A new inference technique lets diffusion and flow matching models skip a large chunk of their generation steps — no retraining required.
Researchers behind arXiv:2607.06114 formalize a property they call "endpoint decodability": during sampling, the intermediate state and its path velocity already contain enough information to estimate the final clean image. They turn this observation into Truncated Jump Sampling (TJS), which simply exits the ODE solver early and decodes the clean sample directly. No distillation, no architecture changes, no modified training trajectories. Tested across SDXL, SD3.5M, Z-Image-Turbo, and three class-conditional benchmarks, TJS reduced neural function evaluations (NFEs) — the main computational cost of diffusion inference — by 20 to 70 percent while keeping output quality close to the baseline.
Most diffusion speedup research comes with a catch: you have to retrain or distill the model, which costs time and compute and means released checkpoints can't benefit without extra work. TJS sidesteps that entirely, making it immediately applicable to any compatible released model. That's a practical gap most acceleration research doesn't close.
The caveat is that "near-matched quality" is doing real work in that sentence — the paper doesn't claim identical output, and how much degradation is acceptable will depend on the use case. Still, for teams running inference at scale on existing checkpoints, shaving 20-70 percent off NFEs without touching the model weights is the kind of result that tends to get quietly integrated into production pipelines before it shows up in a press release.