Ten seconds is long enough to check Slack. That is the real cost of a slow generation: not the wait itself, but the idea you were holding when it started.
Recraft V4.1 Flash is the fastest model on the market at 1.3 seconds from prompt to image. Speed changes the shape of the work. When an image lands about as fast as the next idea arrives, your focus does not break and the session keeps its momentum.
And at a fraction of the cost of V4.1, a long session of variations stays super affordable. Plus, it keeps the Recraft look across photography, portraits, mood and atmosphere, complex scenes, and early concepts, so moving fast doesn't mean settling for less.
You look, you react, you go again.
.webp)
We ran Flash against other fast models on the kind of briefs designers actually send. Notes below are from our creative director.
Flux 1 Schnell returns a clean, evenly lit studio portrait. But it is also the chef you would license from a stock library: pressed whites, toque, centered.
Flash reads like a portrait someone was assigned to take, with uneven light and a face that has some weather in it.
.webp)
.webp)
Grok Imagine styles the shot nicely, loose greens across stone. It just does not count at all. The prompt called for seven carrots exactly, and Grok didn't read that through.
Flash generated exactly seven carrots, which is the whole brief.
Text accuracy is GPT's lane and it holds here. Every sign is legible with the word "Nirvana," which is important. But the signage stacks into a flat wall of type, and the result looks like a poster about a marketplace, making the image look truly "AI-generated" and not like a true photo.
Flash kept the dusk, the haze, and the depth of a crowd you are standing in.
Flux 1 Schnell over-polishes until physics gives out. A flower grows out of the faucet and the vase floats above the table, as if held by magic.
Flash left the surfaces a little imperfect, which is what makes it read as a photograph.
.webp)
.webp)
Flux 2 Klein makes the more saturated, more appetizing burger. Then the headline dissolves into letters that are not words at all.
Flash placed large type confidently over the darker part of the frame, which is the actual craft decision in an ad layout.
Flux 2 Klein builds a warm, golden banquet, and it is a lovely image. It is also a party for twelve.
"Intimate" is the word doing the work in that brief, and Flash read it, closing the composition down to two people and the light between them.
Flux 1 Schnell swaps the high-detail realism for a painted fantasy look, and the lunge collapses into a nose-to-nose standoff. The young elk also arrives with a full set of antlers.
Flash caught the wolf mid-lunge, fangs bared, with motion blur through the trees and a calf that actually reads as young.
.webp)
When an image is headed for publication, the newly added Refine button regenerates it with Recraft V4.1 Pro, our highest-quality model, at 2048×2048. It cleans up small imperfections so the image holds up at a larger format, keeping the same composition, same light, same decisions you already made. Not a new interpretation of your prompt, just cleaned up.
.webp)
.webp)
Refine has two levels. Subtle clears up noise and contrast and stays close to the original. Moderate, the default, rebuilds small details more actively while holding your layout intact. Both work on details from the first generation.
We measured every model the same way: same prompts, same settings, from the moment the request is sent to the moment the image comes back.
The table below shows 95th percentile times, the number that covers almost every request rather than a lucky one. That matters more than an average: what breaks a workflow is the occasional request that takes much longer. Flash leads on it.
| Model | p95, seconds | Price per image, $ |
|---|---|---|
| Recraft V4.1 Flash | 1.8 | 0.007 |
| FLUX.2 klein 9B | 2.1 | 0.0008 |
| Z-Image-Turbo | 2.3 | 0.0013 |
| FLUX.1 schnell | 3.6 | 0.0013 |
| Nano Banana 2 Lite | 5.8 | 0.0336 |
| P-Image | 6.2 | 0.0044 |
| P-Image-Ideogram, very low | 6.8 | 0.0023 |
| P-Image-Ideogram, medium | 7.3 | 0.0075 |
| P-Image-Ideogram, low | 9.0 | 0.0056 |
| P-Image-Ideogram, very low + enhance | 10.1 | 0.0023 |
| Grok Imagine Image | 10.5 | 0.0200 |
| P-Image-Ideogram, medium + enhance | 10.7 | 0.0075 |
| P-Image-Ideogram, low + enhance | 10.8 | 0.0056 |
| Ideogram 3.0 | 13.0 | 0.0600 |
| Krea 2 Medium Turbo | 18.8 | 0.0150 |
| FLUX.2 flash | 20.1 | 0.0053 |
| GPT Image 2, quality=low | 26.4 | 0.006 |
| FLUX.2 turbo | 35.6 | 0.0085 |
| Nano Banana Pro | 41.2 | 0.1380 |
| Ideogram v4 Instant | 62.8 | 0.0080 |
| Ideogram v4 Fast | 63.5 | 0.0111 |
| Seedream 5.0 Lite | 85.1 | 0.0350 (~4MP only) |
| Grok Imagine Image 2.0 | 97.8 | 0.0600 |
Every model ran the same set of 100 prompts, one run each, from US, Kansas, on its default settings. The prompts covered marketing and ads, branding and packaging, UI and information design, photography, illustration, game and concept art, and fine art styles, across a range of aspect ratios and resolutions, ranging from 2 to 217 words.
Most of these models are not sold directly by the labs that built them, so there is no single canonical number to quote. Recraft models were measured through the Recraft API. FLUX.2 flash, FLUX.2 turbo, Ideogram v4 Instant and Ideogram v4 Fast were measured through fal, and every other model through Runware. Results reflect each provider's performance on the test date.
A technical note on default settings: many providers run a prompt enhancer before generation, an extra model call that expands what you typed. It is on by default in many cases and adds time. We left every model as it comes, because that is what you get without changing anything. Where a model offers the enhancer as a setting, we measured it both ways, so you can see what the step costs. Published figures often exclude this step, which is one reason they look faster than ours.
Download time is excluded, so real time to a visible image is roughly 100ms higher across the board.
Measured on 21 September 2026 by the Recraft team.
Separately, the 1.3 seconds we quote for Flash is its median generation time, measured on a larger prompt set.
Flash is live in Recraft Studio. Pick it from the model picker and start exploring. Using the API? Read our docs to learn how to work with it.
Flash goes from prompt to image in about 1.3 seconds with the API. It is the fastest model Recraft has shipped, fast enough that you can react to each image and try again without losing the idea you were working on.
Flash is a separate model, not a mode. You select it from the model picker in Recraft Studio. It returns finished images rather than rough drafts, across photography, portraits, complex scenes, and early concepts.
Refine cleans up small imperfections so an image holds up at a larger format. It keeps your composition, lighting, and layout as they are. Subtle clears noise and contrast and stays close to the original. Moderate, the default, rebuilds small details more actively while keeping your layout intact.
.svg)
