先看這個問題為什麼值得解Why this problem is worth solving
情境先行Start with the situation
從一團雜訊走到一張影像，可以繞遠路，也可以學一條較直的路；但較直不必然較好學。 To move from noise to an image, one can take a winding path or learn a straighter one; straighter is not always easier to learn.
只比較最終樣本，會看不見訓練估計誤差、數值求解與步數之間的交換。 Comparing only final samples hides trade-offs among estimation error, numerical integration, and step count.
先備知識Prerequisite
- 熟悉機率分布與梯度的直覺 Intuition for probability distributions and gradients
讀完你能做到After this path
- 用狀態、時間與動力系統描述生成程序 Describe generation using state, time, and dynamics
- 區分訓練目標與取樣器的貢獻 Separate the training objective from the sampler
- 從品質、速度與穩定度三軸讀實驗 Read experiments along quality, speed, and stability axes
閱讀框架Concept frame
- 核心問題Question
- 如果兩種方法最後都能生成資料，路徑怎麼走還重要嗎？ If two methods generate the same endpoint distribution, does the path still matter?
- 對照軸Comparison axis
- 取樣計算預算 Sampling compute budget
- 判讀方式Evidence to inspect
- 並排閱讀三種相對情境，再回到論文檢查消融實驗、反事實測試與適用範圍。Read three relative scenarios side by side, then inspect the paper's ablations, counterfactual tests, and scope statements.
回答後查看解析Answer, then view the explanation
你的預測Make a prediction
將取樣步數大幅減少時，最先該檢查什麼？ When sampling steps are cut sharply, what should be checked first?
看現象，不急著講故事Observe before narrating
三個診斷訊號Three diagnostic signals
這些是待檢查的診斷假說。下一步用條件比較、消融實驗或反事實測試排除其他解釋。These are diagnostic hypotheses to test. Use condition comparisons, ablations, or counterfactual tests to rule out alternatives.
一次看完三種情境Read three scenarios side by side
三種情境對照Three-scenario comparison
你會如何證明速度改進不是靠更多訓練成本換來的？ How would you show that faster sampling was not bought with greater training cost?
對照軸：Comparison axis: 取樣計算預算 Sampling compute budget
關注路徑直度、穩定 solver 與蒸餾誤差。 Focus on path straightness, stable solvers, and distillation error.
品質與 latency 可以用 Pareto front 比較。 Compare quality and latency with a Pareto frontier.
額外步數的邊際收益與 energy cost 成為問題。 Marginal quality gains and energy cost become the question.
從直覺到機制From intuition to mechanism
把機制講清楚Explain the mechanism
生成模型學的是一套把簡單分布搬到資料分布的交通規則。 A generative model learns traffic rules that move a simple distribution toward the data distribution.
輸入是初始雜訊與條件；輸出是生成樣本；目標是讓生成分布貼近資料，同時維持可接受的取樣成本。 Input: initial noise and conditioning. Output: a sample. Objective: match the data distribution while remaining practical to sample.
- 01 選定中間狀態或機率路徑。 Choose intermediate states or a probability path.
- 02 學習 score、velocity 或轉移規則。 Learn a score, velocity, or transition rule.
- 03 以 solver 或離散更新從起點走到終點。 Use a solver or discrete updates to move from start to endpoint.
把 training objective、model parameterization 與 sampler 混成同一個『方法』。 Treating the objective, model parameterization, and sampler as one indivisible method.
讀生成論文時固定一層再比較另一層，並同時看品質、步數與總計算。 When reading generative papers, hold one layer fixed while comparing another, and inspect quality, steps, and total compute together.
不用回看頁面的自我檢查Check your recall without looking back
三個回想題Three recall questions
先口頭回答，再展開解析。Answer aloud before opening each explanation.
01 為何 endpoint 相同不代表方法等價？ Why do equal endpoints not make methods equivalent?
中間路徑會改變學習難度、數值誤差與取樣成本。 Intermediate paths change learning difficulty, numerical error, and sampling cost.
02 什麼 ablation 能隔離 sampler？ What ablation isolates the sampler?
固定 learned dynamics，替換 solver 與 step schedule。 Hold learned dynamics fixed and vary the solver and step schedule.
03 少步數生成至少要報哪三項？ What three dimensions should low-step generation report?
品質、延遲／步數，以及訓練與推論的總成本。 Quality, latency or steps, and total training/inference cost.
把工具帶回你的研究Take the tool back to your research
最後一個任務One final task
選兩篇生成論文，畫出 objective、parameterization、solver 三欄表，再找真正不同的那一欄。 Choose two generative papers, tabulate objective, parameterization, and solver, then identify the actual point of difference.