The Pulse
Apple’s NTM Generates Images in Four Steps With Exact Likelihood
Apple researchers introduce Normalizing Trajectory Models, a generative method that models each denoising transition as a conditional normalizing flow. Their reported text-to-image results reach four-step generation while retaining exact li
AI.info Team · Written with AI ·
Apple researchers are targeting a weakness in fast image generation: diffusion models can produce images in fewer steps, but compressing many small denoising moves into a handful of large ones makes the usual Gaussian-transition assumption less accurate. Their proposed alternative, Normalizing Trajectory Models (NTM), models each reverse step with a conditional normalizing flow, aiming to keep the speed of few-step generation without giving up an exact likelihood for the modeled trajectory.
The paper was submitted to arXiv on May 8, 2026, and Apple’s research page lists it as published in October. The authors report that NTM generates text-prompted images in four sampling steps, with results that match or outperform selected image-generation baselines on the benchmarks they tested. That is a research result, not evidence of a released Apple product or a demonstrated speed advantage in consumer applications.
Why four large denoising moves pose a problem
Standard diffusion sampling gradually turns noise into an image through a sequence of denoising steps. Each reverse transition is commonly represented as a Gaussian distribution. The paper argues that this approximation becomes less suitable when a sampler uses fewer, larger transitions: the true reverse distribution can be multimodal, while a single Gaussian has limited ability to represent those alternatives.
NTM changes the transition model rather than simply asking a conventional sampler to take longer jumps. An invertible “transporter” maps images into a latent representation, where a predictor models the transition with a Gaussian distribution; the invertible mapping makes it possible to calculate exact likelihoods using the change-of-variables formula. The model can be trained from scratch or initialized from a pretrained flow-matching model, according to the paper.
“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.”
Jiatao Gu et al., “Normalizing Trajectory Models”
Four steps, with different results by test
For text-to-image generation trained from scratch at 256-by-256 resolution, the researchers report a 0.82 score on GenEval and 79.64 on DPG-Bench using four steps. They compare that GenEval result with 0.56 for STARFlow, a previous normalizing-flow model that the paper says uses 256 autoregressive steps. The paper also reports a fine-tuned 512-by-512 version scoring 0.76 on GenEval and 83.38 on DPG-Bench.
A separate, class-conditional ImageNet test measures a different task. There, NTM reports a 3.83 FID score at four steps and 2.80 at 16 steps; the paper compares those results with STARFlow’s 2.67 FID using 256 autoregressive steps. The results suggest a trade-off: the NTM four-step run is far shorter by step count, while its FID is higher than that STARFlow comparison. These benchmark figures do not by themselves establish equal runtime, compute budgets or image quality across every use case.
The model’s limit is the one-step case
NTM does not keep improving simply by cutting the trajectory further. The authors report that one-step generation produces severely degraded images: with only one transition, the shallow transporter must carry too much of the distribution’s complexity, and the predictor has no multi-step trajectory to work across. They identify four to eight steps as the model’s effective range, rather than claiming that NTM eliminates iterative generation.
The paper also describes a learned denoiser that distills trajectory-based refinement into a single forward pass. In the fine-tuned setting, the authors report 1.88 images per second for the predictor-plus-denoiser approach, compared with 0.20 images per second for the full normalizing-flow and trajectory-denoising method, alongside an LPIPS score of 0.121. Those figures describe the paper’s tested setup; they do not establish performance on Apple devices or in a shipping image-generation system.
Exact likelihood comes with a quality trade-off
NTM’s contribution is not simply fewer steps. Its central claim is that an expressive, invertible transition can preserve exact trajectory likelihood while supporting four-step image generation, a combination the authors say prior few-step methods did not provide. The study’s own results also define the limits: the strongest reported text-to-image numbers vary by resolution and training approach, some position and attribute-binding tasks remain difficult, and one-step output degrades sharply.
For now, the paper establishes a benchmarked research method, not a general replacement for diffusion sampling. Its clearest boundary is the one reported by its authors: NTM works best across four to eight steps, while the one-step version produces severely degraded images.