Kuis Pencitraan Generatif Bertopeng Muse

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. Apa yang diprediksi Muse selama pembuatan alih-alih menolak piksel?
   • Nilai piksel RGB mentah secara langsung
   • Token gambar diskrit bertopeng dari buku kode VQGAN
   • Goresan vektor seperti program grafik vektor
   • Peta kedalaman yang kemudian diwarnai
2. What feature contributes to Muse's reported sampling efficiency relative to the evaluated sequential image-token baseline?
   • Generating only monochrome outputs
   • Removing all text-encoder computation
   • Guaranteeing completion of every image in a single pass
   • Predicting multiple masked image tokens in each decoding round
3. How does original Muse condition image prediction on its text prompt?
   • With embeddings from a pretrained frozen T5-XXL encoder
   • With an untrained word-frequency table
   • By reading only the first character
   • By turning off text processing during inference
4. Apa peran remasking berbasis kepercayaan diri di Muse?
   • Ini menghapus gambar pelatihan berkualitas rendah
   • Ini memampatkan file gambar akhir
   • Ini menjaga prediksi token tetap percaya diri dan menutupi kembali prediksi yang tidak pasti untuk putaran berikutnya
   • Ini menilai kualitas cepat pengguna
5. How does Muse's two-stage token pipeline produce a higher-resolution image representation?
   • It enlarges the output only with a blur filter
   • It predicts all resolutions with no additional model
   • It discards the text condition before the second stage
   • A super-resolution token model uses the base representation to predict a higher-resolution token grid
6. Jenis tugas apa yang secara alami cocok untuk Muse karena memprediksi wilayah bertopeng?
   • Transkripsi audio
   • Pengecatan dalam dan pengecatan luar
   • Perhitungan lembar bentang
   • Hanya kompresi video
7. After Muse predicts its image-token grid, which component reconstructs pixels?
   • The text prompt's keyboard input handler
   • An AIS transponder
   • The corresponding VQGAN tokenizer decoder
   • The training-caption file name
8. Why does a quoted Muse image-model parameter count omit some parameters used by the full generation system?
   • The frozen T5-XXL text encoder is an additional component
   • Every frozen model has zero stored parameters
   • The model count measures only PNG file bytes
   • The text encoder becomes the image tokenizer at runtime