Luma AI's TVM cuts image generation to 4 steps

Generating a detailed, high-resolution AI image usually means making a neural network run dozens, or even a hundred, heavy evaluation passes. Luma AI wants to compress that entire slog into just four steps without any visible drop in quality.

Luma AI introduced Terminal Velocity Matching (TVM), a brand-new single-stage pretraining paradigm engineered for ultra-fast generative inference.

When trained completely from scratch, a TVM text-to-image model requires only 4 neural network calls to match the output quality of standard diffusion models running 50 steps with guidance (100 total network calls). That equates to an immediate 25x speedup during inference.

Scaling past the limits of IMM

Diffusion models and Flow Matching have powered today's top image and video generators, but serving them at scale is painfully expensive. The repetitive denoising process has always been the primary bottleneck.

Earlier, Luma tackled this with Inductive Moment Matching (IMM), an algorithm that tried to bake few-step generation directly into pretraining. But IMM had critical flaws that kept it trapped in the lab: it required a multi-sample training objective and strictly demanded FP16 precision, breaking down under the industry-standard BF16 format.

TVM solves those engineering hurdles. It provides a principled mathematical framework that subsumes Flow Matching as a special case and connects directly to distribution matching, all within a single-stage training setup.

Crucially, Luma confirmed that TVM scales smoothly to 10-billion-plus parameter diffusion transformers (DiTs), turning few-step pretraining into a practical reality for frontier-scale models.

Why it matters

For creators and developers, shrinking the inference pipeline to 4 steps is the key to real-time generation. If this architecture makes its way into production, waiting around for render bars could soon become obsolete, and API compute costs will drop significantly. That said, this is currently an engineering preview, meaning you cannot spin it up in an active app just yet.

For the AI landscape, TVM shifts the efficiency debate from post-training distillation to native pretraining. Instead of training an enormous, slow model and spending weeks distilling it into a faster variant, TVM bakes rapid sampling into the weights from day one. That puts serious architectural pressure on rivals like Runway and Midjourney to rethink their own training pipelines.

We think Luma's smartest decision was ditching delicate training constraints in favor of BF16 compatibility. A brilliant algorithm is useless if it chokes modern cluster infrastructure.

The lingering question, however, is whether 4 steps can truly hold up across edge cases. Compressing 100 steps into 4 works well for clean portraits and landscape shots, but it remains to be seen whether TVM sacrifices subtle spatial consistency when scaled to long-form video or highly convoluted prompts.