Text-to-image, compared
Text-to-image tools all promise the same thing and hand back wildly different pictures. We type the same prompts into each one and write down what actually came out — how closely it followed the words, whether the text in the frame is readable, what the hands look like, and what the licence lets you keep.
compare — run one prompt through every text-to-image generator and show me what each one returns
The directory
Each card below is one comparison running across the whole field. Pick the question that decides your purchase — most people only really care about two or three of them.
does it actually draw what I asked for?
The core question. We write prompts with countable objects, specific colours, and awkward spatial relationships, then check whether the picture matches the sentence — and how each generator behaves when you push it with longer, fussier instructions.
object counts · colour terms · spatial words
how far does its style stretch?
Some generators have a signature look they cannot escape; others move from documentary photography to flat vector without complaint. We compare how wide the range really is and how hard you have to work to leave the default aesthetic behind.
photoreal · painterly · vector · 3D
can it render readable text inside the image?
Words on a shopfront, a poster headline, a label on a jar. Legible in-image text is still the sharpest dividing line between generators, so we test short strings, longer strings, and whether the spelling survives a re-roll.
short strings · long strings · fonts
how does it handle hands, faces and bodies?
The classic failure mode. We look at fingers, eyes, teeth, limbs at odd angles, and groups of people, and note how often you need a second attempt to get something you would actually publish.
fingers · eyes · crowds · poses
what sizes and resolutions can I get out?
Whether you can ask for a wide banner or a tall poster natively, what the native pixel dimensions are before upscaling, and how the upscaler treats fine detail when you push an image past its original size.
native ratios · upscaling · file formats
can I fix one part without redoing the whole thing?
Generation is only half the job. We compare inpainting and masked edits, outpainting, seed control, variation strength, and how easy it is to nudge a nearly-right image the last ten percent instead of starting over.
inpaint · outpaint · seeds · variations
what am I allowed to do with the output?
We read the terms so you do not have to: who owns the image, whether the free tier permits commercial work, what the training-data position is, and whether any indemnity attaches to the output.
ownership · commercial rights · attribution
how much do I get before paying?
How generous the free allowance is, how credits are spent by resolution and re-rolls, what happens when you run out, and whether the free tier is a genuine trial or a demo that cannot produce usable work.
allowances · credit maths · queue speed
House rules
Image generation is subjective enough without an invented scoring system on top of it. These are the constraints we work under, written down so you can hold us to them.
Every generator gets the identical prompt text, with no per-tool tuning to flatter a favourite. If a tool needs different phrasing to perform, that is a finding we report, not a correction we quietly make.
Cherry-picking the best of twenty attempts tells you nothing useful. We describe typical output rather than best-case output, and say plainly how many tries it took to get something usable.
You will not find a number out of ten here. Image quality is a judgement call, and a fabricated decimal only disguises that. We say where a tool wins, where it loses, and who it suits.
Some outbound links earn us a commission. They never change a verdict, never buy placement, and never decide which tools we cover. The recommendation reads the same either way.
The run order
Nothing exotic — just the same work done the same way each time, so the comparison holds up when a new model ships and we run it all again.
A fixed set of prompts covering the jobs people actually bring to these tools: a product shot, a character, a poster with words on it, a scene with a specific object count, and a stylised illustration.
The same set goes through each generator at comparable settings, with the same number of attempts, so the differences you see come from the model rather than from us trying harder in one place.
Then we write what changed between the outputs — what each tool got right, where it fell apart, what the editing tools could rescue, and what the licence lets you do with the file afterwards.