Image
Image generation models ranked by human preference, priced per generated image.
Updated · Sources: LMArena, LiteLLM, models.dev
Tiers are ours: a model's place within the board counting only the models whose whole confidence interval is above it. S — top 10%, A — up to 30%, B — up to 60%, C — up to 85%, D — the rest.
What is ranked here
This leaderboard covers models that generate images from a text prompt. It is based on the LMArena text-to-image arena: a voter writes a prompt, receives two pictures from two anonymous models and picks the one they prefer. Many such votes produce an Elo rating with a 95% confidence interval for each model.
Prices per image
Image models are sold in different units. When a provider charges per generated image, that is the price you see; otherwise the table falls back to the price per megapixel or per million tokens, and the unit is written in the column header. The price is the lowest current offer we found in models.dev or LiteLLM.
The "Low price" chip marks models under 0.02 USD per image. Real costs depend on the settings you choose: a larger size or a higher quality level usually costs more than the base price shown.
Reading the tiers
Tiers work as on every leaderboard here. A model needs at least 300 votes; its place is determined by how many models have their entire confidence interval above its own, and the share of the board gives the letter — S under 10%, A under 30%, B under 60%, C under 85%, D for the rest. Several models can share S when their intervals overlap with the leader's.
What the votes do and do not say
- They rate how the picture looks to people, on the prompts people sent.
- They do not rate how well a model follows long, detailed instructions, or how consistent it is across a series.
- They say nothing about the licence terms of the output: check the provider's terms before commercial use.
Use the chart of quality against price to find models that sit high on the rating without costing the most, and tick up to 4 of them to compare the details side by side.