How AI image generation works, in the ten seconds after you press generate
The wait is not the model thinking. It is a queue, a text encoder, a handful of denoising steps and a file transfer, and we can show you how long each one takes.
Model filegetting started·FreeImgGen Team·Updated ·Tested on Z-Image Turbo
The short answer
AI image generation works in three stages. Your prompt is turned into numbers, a square of random static is created, and the model repeatedly estimates what noise to remove so that the result matches those numbers. A fast model does that in a handful of passes rather than the twenty or thirty older ones needed, which is why the whole round trip here takes about ten seconds rather than a minute. Most of the variation you feel is queueing, not computing.
MethodTimings measured end to end from our own machine across two batches of real generations on 2026-08-11 and 2026-08-12, submit to downloadable file, over the public endpoint. Stage attribution is from how the pipeline is built, not from instrumenting each stage separately, and is marked as an estimate where it is one.
Stage
What is happening
Share of the wait
Queue
Your job waits for a free slot on the GPU
Near zero when quiet, the entire variance when busy
Text encoding
The prompt becomes a sequence of vectors the model can condition on
Well under a second (estimated)
Denoising
A few passes over a 1024x1024 latent, each removing part of the noise
The bulk of it (estimated)
Decode
The latent becomes actual pixels
Under a second (estimated)
Transfer
The file crosses the network to your browser
Depends on your connection, not on us
Measured total
Submit to downloadable result, typical
About 10 seconds
Measured total
Submit to downloadable result, slowest we saw
Just over 14 seconds
It starts with static, not with a blank canvas
The mental model most people have is that the model draws, beginning somewhere and adding detail. It does not.
It begins with a square of random noise and asks a question over and over: given this noise and given what the prompt describes, what would this look like with a little of the noise taken away? Each pass produces a slightly less noisy version. After enough passes there is no noise left, and what remains is an image.
That is the whole trick, and it explains several things people find puzzling. Why you cannot get the same picture twice without the same starting noise. Why "a bit more to the left" is not a thing you can ask for. Why the model has no idea how many apples it drew: it never counted, it converged.
Where the prompt actually enters
Your words never reach the image directly. A text encoder turns them into vectors, and those vectors steer every denoising pass.
This is why word order matters less than people expect and why word choice matters more. It is also why negative phrasing does so little: "no cars" still puts the concept of cars into the conditioning, and there is no subtraction step anywhere to act on the "no". We tested that across forty images rather than assert it.
And it is why a word that names a physical condition beats a word that names an opinion. "In fog" corresponds to a consistent visual pattern in the training data. "Cinematic" is attached to everything, so it pulls nowhere in particular.
What "Turbo" buys and what it costs
The model we run, Z-Image Turbo, is distilled to reach a finished image in a handful of denoising passes instead of the twenty to fifty a standard model wants. That is the single reason the wait is ten seconds and not a minute.
The bill comes due in the last few percent of detail, and it shows up exactly where you would expect: fine texture, small text, anything requiring the model to reconsider a region several times. Our model file has the failure cases side by side with the successes.
For most uses that trade is obviously right. If you are producing one hero image and can wait two minutes, a slower model on your own hardware will beat it. If you are running twenty variations to find the composition, faster wins, and it is not close.
Ad
Why your wait varies and ours does too
Across two batches of real generations, almost everything landed within a second of ten. The outliers were slower by four seconds or more, and they were all queueing behind other traffic rather than working harder.
We run this on one RTX 4090 of our own rather than reselling a cloud API, which is what makes it free with no account and no credit counter. It also means that when a lot of people arrive at once, there is a queue, and we would rather slow you down gently than start charging or start capping.
Editing is the exception to the ten-second rule. The editing model is larger and has to encode your uploaded image before it can start, and our six-edit run averaged just under a minute. That is measured in the editor model file.
The one number that surprises people
A generation costs roughly a hundredth of a penny in electricity. Not a figure we can invoice, and it is the entire reason a site like this can exist without an account system.
When a free tool has no obvious business model, the reasonable question is what you are paying with instead. Here it is page ads and nothing else: no account, so nothing to sell; no uploads retained for training; no watermark; no credits.
The honest caveat is the one above. One card means a queue at peak, and at peak we rate limit gently rather than degrade quality or ask for a card. That is the trade, stated plainly so you can decide whether it suits you.
Watch the ten seconds yourself
Free, no account. The progress bar you see is the denoising passes finishing, then the file coming down.
A text encoder turns your prompt into vectors. The model starts from random noise and repeatedly estimates what noise to remove so the result matches those vectors. After a handful of passes the noise is gone and an image remains. There is no drawing step and no counting step anywhere in it.
Why does it take about ten seconds?
Most of that is the denoising passes over a 1024x1024 latent, plus a decode into pixels and the file transfer. Our measured typical is around ten seconds end to end, with the slowest we saw just over fourteen, and that outlier was queueing rather than computing.
Why can I never get the same image twice?
Because each run starts from a different square of random noise, and that starting point determines the result as much as the prompt does. Tools that let you fix a seed are fixing exactly that starting noise. Ours does not expose one.
Why do negative prompts not work?
The prompt becomes conditioning vectors and there is no subtraction anywhere in the pipeline. Writing "no cars" still puts cars into the conditioning. We ran twenty pairs to check rather than take it on faith, and the instruction was ignored more often than not.
Does editing take the same time as generating?
No, it is roughly six times slower. The editing model is larger and has to encode your uploaded image before it can begin. Our six edits ranged from about twenty seconds to about eighty-eight, averaging just under a minute.