Luma unveils IMM to break the pre-training ceiling of diffusion models

Inductive Moment Matching (IMM) is an open-source pre-training algorithm from Luma designed to fix the biggest headache with diffusion models: agonizingly slow sampling.

Instead of relying on conventional denoising score matching or complex SDE formulations, IMM replaces the standard diffusion engine with a unified moment matching objective. It slashes sampling steps by up to 30x while actually beating standard diffusion models and Flow Matching in output quality.

How it bypasses diffusion bottlenecks

Continuous generative modeling has been stuck on diffusion pipelines since mid-2020. The fundamental problem is inference efficiency: even if you scale model capacity, diffusion inherently demands granular, iterative steps to converge.

Take the standard DDIM sampler, for instance. Each iteration makes a prediction based on the current timestep and linearly interpolates toward the next. That linear behavior strictly caps how expressive each step can be.

IMM changes the math with an intuitive twist: the network processes both the current timestep and the target timestep it wants to jump to. By allowing the network to know its destination ahead of time, each inference step gains far more expressive freedom.

To make this stable, IMM incorporates Maximum Mean Discrepancy (MMD), a moment matching tool developed over 15 years ago. Unlike consistency models, which are notorious for unstable training and hypersensitive hyperparameters, IMM trains reliably with a single objective across architectures.

Benchmark gains and open-source release

Across standard tests, IMM delivers over a tenfold leap in sampling efficiency compared to diffusion models.

On ImageNet 256x256, IMM achieved a 1.99 FID score using 30x fewer sampling steps, beating diffusion baselines at 2.27 and Flow Matching at 2.15. On CIFAR-10 trained from scratch, it hit a state-of-the-art 2-step FID of 1.98.

Luma has open-sourced the entire project. You can find the code and pre-trained checkpoints on GitHub under lumalabs/imm, along with two ArXiv papers detailing the mathematics and the case for inference-first pre-training.

Why it matters

Inference latency and compute costs remain the biggest roadblock for commercial image and video generation. Most teams attempt to patch this after the fact using distillation or aggressive truncation, which usually hurts fidelity.

IMM shows that rethinking pre-training from an inference-first perspective removes the need to trade quality for speed. If you train visual generative models, this is a repository worth cloning today: it proves continuous generation does not have to remain locked to diffusion orthodoxy.