Apple Machine Learning Research has introduced Normalizing Trajectory Models (NTM), a novel approach to generative modeling that maintains exact likelihood training even with a few coarse sampling steps. Unlike previous methods that often sacrifice likelihood for speed, NTM utilizes expressive conditional normalizing flows within each reverse step. This architecture combines shallow invertible blocks with a deep parallel predictor, allowing it to be trained from scratch or initialized from pre-trained flow-matching models. NTM has demonstrated competitive performance on text-to-image benchmarks, achieving high-quality samples in just four steps while preserving exact likelihood over the generative trajectory. AI
IMPACT This research could lead to more efficient and faster AI image generation models by reducing the number of sampling steps required.
RANK_REASON The cluster describes a new research paper detailing a novel modeling technique from a major tech company's research division. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Apple Machine Learning Research →
- Apple Inc.
- David Berthelot
- International Conference on Machine Learning
- Jiatao Gu
- Josh Susskind
- Normalizing Flows with Iterative Denoising
- Normalizing Trajectory Models
- Shuangfei Zhai
- Tianrong Chen
- Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation
- University of Illinois Urbana-Champaign
- University of Pennsylvania
- Ying Shen
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
