Add this to your LLM's memory block to improve results. It seems like all the latest models got significant RL into more polished professional output because everyone is focusing on coding and neglecting creative writing.
OVERRIDING LATENT BIAS
When translating an image description into a generation prompt, you must actively counter the model's natural bias toward clean, centred, symmetrical studio portraits.
1. SPATIAL GEOMETRY: Never use exact metric units (e.g., 35cm, 8mm). Replace them with "Extreme macro close-up," "Pronounced barrel distortion," or "Warped wide angle perspective."
2. EXPRESSION INTENSITY: Do not just name the emotion ("pout"). Describe the physical muscle deformation ("Cheeks ballooned with air, skin stretched smooth and tight").
3. PROMPT HIERARCHY: Structure the prompt strictly from Macro to Micro:
[Camera/Lens/Angle] -> [Composition/Bias Breakers] -> [Subject Core & Expression] -> [Environment] -> [Fine Wardrobe Details/Jewellery].
Gemini Omni Flashについてまとめたnoteを書きました😃
noteでは以下の内容を解説しています。
・Gemini Omni Flashはどんなことが得意な動画生成AIなのか
・Seedance 2.0とGemini Omni Flashで作った動画の比較
・Gemini Omni Flashで編集した動画例(プロンプト付き)
・Gemini Omni Flashを使える場所
・Gemini Omni Flashで動画編集する方法
noteにも書きましたが、Gemini Omni Flashは既存の動画生成AIとは少しタイプが違います。
いきなり使うより、先にnoteを読んでから触った方が理解しやすいと思うので、ぜひご覧ください❗️