Kuis KontrolNet

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. What does ControlNet add to a pretrained text-to-image diffusion model?
   • A spatial conditioning path that can guide structure using inputs such as edges or pose maps
   • A guarantee that every generated pixel matches the input map exactly
   • A replacement for the text encoder and denoiser
   • A tool that predicts the original training image
2. Apa peran lapisan 'konvolusi nol' di ControlNet?
   • Mereka menghapus bobot model asli
   • Mereka memulai dengan bobot nol sehingga pelatihan dimulai dari perilaku model aslinya
   • Mereka mengubah gambar menjadi skala abu-abu
   • Mereka menambahkan suara acak ke prompt
3. Manakah dari berikut ini yang merupakan input pengkondisian ControlNet yang valid?
   • Spreadsheet angka
   • Bentuk gelombang audio
   • Kerangka OpenPose atau peta kedalaman
   • Skema basis data
4. Bagaimana ControlNet berhubungan dengan model difusi dasar seperti Difusi Stabil?
   • Ini sepenuhnya menggantikan model dasar
   • Ini hanya berfungsi tanpa model dasar apa pun
   • Ini berjalan di komputer terpisah melalui internet
   • Itu menempel sebagai cabang yang dapat dilatih sementara model dasarnya tetap dibekukan
5. In a ControlNet workflow, what roles can text and spatial conditioning play?
   • Text can describe semantic/style content while a spatial map supplies structural guidance
   • The spatial map proves the text prompt is factually true
   • Text always supplies exact pixel positions and the map supplies the object name
   • The text prompt is ignored whenever a map is present
6. What training-data scale does the ControlNet paper report as robust for training its controls?
   • Only datasets with exactly 50 examples
   • Both small datasets under 50,000 examples and large datasets over 1 million examples
   • Only datasets larger than 100 million examples
   • No training data is used
7. What does the ControlNet paper report about combining conditioning controls?
   • It permits only one edge map and requires no prompt in every case
   • It accepts text prompts but no spatial controls
   • It tests single or multiple conditions, with or without a text prompt
   • It requires exactly two depth maps for each image
8. Which pretrained part does ControlNet reuse as a strong backbone for learning spatial controls?
   • A database of user prompts
   • A language-model tokenizer only
   • A separately trained classifier for every edge map
   • The deep encoding layers of a pretrained diffusion model