Normalizing Trajectory Models
https://machinelearning.apple.com/research/normalizing-trajectory-models📌 【Apple ML @ NeurIPS】用「正規化流」重寫擴散軌跡，四步生成也能保留精確似然
TL;DR：NTM 把擴散模型每一步換成可逆 normalizing flow，四步生圖還能算精確機率。
擴散模型（diffusion model）能生成極高品質的圖片，代價是要跑很多次小幅度的去噪（denoising）步驟。當步驟數被壓縮到只剩幾步時，現有加速方法大多得靠蒸餾（distillation）、一致性訓練（consistency training）或對抗式目標（adversarial objective），但這些方法都得放棄擴散模型最珍貴的特性：精確的似然（likelihood）計算。Apple 機器學習團隊發表於 NeurIPS 的新論文，提出 Normalizing Trajectory Models（NTM），試圖在「少步生成」與「精確似然」之間兩者兼得。
🤔 擴散模型的假設，在少步時會失效
擴散模型的理論基礎，是把生成過程拆成許多個高斯去噪步驟，每一步都很小、很接近線性。但當生成被壓縮成少數幾個「粗粒度」的轉移步驟時，這個假設就不再成立，原本依賴小步近似所建立的似然框架也隨之瓦解。
🧩 每一步換成一個條件式 normalizing flow
NTM 的做法，是把擴散的每個反向步驟（reverse step）都改建模成一個具表達力的條件式 normalizing flow，並維持精確的似然訓練。架構上，NTM 結合兩個層次：每一步內部用淺層的可逆區塊（invertible block），整條軌跡（trajectory）上則疊加一個深層的平行預測器（parallel predictor），兩者組成一個可端到端（end-to-end）訓練的網路。這個網路可以從零開始訓練，也可以用預訓練好的 flow-matching 模型來初始化。
由於整條軌跡的似然是精確可算的，NTM 因此支援一種自我蒸餾（self-distillation）機制：用模型自身誘導出的 score function 訓練一個輕量化的去噪器（denoiser），就能在四步之內產生高品質的生成結果。
📊 文字生圖任務上，四步打平甚至超越強力基準
論文指出，在文字生圖（text-to-image）的基準測試上，NTM 僅用四個取樣步驟，就能打平或超越強力的圖像生成基準模型，而且是在所有基準方法都放棄的「精確似然」這一點上獨家保留下來。論文摘要並未提供具體的基準名稱或數值分數，僅陳述上述結論。
⚠️ 摘要未揭露的部分
目前公開的素材僅為論文摘要，並未附上實驗章節的具體資料集名稱、訓練成本或消融實驗（ablation）細節，這些留待完整論文公開後再確認。
🎯 對從事生成模型工程的啟示
如果你的應用場景同時需要「快速取樣」與「可計算似然」（例如需要用似然做異常偵測、資料評分或不確定性估計的場景），NTM 提出的「每步可逆 flow 加跨步驟平行預測器」架構，是值得關注的設計方向，尤其是它宣稱可以直接從現成的 flow-matching 預訓練模型初始化，降低了重新訓練的門檻。
🔗 來源
- 標題：Normalizing Trajectory Models
- 作者／機構：Jiatao Gu, Tianrong Chen, Ying Shen, David Berthelot, Shuangfei Zhai, Josh Susskind（Apple Machine Learning）
- 連結：https://machinelearning.apple.com/research/normalizing-trajectory-models
#AppleML #NeurIPS #NormalizingFlow #DiffusionModel #GenerativeAI #TextToImage #LikelihoodModel #ComputerVision #FlowMatching #DeepLearning
原始資料 Apple ML · 收集於 2026-10-08
摘要原文
research area Computer Vision , research area Methods and Algorithms conference NeurIPS content type paper published October 2026 Normalizing Trajectory Models Authors Jiatao Gu†, Tianrong Chen, Ying Shen‡**, David Berthelot, Shuangfei Zhai, Josh Susskind View publication Copy Bibtex Diffusion-based models decompose sampling into many small Gaussian denoising steps, an assumption that breaks down when generation is compressed to a few coarse transitions. Existing few-step methods address this through distillation, consistency training, or adversarial objectives, but sacrifice the likelihood framework in the process. We introduce Normalizing Trajectory Models (NTM), which models each reverse step as an expressive conditional normalizing flow with exact likelihood training. Architecturally, NTM combines shallow invertible blocks within each step with a deep parallel predictor across the trajectory, forming an end-to-end network trainable from scratch or initializable from pretrained flow-matching models. Its exact trajectory likelihood further enables self-distillation: a lightweight denoiser trained on the score function induced by the model itself produces high-quality samples in four steps. On text-to-image benchmarks, NTM matches or outperforms strong image generation baselines in just four sampling steps while uniquely retaining exact likelihood over the generative trajectory. † University of Pennsylvania ‡ UIUC ** Work done while at Apple Related readings and updates. Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation September 16, 2026 research area Methods and Algorithms , research area Speech and Natural Language Processing Discrete flow matching generates text by iteratively transforming noise tokens into coherent language, but may require hundreds of forward passes. Distillation uses the multi-step trajectory to train a student to reproduce the process in a few steps. When the student underperforms, the usual explanation is insufficient capacity. We argue the opposite: the trajectory is the bottleneck, not the student. Each training trajectory is built through a… Read more Normalizing Flows with Iterative Denoising May 6, 2026 research area Computer Vision , research area Methods and Algorithms conference ICML Normalizing Flows (NFs) are a classical family of likelihood-based methods that have received revived attention. Recent efforts such as TARFlow have shown that NFs are capable of achieving promising performance on image modeling tasks, making them viable alternatives to other methods such as diffusion models. In this work, we further advance the state of Normalizing Flow generative models by introducing iterative TARFlow (iTARFlow). Unlike… Read more Discover opportunities in Machine Learning. Our research in machine learning breaks new ground every day. Work with us
由 claude-code/sonnet 自動生成