引導式測驗 · 中等 等級
部分通路自迴歸影像測驗
測試 Parti 自回歸影像標記、ViT-VQGAN 重建、觀察到的縮放結果、PartiPrompts 評估和提示忠實度限制。
問題 1 的 8
Parti 如何產生影像?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- Parti 如何產生影像?
- 透過多個步驟對隨機雜訊進行去噪
- 透過一次預測一個圖像標記,就像寫一個句子一樣
- 透過將照片拼接在一起
- 透過繪製向量路徑
- Parti 使用什麼整體框架來執行文字到圖像任務?
- 聚類問題
- 強化學習遊戲
- 最近鄰查找
- 序列到序列的翻譯問題
- What did the Parti researchers observe when comparing their model scales up to 20 billion parameters?
- Consistent quality improvements in their reported evaluations
- A proof that every larger image model always answers every prompt correctly
- Identical outputs and quality at every model scale
- An elimination of every counting and spelling error
- 是什麼將影像轉換為 Parti 的離散標記?
- 試算表
- 擴散 U 網
- ViT-VQGAN 分詞器
- 語音辨識器
- What generation dependency creates sequential decoding work in Parti?
- Every pixel must be manually drawn by the user
- The text encoder can only read a grayscale image
- The tokenizer requires an AIS position message
- Each new image-token prediction depends on the previously generated token prefix
- 什麼是 PartiPrompts?
- 音樂數據集
- 超過 1,600 個具有挑戰性的文字提示的基準
- 一張照片編輯應用程式
- GPU 的一種
- In Parti, what turns the generated visual-token sequence into image pixels?
- The PartiPrompts category label alone
- The user's text file extension
- The image tokenizer's decoder
- A fixed list of object names
- Does original Parti's reported scaling improvement establish perfect counting or spelling on every prompt?
- No; the paper also documents prompt-dependent failures and inconsistencies
- Yes; the largest scale guarantees exact counts for every scene
- Yes; a benchmark removes all spelling errors by definition
- Only if the prompt has no text