Apple's NTM Keeps Exact Likelihood in Few-Step Text-to-Image Generation

Apple researchers introduced Normalizing Trajectory Models (NTM), a generative approach that models each reverse step of a diffusion-style trajectory as a conditional normalizing flow, trained with exact likelihood. The method targets the common failure of diffusion models when generation is compressed to a few coarse transitions, where the Gaussian denoising assumption breaks down.

NTM combines shallow invertible blocks within each step with a deep parallel predictor across the trajectory, forming an end-to-end network. It can be trained from scratch or initialized from pretrained flow-matching models. The exact trajectory likelihood 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. The paper was published in October 2026 and presented at NeurIPS.

What it means for inference engineers

For teams serving text-to-image models, the practical takeaway is that NTM offers a path to fewer sampling steps without abandoning the likelihood framework. Fewer steps directly cut the number of forward passes per generation, which lowers latency and compute cost. The ability to initialize from pretrained flow-matching models also means existing checkpoints could be adapted rather than trained from scratch.

Self-distillation is another lever: the lightweight denoiser trained on the model's own score function reduces the per-step cost, making four-step generation more feasible on constrained hardware. Because NTM retains exact likelihood, it preserves a principled way to evaluate and compare generative models, which is often lost in distillation or adversarial approaches.

Context in Apple's research line

The paper sits alongside related Apple work on few-step generation. A September 2026 paper argued that in discrete flow matching, the training trajectory is the bottleneck, not the student model, and proposed energy-navigated distillation. Earlier, in May 2026, Apple advanced normalizing flows with iterative TARFlow (iTARFlow) for image modeling. NTM extends that line by making normalizing flows work within a trajectory model, rather than as a standalone generative family.

The authors are Jiatao Gu, Tianrong Chen, Ying Shen, David Berthelot, Shuangfei Zhai, and Josh Susskind, with affiliations at Apple, University of Pennsylvania, and UIUC. The work was done while Shen was at Apple.