How to Edit an Image with Wan 2.7 (Prompt-to-Output Workflow)

Wan 2.7 image editing is not Photoshop with a text box. You don't mask a region, open a layer, or match a clone stamp to the surrounding texture. You hand the model your image, describe the change in ordinary language, and it re-renders the whole picture with that edit applied and the rest held as stable as you asked it to. That one mechanic — describe, re-render, keep — is the whole tool, and getting good results is about phrasing the request so the model changes what you want and leaves alone what you don't.

The workflow we use: how to frame the prompt, when to add references, which tier and size to pick, and the failure modes that make most first edits land wrong.

What "editing" actually means here

The model takes your base image and produces a new image conditioned on it. It is re-drawing the frame, not painting pixels into a fixed canvas. Two consequences follow from that, and they explain most surprises:

• You can ask for changes that would be painful by hand — "make it golden hour," "replace the plain wall with a bookshelf," "remove the person on the left" — because the model regenerates the affected area with plausible lighting and detail instead of leaving a hole.
• You can also lose detail you didn't mention. If you don't tell the model what to keep, it feels free to restyle it. The fix for almost every bad edit is the same: name the keep-list.

The base image is optional. With no input image, the same engine does text-to-image generation; with one, it edits. Add more images and it composes from them, covered below.

Step 1: State the change, then name what stays

The single most useful habit is ordering your prompt as change first, keep second. The model reads your sentence roughly in that priority, so lead with the edit and follow with the things that must survive it — pose, framing, the subject's face, the product label, the background you liked.

Weak: A photo of a coffee cup on a table.

Strong: Replace the plain white wall behind the coffee cup with a warm wood bookshelf. Keep the cup, its handle facing right, the saucer, and the table exactly as they are.

The second prompt is longer, but every added clause is a constraint that stops the model "improving" something you wanted untouched. Be specific about the things that matter most: not "keep the product" but "keep the label text and the bottle angle identical."

A couple of extra levers worth knowing:

• One change at a time. A prompt asking to swap the background, change the outfit, retime to night, and add a dog tends to do some of them badly. Chain two or three focused edits and you control each one.
• Match the photo, not your memory. Describe what's actually in the frame the model can see ("the man in the grey hoodie on the left") so it edits the right element.

Step 2: Use reference images for things words can't pin down

A prompt describes; a reference image shows. Wan 2.7 Image accepts one base image plus up to eight reference images, so you can feed it several sources and combine them in one edit — put your subject into a specific scene wearing a specific garment, each from its own photo.

This is the part most people under-use. If you keep regenerating trying to describe "this exact style" or "this exact logo treatment," stop and upload it as a reference instead. The catch: when the model has several images, order matters, because prompts that point at sources use Image 1, Image 2, and so on in the order you attached them. Reference a clean, well-cropped source — a reference full of its own background drags that background into the result.

For a holistic style or mood, one reference often does the job ("match the lighting and color grade of Image 1"). For a structural swap — this face onto that body — give one subject reference and one scene reference and let the composition do the rest.

Step 3: Pick the tier and size deliberately

Both variants run on Moosky with credit-based pricing (100 credits per dollar), and the exact cost is shown before you generate, so the choice is informed, not a guess.

A clean draft-to-final pattern works well here: run the edit on Regular to lock the look, then re-render the winning prompt on Pro once you know it's worth it. Resolution is set per run, so match the shape you need — vertical 9:16 for stories, square for feed, landscape for banners — instead of cropping a square render afterward and losing edge detail.

Step 4: Spot the failure modes before you spend credits

Most disappointing edits are one of a handful of recognizable problems. Knowing them saves a regenerate.

The whole image drifted, not just the target. The model re-rendered everything and your face or product is subtly different. Fix: the edit is fine, your keep-list was too vague. Add "keep [X] exactly as it is."

Text went wrong. Wan 2.7 is a pixel-space model, so readable text is its weakest area — logos, signs, and captions can come out misspelled or warped. If the edit touches on-image text, don't ask this model to render it: blank or carry the text region, then add the real lettering in an editor. For work whose whole point is crisp headline text, a typography-focused editor is the better tool.

A small cleanup turned into a big re-render. For one targeted local spot-fix, a holistic model can over-reach. Describe the spot tightly and forbid the rest ("remove only the power cable along the left edge; change nothing else").

A face or identity slipped. Push an edit too far and identity drifts, worst in tight crops. Keep the change small and explicit ("her exact face, just smiling"), and lean on a clear subject reference.

A worked example: a product shot that needs three looks

Say you have one studio photo of a bottle on a white background and you need it across a launch week. The efficient path is a series of tight edits off that single frame, each keeping the product fixed:

1. Lifestyle placement. "Place the bottle on a sunlit bathroom counter with a folded towel and a plant behind it; keep the bottle, label, and cap identical." The keep clause protects the only thing that must not move — the label.
2. Seasonal restyle. "Same bottle, relit as a warm holiday scene with soft bokeh lights behind it; keep the bottle unchanged."
3. Crop for stories. Re-render the lifestyle shot at vertical 9:16 so it fills a phone screen natively instead of cropping.

Three on-brand assets, one source image, each edit a one-sentence prompt with a keep-list, from a model that adds no watermark. The same loop covers portraits (background swaps, restyle), real estate (relight, declutter), and any "right photo, wrong detail" job.

Frequently asked questions

Can Wan 2.7 really edit, or does it just regenerate from scratch? It edits. Give it a base image and your requested change and it re-renders that image with the change applied while holding what you ask it to keep. The quality of the "keep" comes almost entirely from how clearly you state it.

Is there a free trial? New Moosky accounts get signup credits to run your first edits with no card, then it's a couple of dollars' worth of credits per batch. No subscription, and unspent credits don't expire.

How many images can I use? Up to eight references on top of your base image — enough to place a specific person in a specific scene wearing a specific outfit. Keep them in prompt order and use clean crops.

What resolution does it reach? Regular tops out near 2K. Pro reaches 4K landscape or portrait and is the tier for finals and large-format output.

Try it

Open the Wan 2.7 Image generator for text-to-image and prompt-only work, or the Wan 2.7 image editor to transform a photo you already have: upload the base image, describe the change and the keep-list, set tier and size, and check the credit cost before you run. No watermark, no install, no subscription.