Flux ArtAn annotation box is an input cue, not a promise of pixel protection
When an image contains two similar headings, an instruction to replace the heading may affect the wrong one. Flux Art users who separately evaluate Black Forest Labs' FLUX.1 Kontext can refer to its annotation-box method: https://docs.bfl.ai/kontext/kontext_image_editing . A colored box within the image helps identify the target text region. It is not a mask and does not guarantee that pixels outside the box remain unchanged. This page explains the provider's method and review steps; it does not claim Flux Art offers the same tool, and no comparative generation test was performed.
First confirm that the task needs generative editing
If you have an editable source document and only need to change the copy, editing the typography directly is usually easier to control. If only a flattened image remains and the text is baked into the background, assess whether generative editing is appropriate. The provider currently describes Kontext as a previous-generation model with corresponding access still available. Do not repeat an old draft's claim that it is the latest flagship. See https://help.bfl.ai/articles/5186006235-what-is-flux-1-kontext for version context. This page addresses targeting within an existing Kontext workflow, not a recommendation to replace every new project's approach.
Save the original, annotated input and generated result separately
Copy the approved, clear original image. On the copy, draw one easily recognizable box around the text to change and save it as a separate input file. Do not mark the original or overwrite it with a result. Record the line the box identifies, the original wording, the replacement wording and whether other text must remain unchanged. Do not treat a compressed screenshot from a client chat as the production original. If a screenshot shows only part of the image or obscures a key area, return to the complete source image to locate the target. If its provenance is unknown, obtain supporting evidence first.
| Where to check | Supporting evidence | What constitutes a failure |
|---|---|---|
| Actual input file | Unmarked original and annotated copy | A compressed annotation screenshot is mistaken for the original |
| Target text inside the box | Approved original and replacement wording | The wrong line changes, characters are missing or a price changes |
| Other similar text | Comparison of areas not intended for editing | The second identical heading also changes |
| Edges of the guide box | Complete result and enlarged details | Box lines, corners or color traces remain |
| Product and packaging | Approved original showing the real product | Details or structures change outside the intended scope |
| Final placement size | Actual downloaded dimensions and final image | Text becomes unreadable or cropped after resizing |
The box marks a location; written instructions specify the action
If a box contains both a heading and a price, 'change this' is not enough. Specify the original wording, replacement wording and any movement or resizing separately. For example, request that only the old heading inside the box be replaced with the approved new heading, without intentionally modifying the other numbers inside the box or content outside it. Ask for the guide box to be absent from the output, but verify the actual result. Start by testing one clear change. Do not combine it with a complete style change or major rearrangement, which makes it harder to tell whether an error came from an ambiguous target or an overly broad edit.
With two identical headings, verify which one changed first
Suppose a campaign poster repeats the same heading at the upper left and lower right, but only the upper-left change has business approval. Mark the upper-left heading on the input copy, and provide both the original and replacement wording. In the output, first check that the upper-left change is correct, then verify the lower-right heading, and finally inspect for box residue. This is a hypothetical scenario, not a test result or a guarantee that marking will restrict the change to one location. If both headings change, cropping out the lower-right one does not make the task successful. Return to the original and adjust the method, or use a workflow with precise typography controls.
Reject box residue and product-edge changes separately
The provider's examples include results with the annotation box removed, but one example does not establish successful automatic removal on every attempt. A result may retain colored edges or corners, or show an unnatural transition near the original box. Record the target text, remaining guide marks and changes to other products or text separately. A missing box does not make the result acceptable if the product packaging has been redrawn. Inspect both at the intended viewing size and enlarged at the file's original resolution, rather than relying only on the generation page's thumbnail.
Do not overlook output-size changes in a local edit
Kontext's official documentation describes output of approximately one megapixel with corresponding aspect-ratio handling. Do not promise identical input and output pixel dimensions. Record the actual downloaded width and height, then inspect the text and details at the intended placement specifications. Tasks that require original product pixels, fine packaging text or engineering structures to be preserved should consider manual typography or compositing that retains the original. Stronger prompt wording is not grounds for approval. If the output needs resizing, check the resized final image again.
After a failure, return to the approved baseline
If an attempt changes another heading or leaves box marks, keep the failed image for diagnosis, but use the approved baseline or an accepted version as the next input. For defining the task before editing, refer to the version-difference work order at Related workflow . That link provides a task-recording method; it does not mean the two models share the same controls. Record seam problems against their actual boundaries as well. Repeatedly editing a failed result can lose the original basis for comparison.
Explain the relationship between input marks and final output
Keep the unmarked original, annotated input, explicit edit instructions, generated candidates and conclusions for all three checks in the handoff package. The recipient should understand that the box is guidance, not design content. Mark unverified areas that should have been preserved as pending; an attractive overall image does not replace item-by-item comparison. Operated by MORNING STAR INDUSTRY LIMITED, Flux Art is a multi-model AI visual creation and production platform, not a single Black Forest Labs model. Its entry point is https://flux-art.net/en , and generated content can be used commercially. Separately check the provider's terms for each model or service, third-party input rights and product accuracy. Do not conflate the licenses of different versions.