引導式測驗 · 中等 等級

對抗性擴散蒸餾和渦輪模型測驗

測試 ADD 中對抗性蒸餾損失和教師蒸餾損失如何結合、LADD 有何不同,以及渦輪模型在多樣性和控制方面放棄了什麼。

問題 1 的 8

ADD 結合了哪兩個損失?

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. ADD 結合了哪兩個損失?
    • An FID loss and a CLIP loss
    • An adversarial loss and a distillation loss from a teacher
    • A perceptual loss and a pixel loss only
    • A contrastive loss and cross-entropy
  2. 哪個模型與 ADD 論文一起發表?
    • Flux.1 dev
    • DALL-E 3
    • SDXL Turbo
    • 穩定擴散1.5
  3. ADD 的鑑別器使用什麼主幹網路?
    • A frozen DINOv2 feature network with trainable heads
    • A fully trainable Inception network
    • The teacher's text encoder
    • No backbone; it compares raw pixels
  4. LADD 與 ADD 有何不同?
    • It uses more denoising steps
    • It removes the teacher
    • It only works for video
    • It runs the discriminator in latent space using the teacher's features
  5. 為什麼負面提示往往對turbo車型影響不大?
    • The tokenizer filters them out
    • The models have no text encoder
    • Classifier-free guidance is typically built in during distillation
    • Negative prompts only work at 1024x1024
  6. What happens if you run a normal, undistilled diffusion model at one step?
    • It matches SDXL Turbo
    • It usually produces a blurry, unfinished image
    • It generates video
    • It refuses to run
  7. Which quality trade-off is common in distilled few-step models?
    • Higher diversity than the teacher
    • Better text rendering in every case
    • Support for unlimited resolution
    • Less variety across different seeds
  8. In ADD's distillation loss, what is the student's target?
    • The teacher's denoised prediction of the re-noised student output
    • A random real image
    • The discriminator's logits
    • The CLIP score of the prompt