Kuiz Suntingan Tune-A-Video One-Shot

Jawapan dan penjelasan muncul semasa anda mengambil kuiz.

1. Apakah model asas yang Tune-A-Video bermula sebelum menyesuaikannya untuk video?
   • Model resapan teks-ke-imej terlatih seperti Resapan Stabil
   • Model bahasa yang besar
   • Model video yang dilatih dari awal pada jutaan klip
   • Penjana imej berasaskan GAN
2. Apakah yang dimaksudkan dengan 'one-shot' dalam Tune-A-Video?
   • Menjana dalam satu langkah denoising
   • Menala hanya satu sumber video
   • Menggunakan satu gesaan untuk keseluruhan set data
   • Berjalan pada satu GPU
3. How does Tune-A-Video’s sparse causal attention connect video frames?
   • Frames are generated fully independently
   • Every frame is averaged into one image
   • Each frame can attend to the first frame and the preceding frame
   • Attention is disabled across all frames
4. What does DDIM inversion provide before Tune-A-Video samples an edit?
   • A compressed audio track
   • A new video caption
   • A camera calibration file
   • A source-conditioned latent that guides generation from the input video
5. Semasa penalaan, apakah yang paling utama dikemas kini oleh Tune-A-Video untuk kekal cekap?
   • Satu set kecil perhatian/parameter temporal
   • Setiap berat dalam model penuh
   • Hanya perbendaharaan kata pengekod teks
   • Penyahkod imej sahaja
6. What limitation motivated later tuning-free video-editing methods?
   • It can only edit audio
   • Tune-A-Video requires per-video tuning before editing a new clip
   • It supports no text prompt
   • It was trained only on black-and-white images
7. Which editing capability did the Tune-A-Video paper demonstrate?
   • Automatic speech-quality scoring
   • 3D camera-pose recovery from photos
   • Text-guided style transfer on a source video
   • Instrument-stem separation
8. Which content change can Tune-A-Video apply to a source clip using text?
   • Measuring packet latency
   • Estimating pitch from a monophonic signal
   • Recovering 3D scene geometry from uncalibrated photos
   • Changing a subject or background while trying to retain source motion