characters doing v sign. Smiling at Character A's smarphone camera.
- All characters standing at [35.02072975402742, 138.8974273304328].當中Attachment A, B, C, D就是上面4個角色在SDXL(Illustrious)生成過的立繪。
或者這個：
All the characters (as the attachment) are preparing for the christmas party.
Personaility in one sentence:
Attachment A: have sense of justice
Attachment B: polite, and loyalty to master
Attachment C: gentle but sensitive at sex
Attachment D: super optimistic, super energetic, super happy
Attachment E: Very wary of people especially strangers, but engaged and competitive when playing games
Clothing:
A, C: santa costume, keep original gloves
D, E: reindeer costume, brown mittens
B: add santa hat only
Location:
A dining area of a room inside a castle, looks majestic and noble. Add some christmas decoration and chrtistmas tree. Christmas food are on a round table with tablecloth. Outside window looking night and snowly. Living in Middle Ages.連改變服裝也可以！
網上也有比本人範例更誇張的例子。
NanoBananaPro的容錯率很細，只要輸入的提示詞用語正確的話，就可以準確地生成可立即成為「成品(production)」的圖像。不需要像其它模型一般需要反覆生成以測試除錯，十次以內的生成已經可以達「想像中」滿意的結果。
下一篇就不如說平時用什麼平台的NanoBananaPro生成的…… (在香港不能直接用Google挑戰更大)
發佈留言