Apple's NTM hits strong image quality in four sampling steps

Diffusion-based models break generation into many small Gaussian denoising steps, an assumption that weakens when generation is compressed to a few coarse transitions. Existing few-step approaches, such as distillation, consistency training, and adversarial objectives, address this compression but give up the likelihood framework in the process.

NTM keeps that likelihood framework by modeling each reverse step as an expressive conditional normalizing flow with exact likelihood training. Architecturally, it pairs shallow invertible blocks within each step with a deep parallel predictor across the trajectory, so the whole network can be trained end-to-end from scratch or initialized from pretrained flow-matching models.

The exact trajectory likelihood also opens the door to self-distillation, where a lightweight denoiser is trained on the score function induced by the model itself. According to the authors, this produces high-quality samples in four steps, and on text-to-image benchmarks NTM matches or outperforms strong image generation baselines at that step count.