引導式測驗 · 中等 等級

AI 影像中的字元一致性測驗

測試參考影像、字元 LoRA 和身分編碼器如何保持字元一致,以及每種方法失敗的具體方式。

問題 1 的 8

Why does reusing the same seed not guarantee the same character once the prompt changes?

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。

  1. Why does reusing the same seed not guarantee the same character once the prompt changes?
    • Seeds expire after one use
    • Changing the prompt changes the denoising path, so the output changes even from the same starting noise
    • Seeds only control the image resolution
    • The model randomly reassigns seeds between runs
  2. Which technique trains small weight updates on a set of images and ties them to a trigger token?
    • Identity encoder
    • Reference-image conditioning
    • Character LoRA
    • Inpainting
  3. What is the main limitation of face identity encoders like InstantID?
    • They preserve the face but not hair, clothing or body shape
    • They require hundreds of training images
    • They only work on cartoon characters
    • They cannot change the background
  4. How does IP-Adapter inject a reference image into generation?
    • By replacing the text prompt entirely
    • By retraining the full model on the image
    • By copying reference pixels into the output
    • Through decoupled cross-attention with separate key and value projections for image features
  5. A character LoRA keeps producing the same pose and background as its training images. What is the likely cause?
    • 欠擬合
    • 過擬合
    • A low reference weight
    • A missing seed
  6. What does DreamBooth's prior-preservation loss prevent?
    • The class word (like 'dog') collapsing into your specific character
    • The model from generating backgrounds
    • Face embeddings from being computed
    • Two LoRAs from loading at once
  7. When captioning LoRA training images, what should the captions describe?
    • Only the character's fixed traits
    • Nothing, captions should be left blank
    • Only the trigger word repeated
    • The things that should vary, such as background, pose and lighting
  8. Why do identity encoders often struggle with stylized cartoon characters?
    • Cartoons have too many colors
    • Cartoon images are too small
    • The face recognizers behind them were trained on real photographs
    • Cartoon characters cannot have seeds