Prompt editing is a different job from generating. You are not asking for a picture, you are asking for one change to a picture that already exists, and the model's real skill is leaving everything else alone. OpenModels runs one edit model on our own GPUs, payable in green credits, and several hosted ones that bill in gold. This is how to drive them and how to tell which is which before you press anything.
An edit model takes your picture as the starting point rather than as inspiration. The attached image drives the latent the model works in, which is also why it usually decides the output size: on the local Qwen graph the first attached image sets the dimensions the result comes back at. Everything you write is applied on top of that starting point, so the picture is the constraint and the prompt is the instruction.
That is why edit prompts read differently from generate prompts. "A woman in a red coat on a bridge" is a generation. "Change the coat to red, keep the same pose, background and lighting" is an edit. The second half of that sentence does most of the work, because a model given only the change will often re-render the scene around it and hand you a different photo of a similar thing.
| Model | Runs on | Price | Reference images | Plan |
|---|---|---|---|---|
| Qwen Image Edit (Local) | Our own GPU | from ⚡1 | 0 to 3 | Free and up |
| Qwen Image Edit | Wavespeed | ✦1.4 | 0 to 3 | Starter and up |
| Qwen Image Edit 2.0 Pro | Wavespeed | ✦7 | 1 to 3 | Starter and up |
| Qwen Image Edit Max | Wavespeed | ✦7 | 1 to 6 | Starter and up |
| Nano Banana Pro | Wavespeed | ✦4.7, ✦8.1 at 4K | 0 to 6 | Starter and up |
| Seedream 4.5 | Wavespeed | ✦1.5 | 0 to 6 | Starter and up |
What it costs: Own-GPU prices start at ⚡1 an image. They 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 in ✦ gold instead. Treat ⚡1 as a floor rather than a quote.
The local model is the one to reach for first, and not only because it is cheap. Its graph exposes three image slots, all of them optional — image one normally carries the photo being edited, and images two and three carry whatever the instruction refers to. With no image attached at all it falls back to plain text-to-image on the same checkpoint, so a mixed run of edits and generations does not have to be split into two presses.
Try it on the local edit model — Fills the composer with the instruction and selects Qwen Image Edit (Local). Attach your photo and press Generate yourself.
One habit that pays for itself: keep the successful instruction, not just the picture. A generation card stores the full input set, and Regenerate on it restores the prompt, the attached images and the model — so an edit recipe that worked on one photo is one press away from being applied to the next.
Prompt editing is not masking. If you need one precise region changed and nothing else touched at all, the lightbox Inpaint tool paints a mask over exactly that area and fills it — a different, more surgical job, billed in gold. Prompt editing is the better tool when the change is semantic ("make it winter") and the worse tool when the change is spatial ("this one object, nothing else").
Two edit features on the local model are Ultra-only: the face-replace recipe, and the batch version that applies one face to up to thirty base photos in a single run. Below Ultra those chips are not rendered at all, so if you cannot see them that is the reason.
Which plan: Free and Starter output carries an OpenModels watermark, applied at save time. The switch to turn it off is a Creator entitlement and applies to new results only — it does not go back and re-save older ones.
Yes. Qwen Image Edit (Local) runs on our own GPUs, is open on the free plan, and starts at ⚡1 an image against the ⚡20 a day a free account refills to. The hosted edit models are the ones that need gold credits.
The composer tray holds six. What each model actually consumes differs: the local Qwen edit graph takes up to three, Nano Banana Pro and Seedream take up to six, and Qwen Image Edit 2.0 Pro requires at least one.
Almost always because the instruction named the change without naming the constraint. Add the clause that lists what must not move — pose, framing, background, lighting — and re-run.
The first attached image drives the latent on the local graph, so the output tracks it. Very large inputs are downscaled first, and if you need the result bigger the lightbox Upscale tool takes it to 2K, 4K or 8K in gold.
The same model can be served over a paid API instead, and that run is billed in ✦ gold rather than green. This is why the price is quoted as a floor and the button shows the number before you commit.