WHY THIS SECTION EXISTS
Most of this document covers the production state of the field, the models you can actually use today, how they work, and what they cost. The research frontier is something different. It is the set of techniques that are being published in papers right now and that will shape the production models of 2027 and 2028. Knowing what is at the frontier is not just academic curiosity, it tells you which limitations of today's models are about to be solved, which strategic bets are about to look obvious in hindsight, and which categories of product are about to become possible. This section covers four areas where the frontier is moving fastest in 2026.
A note on where this part sits. Everything up to here has been either the machine or the medium built on it. This closing part looks forward, and it now spans all three pillars. The threads that follow, faster generation, character consistency, long-form video and world models, and the audio frontier, are the research that will define the next edition of this book. Read it as the horizon the whole field is walking toward, not a detour back into the lab.
the distillation and Mixture of Experts chapter introduced distillation as the general technique for making slow models fast by training a student to mimic a teacher. This section covers the current research state, the specific techniques, and the business implications, without retelling the conceptual introduction. If distillation is still a new concept to you, read the distillation and Mixture of Experts chapter first and come back here.
The biggest research story in image generation right now is not about quality, it is about speed. The standard diffusion process requires 20 to 50 denoising steps per image, each of which is a full forward pass through a multi-billion-parameter model. Each generation takes a few seconds even on the best hardware. At scale, those few seconds dominate the cost structure of every generative product. Research labs have been competing for the past two years to compress that 20-to-50-step process down to fewer steps, ideally to one, without losing quality.
The technique that makes this possible is called distillation. The idea is to train a small 'student' model that learns to mimic the outputs of a large 'teacher' diffusion model, but in fewer steps. The student does not need to learn how to generate images from scratch, that knowledge is already in the teacher. The student just needs to learn how to take a noise vector and produce, in one step or four steps, what the teacher would have produced in fifty steps.
There have been several generations of this work, each one closing the gap further. The first wave was Latent Consistency Models (Latent Consistency Model (LCM)) in 2023, which got Stable Diffusion to four-step generation but with visible quality loss. The second wave included InstaFlow, UFOGen, and SDXL-Turbo, all of which targeted one-step generation but produced blurrier results than the multi-step originals. The third wave is Distribution Matching Distillation (DMD), introduced by Tianwei Yin and colleagues at MIT in 2024, which closed almost all of the quality gap, DMD on SD 1.5 reaches within 2.7 Fréchet Inception Distance (FID) of the original model while running roughly 30 times faster, generating images at 20 frames per second on a single Graphics Processing Unit (GPU). DMD2, released later in 2024, removed the regression-loss requirement that made the original DMD expensive to train.
The current state of the art is something like SANA-Sprint, which uses continuous-time consistency distillation to achieve one-step generation at 0.1 seconds per image on an H100 (and 0.31 seconds on a consumer 4090) while matching the quality of the original multi-step model on most benchmarks. This is 10 times faster than typical 28-step inference, with no meaningful quality loss. It is the kind of speedup that changes what is economically possible.
Why this matters for your business
If you are building a generative imagery product, distillation is the most important research trend to track. Here is the implication: the cost per image in 2027 will be roughly 10x lower than in 2026 because of distillation alone, on top of any hardware improvements. If your business model assumes that inference cost is a major component of your unit economics, the cost will fall faster than you expect, and competitors who lock in low costs early will be hard to dislodge later. Conversely, if your business model assumes high inference costs are a barrier to entry that protects you, that barrier is going to crumble.
The other implication is that real-time generation becomes possible. Once you can generate an image in 100 milliseconds, you can build interactive products that were not feasible before, live AI avatars on video calls, real-time AR effects, generative game graphics that respond to player actions. This is already starting to happen with products like fal's Real-Time Application Programming Interface (API) and Krea's real-time canvas. By 2027, the dominant consumer product category in generative imagery may be real-time interactive experiences rather than the current 'type prompt, wait, see image' paradigm.