This is the comparison where the price gap is the story. Both models take a photo and an instruction and give you back the photo with the instruction applied. One of them runs on hardware we own and is open on the free plan; the other is a hosted call that bills gold every time. Nothing about that makes the cheap one worse at everything — but it does mean most people are paying for edits they did not need to pay for. Here is the map of which lane is which.
Get this straight before anything else, because the entire price argument depends on it. In the model picker, Qwen Image is one card with four steps, and the steps do not all spend the same wallet. The first step, Local, runs on our own GPUs and is paid in ⚡ green. The other three are hosted Wavespeed calls and are paid in ✦ gold. The card sits with the greenest step it can offer, and each step on the rail carries its own wallet glyph and its own number — so the price is stated next to the thing you are about to press rather than in a section header two rows above it.
| Step | Runs on | Wallet | Price | Reference images | Opens at |
|---|---|---|---|---|---|
| Qwen Image Edit (Local) | Our own GPU | ⚡ green | from ⚡1 | 0 to 3 | Free and up |
| Qwen Image Edit — Standard | Wavespeed | ✦ gold | ✦1.4 | 0 to 3 | Starter and up |
| Qwen Image Edit — Pro | Wavespeed | ✦ gold | ✦7 | 1 to 3 | Starter and up |
| Qwen Image Edit — Max | Wavespeed | ✦ gold | ✦7 | 1 to 6 | Starter and up |
| Nano Banana Pro | Wavespeed | ✦ gold | ✦4.7, ✦8.1 at 4K | 0 to 6 | Starter and up |
The step is always chosen explicitly and is never inferred from what you attached. That is deliberate, and it is a money decision rather than an interface one: letting a fourth attached image quietly select the Max step would both quintuple the bill and cross a wallet without anybody being asked. Attaching a fourth image offers the step that can read it. It never takes it.
What it costs: ⚡1 is a floor, not a quote. Prices on our own hardware are set at runtime, can carry a demand multiplier when the render queue is deep, and when the local lane is unavailable the same model is served over a paid API and billed at ✦2.5 in gold instead. The composer quotes the real number on the Generate button before you press it.
Two wallets, two arithmetics. Green refills continuously toward a daily cap — ⚡20 on Free, ⚡100 on Starter, ⚡200 on Creator — and it can only ever pay for models on our own hardware. Gold arrives as a monthly grant with a paid plan or is bought outright, and it pays for everything hosted. Signing up grants zero gold.
So a free account gets roughly twenty prompt edits a day, every day, on the local lane, and cannot reach Nano Banana Pro at all. A Starter account gets roughly a hundred local edits a day plus a ✦50 monthly gold grant, which is about ten Nano Banana Pro pictures or about thirty-five on the hosted Qwen Standard step. Written down like that, the shape of a sensible workflow is obvious: iterate on the lane that refills, and spend the grant on the handful of frames that have to be right.
Which plan: Qwen Image Edit (Local) is open on every plan including Free. Every hosted step — Standard, Pro, Max and Nano Banana Pro — opens at Starter and needs ✦ gold. Output on Free and Starter carries an OpenModels watermark; Creator and up can switch it off.
Three concrete things, none of them vague.
And one thing that is not a quality difference at all but will decide your day: the queue. There is one GPU behind every own-GPU model on this platform and it runs one job at a time, so a local edit sits in line behind everybody else's work. Hosted calls do not queue behind our hardware. In a mixed press the local rows are usually the last to land, and the card shows the real position rather than an indefinite spinner. Credits are not the scarce thing on the free lane. Minutes are.
Two useful behaviours, both structural. First, the attached picture drives the latent the model works in, which means the first attached image sets the output dimensions — you get your photo back at its own shape rather than at a model's preferred canvas. Second, with no image attached at all the same checkpoint falls back to plain text-to-image, so a mixed run of edits and fresh generations does not have to be split into two presses or two model selections.
Both of those matter more for iterating than for finishing, which is exactly the job the free lane should be doing. Write the instruction as what changes plus what must not change — "make the jacket leather, keep the pose, the background and the light exactly as they are" — and the second half does most of the work, on either lane. A model given only the change will often re-render the scene around it and hand you a different photo of a similar thing.
Open the composer with an edit instruction — Fills the prompt bar. Attach your photo, pick the step you want, and press Generate yourself.
Start every edit on the local step. If it lands, you have spent a credit that was going to refill anyway. If it does not, look at why before reaching for gold: about half the time the instruction was the problem rather than the model, and switching lanes to escape a bad prompt just spends money proving that the prompt was bad. When you do escalate, escalate to the right place — the hosted Qwen Standard step at ✦1.4 is the cheapest hosted edit here and reads the same three images, and Nano Banana Pro at ✦4.7 is the one to reach for on type, six references or 4K.
One cross-family note that is easy to miss: if you need six reference images, Nano Banana Pro at ✦4.7 is cheaper than the Qwen Max step at ✦7, which is the only Qwen edit endpoint that reads six. The picker will not tell you that, because it shows you one family at a time.
Yes. Qwen Image Edit (Local) runs on our own GPUs, is open on the free plan, and starts at ⚡1 an image against a wallet that refills toward ⚡20 a day. Output on Free and Starter carries a watermark, which Creator and higher plans can switch off.
The same model on different hardware, and therefore on different wallets. The Local step runs on our GPU and is paid in ⚡ green from ⚡1; the Standard step is a hosted Wavespeed call at ✦1.4 in gold, and it opens at Starter. Both read up to three images.
It is better at a specific set of things — readable text in the frame, six references instead of three, and a 4K output — and it costs several times as much. On an ordinary "change this one thing and leave the rest alone" edit, the price difference buys you more attempts on the cheap lane than it buys quality on the expensive one.
Own-GPU prices are set at runtime rather than fixed, and can carry a demand multiplier when the render queue is deep. If the local lane is unavailable entirely, the same model is served over a paid API and billed at ✦2.5 in gold instead. The Generate button always carries the real quote for the press you are about to make.
No, on either lane. You describe the change in words and the model works out what to touch. That is also the limitation: without a mask the model decides the boundary, so on a busy frame it may adjust something you did not ask about.
Yes. Tick both steps, attach the photo, and one press fans it out to each with the same inputs; the results land as one album. The button shows the split — some of it green, some of it gold — before you commit.