Models & Parameters

Autoregressive Models: Generating One Token at a Time

Also called ar model, autoregressive, next token prediction, autoregressive image model

An autoregressive model produces its output one element at a time, each new element conditioned on everything already emitted. Language models work this way, and so does a growing set of image and video models that predict picture tokens or frames in sequence rather than denoising a whole canvas in parallel.

What sequential generation means in practice

Take a picture and split it into a grid of discrete codes using a learned tokenizer, the same way a sentence is split into word pieces. Now you can generate an image the way you generate a paragraph: predict code 1, then predict code 2 given code 1, and so on to the end, then decode the whole sequence back into pixels.

That is the shape of an autoregressive model applied to visual output. Nothing about it is new; the earliest neural image models worked this way, lost to GANs and then to diffusion on quality, and came back once the language model stack got good enough to carry the sequence.

The return matters because the two families fail in opposite directions, and knowing which one you are talking to changes how you prompt it.

What sequential order buys you

Instructions are followed, not approximated. A prompt with six constraints is just more context, and context is what this architecture is built to condition on. Constraints that diffusion models tend to average into a vibe ("exactly three figures", "the text on the left, the logo on the right") survive far more often.

Text and symbols come out right. Spelling, numbers, dates, code on a screen, labels on a diagram. This is the single most reliable practical difference, and it is worth reaching for even when you prefer another model's look.

Editing is conversational. Because the model consumes images and words in the same token stream, you can hand it a picture and a change request, then another change request, and it keeps track of what stayed the same. That is how character consistency across a series of images is done without training anything.

Length is not fixed. For video, generating frame by frame conditioned on the past means there is no architectural clip limit, and the frames can be streamed as they are made. Real time and interactive video systems are built on this property.

What it costs you

  • Latency scales with output. No step count to lower, no fast sampler to swap in. Bigger output means proportionally more sequential work.
  • Quantization softens texture. Going through a discrete codebook throws away micro-detail. Output often reads slightly cleaner and flatter than a diffusion render, which is fine for graphic work and noticeable on skin and film grain.
  • Errors accumulate. Nothing revisits an earlier decision. In video this shows up as slow drift in colour, identity and geometry over a long take.
  • Less structural control. The adapter ecosystem for depth maps, pose skeletons and edge guidance grew up around diffusion backbones. With a sequential model you steer with words and reference images instead.
  • No familiar guidance dial. There is no classifier-free guidance scale to push. Sampling temperature and reference images are the controls, so advice about raising or lowering guidance does not transfer.

Choosing between the two families

The practical split, once you have used both for a while:

Reach for a sequential model when the deliverable has requirements: exact wording in frame, a specific number of objects, a described layout, an edit that must preserve everything else, or a character who has to look the same in the next twelve images. Reach for a diffusion family when the deliverable has a look: photographic texture, a particular film stock, a style adapter you already own, structural control from a depth map, or a batch of twenty cheap variations to pick from.

The line between the families is also blurring. Several current video systems are diffusion backbones distilled into a causal, frame by frame sampler so they can stream, which means they behave like an autoregressive model at inference while carrying diffusion quality. When comparing models on a catalog, judge the behaviour that matters to you rather than the label on the architecture.

The prompt for this

A starting point that reliably produces the effect. Adjust the subject and setting; keep the technical clauses.

A vintage travel poster for Reykjavik with the headline VISIT ICELAND in bold sans serif, three colour screenprint, correct spelling

Try Autoregressive Model yourself

Open the generator with a starting point already filled in.

Frequently asked questions

What is the difference between an autoregressive model and a diffusion model?
Order of construction. A diffusion model starts from noise covering the whole canvas and refines all of it in parallel over a fixed number of steps. An autoregressive model commits to one piece of the output at a time and never revisits it. That makes diffusion better at texture and variation, and sequential models better at following instructions exactly.
Why do sequential image models spell words correctly so much more often?
Because letters are handled the way text is handled: as tokens in a sequence, where each one is predicted with the previous ones fully visible. The model is running the same machinery that already learned spelling from trillions of words, rather than treating letterforms as a texture to be denoised into place.
Why is the output slower, and why does it appear progressively?
Sequential generation cannot be parallelised across the output. Every token waits for the one before it, so latency scales with how much you are producing. The top to bottom reveal you see in some products is not a loading animation, it is the actual generation order.
Can an autoregressive model generate long video?
This is its structural advantage. Because frames are produced in order conditioned on the past, there is no fixed clip length: you can keep going, and you can stream the result while it is still being made. The cost is drift. Small errors compound, so quality and identity degrade the further you go from the start.
Is a large language model an autoregressive model?
Yes, and that is the origin of the term in current usage. The same next token objective that trains a chat model is what these image and video systems adopt, which is why they inherit instruction following and world knowledge along with the sequential latency.

Related terms