Flux Kontext Outfit Swap: Prompts and Quality Checks

For a Flux Kontext outfit swap, describe the garment you want to change and name the features you want to preserve. Then compare the result with the source image. A preservation instruction is a request to the model, not a guarantee that the face, pose, or background will remain identical.

Correction, September 19, 2026: The previous version reported experiments and success rates without supporting records. Those claims have been removed. This page now provides an editing procedure and an original quality-control worksheet; it does not report a completed benchmark.

Start With One Garment Change

Use an image you have permission to edit. Pick a source where the garment boundary and the person's face are visible. For your first attempt, keep the pose, background, camera angle, and lighting request unchanged. That makes it easier to identify what the clothing edit affected.

Here is an example prompt to adapt, not a tested result:

Replace the person's gray hoodie with a navy cotton overshirt, worn open over a plain white crew-neck T-shirt. Retain the person's facial appearance, hairstyle, expression, body position, framing, and room background. Keep the existing direction of the light.

Specify the garment, color, material, and fit you actually need. “A formal look” leaves more choices to the model than “a charcoal single-breasted jacket over the existing white shirt.” If the current task is changing the shirt, adding a new room and a different pose creates extra changes to inspect.

The official ComfyUI Kontext Dev workflow accepts an input image and an editing prompt. Its guidance recommends specific change instructions and explicit preservation requests. Use that documented workflow as your starting point, checking that its model, text encoders, and VAE load correctly. Pro, Max, and Dev are distinct variants; record which one you used rather than treating their outputs as interchangeable.

Check the Face Before Selecting a Result

A convincing outfit can distract from a changed face. Open the source and output side by side at the same display size. Inspect them in the same order each time:

This is a visual acceptance checklist. It is not biometric identity verification, and it does not establish a measurable preservation percentage.

For clothing sold to customers, compare the output against actual product photography and specifications as well. A plausible invented pocket, logo, seam, or fabric texture can still misrepresent the item. Do not use an attractive generated result as evidence of real garment fit or construction.

Recover From a Changed Face or Background

Return to the original image if an edit changes the face. Simplify the request to one clothing change and make the preservation instructions explicit. Keep the failed output so you can compare what changed; do not record only the best result.

Several sequential edits can accumulate differences. If you need a jacket, new background, and different framing, save each accepted stage separately. When a later stage fails, restart from the last acceptable source instead of editing a degraded result again.

If a tightly restricted clothing region is essential, consider a workflow with an explicit mask and verify how that workflow treats pixels outside the mask. Do not assume a text instruction creates a hard mask. Any masked workflow has its own model and compositing behavior to check.

There is no universal denoise value or wording that ensures success across every Kontext workflow. Preserve the working graph and settings with your output. If you change the workflow, document that change before comparing results.

Keep a Small Editing Record

Use one row per attempt. This makes a later comparison possible without relying on memory:

Source image identifier:
Permission or license record:
Model and variant:
Workflow version:
Prompt:
Seed and exposed settings:
Output identifier:
Face check: pass / fail / uncertain
Garment check: pass / fail / uncertain
Scene check: pass / fail / uncertain
Decision and reason:

Define “accepted” for the intended use before counting results. A mood-board sketch and a customer-facing catalog image need different scrutiny. Include rejected attempts and any generation charges when calculating the cost of an accepted image.

For a batch of outfit variants, apply the same checklist to every output and preserve the original source. Reusing a reference or a saved character setup can help organize the work, but neither removes the need to inspect the result. Apatero publishes this guide; availability of a particular editing model should be checked in the product before planning a paid batch around it.