All articles

Explainers

SDXL Turbo Explained: AI Images in a Few Steps

How adversarial diffusion distillation lets SDXL Turbo produce an image in one to four denoising steps, and what you trade for that speed.

By · · Updated · 7 min read

For most of the history of diffusion models, quality and speed were the same dial. A good picture meant running the model for 20, 30, or 50 rounds of noise removal. (Each round is called a denoising step.) Every round you cut made the result visibly worse. SDXL Turbo broke that link. It is Stability AI’s SDXL model, retrained so that one to four steps produce an image that older models needed dozens of steps to reach.

This article covers three things. First, what SDXL Turbo is. Second, how the training trick behind it works, in plain terms. That trick is called adversarial diffusion distillation. Third, what you give up in exchange for a 10x cut in compute. That last trade matters far more on a phone than it ever did on a server.

Where SDXL Turbo comes from

The family tree is short. Stable Diffusion XL (SDXL) is Stability AI’s larger follow-up to the original Stable Diffusion. It has a bigger main network for stripping noise out of an image (the denoising backbone). It uses two networks to read your prompt instead of one (they are called text encoders). And it makes 1024×1024 images natively. Run the ordinary way, SDXL needs tens of steps per image. A server GPU finishes a step in a few tens of milliseconds, so that is fine there. Anywhere slower, it hurts.

SDXL Turbo arrived in late 2023. It is the same kind of network with the same job: look at a noisy image and predict the noise in it. What changed is the training. A second phase squeezes the whole multi-step process into a handful of passes. That phase is called adversarial diffusion distillation, and it stitches together two older ideas from machine learning.

Adversarial diffusion distillation, in accessible terms

Distillation is the teacher-and-student idea. The teacher here is regular SDXL running many steps: slow, but high-quality. The student is a second model, trained to jump straight to the answer the teacher would eventually reach. The student sees a noisy image and makes its best single guess at the finished picture. Then it gets corrected toward what the teacher produced from the same input. Over millions of examples, the student learns to cover in one leap the path the teacher walks in dozens of small steps.

Distillation on its own tends to produce soft, averaged-looking images. A student that has to sum up many possible outcomes in one jump will hedge. That is where the adversarial half comes in, borrowed from a family of models called GANs. A third network trains alongside the student with one job. It looks at an image and decides whether it is a real photograph-like training image or the student’s output. That judge is called the discriminator. The student is penalized every time the judge can tell the difference. Hedging stops working. To survive, the student has to commit to crisp textures, sharp edges, and plausible detail, because blurriness is exactly what gives a fake away.

So two forces pull on the student at once. The teacher keeps it faithful to the prompt and to what diffusion models know about the world. The judge keeps it sharp. The result is a model that produces a committed, detailed image in a single step, then refines it slightly over two or three more.

Why one to four steps is enough

A conventional diffusion model needs many steps by design. Each step was trained to remove only a thin slice of noise, so the model is always making a small, cautious correction. Force it to jump from pure noise to a finished image in one step and you get mush. It was never trained to make that leap.

SDXL Turbo was trained to make exactly that leap. Its first step is a full attempt at the final image. The remaining steps, if you use them, work more like an editing pass. The model adds a little noise back to its own output and cleans it up again, firming up hands, fabric, and small structures. That is why the quality gap between one step and four is modest. For a closer look at what a single step computes, see our article on denoising steps.

Turbo models run without the usual prompt amplifier. In a normal model, that mechanism strengthens your prompt on every step, and it is what makes negative prompts work. (Its technical name is classifier-free guidance.) With it switched off, negative prompts do nothing at all. Anything you want steered belongs in the positive prompt.

What you trade for the speed

The compression is not free.

  • Resolution. SDXL Turbo was distilled at 512×512, not the 1024×1024 its teacher is known for. Its images are sharp for their size. You are still not getting full-resolution SDXL output in four steps.
  • Diversity. A many-step model wanders. Small differences early in the process branch into very different images, so one prompt yields a wide spread of results across seeds. A distilled model commits early instead. The adversarial training also pulls it toward outputs the judge scores as typical. The spread across seeds is narrower.
  • Fine control. With the prompt amplifier gone, there is no guidance-scale dial and no negative prompt.
  • Ceiling. Give full SDXL unlimited steps and a good sampler and it can still edge out its distilled student on intricate scenes. The student matches the teacher’s typical output, not its best-case output.

On a server billed by the second, these trade-offs buy you throughput. On a phone, they buy you the ability to do it at all.

Why turbo models matter specifically on phones

Nearly all of a generation’s time and energy goes into that main noise-removing network. Double the steps and you double the cost. A server GPU that runs 30 steps in three seconds has little reason to care. A phone GPU that needs one to two seconds per step faces a different arithmetic: on a device like that, 30 steps is close to a minute of sustained, full-power GPU work. That warms the device and drains battery. Four steps turns the same job into seconds.

Heat is the other reason. Phones slow themselves down under sustained load (they throttle). So a long generation gets slower as it goes, and it slows down the next one too. Short bursts of work finish before heat becomes a factor. Cutting the step count is the biggest reason interactive on-device image generation exists. Shrinking the model down to smaller numbers (quantization) trims the download, but turbo distillation is what makes the wait tolerable.

SDXL Turbo in Onhand

SDXL Turbo appears as Detailed Images in Onhand. Its prepared package is about 3 GB. It is one option in the model library, alongside Turbo, Detailed Scenes, and other looks. The Fast, Balanced, and Detailed quality choices use one, two, and four steps for this model.

Detailed Images runs through MLX on the device GPU and supports generation and photo editing. It does not use a user-supplied negative prompt. Generation time depends on device, loading, quality, and temperature; a fixed seconds-per-image estimate is not a compatibility guarantee.

SDXL Turbo is not the only fast SDXL-class model on iPhone: Onhand vs Conjure covers an app built around DreamShaper XL Lightning. How Onhand’s fast models behave with no connection is on the offline AI image generator for iPhone page.

Sources

Primary sources for the claims above, so you can check them rather than take our word for it.