Dream

Generation Intermediate Hosted inference

Text-to-image with a fixed seed, so you can change one word in the prompt and see precisely what that word was doing.

Model: SDXL 1.0
Parameters: 3.5B · 2.6B UNet + text encoders
Input: Text
Output: Image
Typical time: 10–40s
Last run: —

01 Prompt

Terms near the beginning carry more weight. Word order matters more than word count — a long prompt is not a more specific one.

Negative prompt

Terms to steer away from. It nudges the guidance; it is not a filter and does not guarantee absence.

02 Style

A style is just extra words appended to your prompt. Open the inspector after generating to see the full text that was sent.

03 Frame and sampling

Aspect ratio

Denoising steps 28

How many times the model subtracts a portion of the noise. Fewer is faster and softer; more rarely helps past about 40.

Prompt adherence 7.5

Low values drift from your prompt but look more natural. High values obey it and start to look overcooked.

Seed locked

The seed determines the field of random noise the image is carved out of. Keep it fixed and every difference you see comes from your prompt. Change it and everything changes at once.

Result Ready

Gallery 0 images

Generated images collect here for this session only. Nothing is uploaded or stored on a server — refresh the page and they are gone.

What is actually happening The model begins with pure random noise determined by the seed. At each step it predicts which part of what it sees is noise, and removes a fraction of it, pulled toward your prompt by classifier-free guidance. After the steps run out, an image is what remains. It never had a plan, and there was never a moment when it decided what to draw.

How it works

It begins with a field of pure random noise. Guided by your prompt, the model repeatedly predicts which part of what it sees is noise and subtracts it. After a few dozen steps an image is left. The seed decides the starting noise, which is why the same seed and prompt reproduce the same picture.

Input Your prompt, encoded by two text encoders

Noise A random field determined entirely by the seed

Model 2.6B-parameter UNet predicts and subtracts noise, repeatedly

Output Whatever is left after the last step

Full identifier: stabilityai/stable-diffusion-xl-base-1.0
Where it runs: Hosted inference API
Parameters: 3.5B

Trained on a large scrape of web images, with the aesthetic and demographic skew that implies. Counting, spelling and precise spatial relationships remain weak.

What to notice

Word order and emphasis matter more than word count. Terms near the start of the prompt carry more weight.
Same seed, one word changed, is the only clean way to see what a word contributes. Everything else changes too much at once.
It has no model of physics or anatomy — it has a model of what pictures of those things tend to look like. Hands and text are the usual tells.

A note on language

This page says “the model predicts”, not “the model knows”. That is deliberate. None of these systems understand the images or sentences you give them; they map inputs to outputs using patterns fixed at training time. The difference matters most exactly when the output is impressive.