引導式測驗 · 簡單 等級
AI 影像產生測驗中的種子和可重複性
測試種子如何設定起始雜訊、哪些設定必須匹配才能重新建立影像,以及為什麼硬體或模型變更會破壞精確再現。
問題 1 的 8
What does the seed directly control in a diffusion image generator?
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。
- What does the seed directly control in a diffusion image generator?
- 模型的訓練數據
- The wording of the prompt
- 輸出解析度
- The pseudorandom starting noise
- Why does keeping the seed but changing the resolution produce an unrelated image?
- Seeds expire after one use
- The noise tensor changes shape, so the same numbers form a different pattern
- Changing resolution rewrites the prompt
- The sampler resets to its defaults
- Which aspect of an image does the seed strongly influence because it is decided in early steps?
- The overall composition, such as where a figure stands
- 文件格式
- 文字編碼器
- The image license
- Why can the same seed give different noise on a CPU and a GPU?
- GPUs ignore seeds
- The CPU secretly adds a watermark
- CPU and GPU random generators produce different sequences from the same seed
- Seeds are rounded to the nearest thousand on GPUs
- In a batch of four images starting at seed 500, which seed do many tools assign to the fourth image?
- 500
- 2000
- 504
- 503
- What does the guide say about 'good seeds'?
- Certain seeds always produce better images
- A seed that works for one prompt and model has no special quality elsewhere
- Seed 42 is the best starting point for any model
- Good seeds carry over between Midjourney and Stable Diffusion
- Why do variation features use spherical interpolation (slerp) instead of simple averaging?
- Slerp is faster to compute
- Slerp changes the seed number automatically
- A linear average of Gaussian noise lowers its variance and produces washed-out images
- Slerp adds text to the image
- How do ancestral samplers differ from DDIM in how they use randomness?
- They draw fresh noise at each step, not only at the start
- They use no randomness at all
- They ignore the seed completely
- They use the seed to choose the prompt