> For the complete documentation index, see [llms.txt](https://help.intangible.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://help.intangible.ai/visualize/edit-images.md). # Edit images The pencil-tool fix-it-in-place workflow. Recolor, remove, or modify specific parts of a generated image without regenerating from scratch. A render came back almost right. The composition is good, the subject is right, but one element is off – a stubborn extra cop car, a sign that should be a different color, a character whose shirt should be red instead of blue. The pencil tool on a generated image lets you describe the change in plain language and the model edits the image in place. ## Watch {% embed url="" %} ![Edit Image model picker dropdown open showing the provider sections and the available image models, with Nano Banana Pro selected](https://1179478100-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FSWrjIqlgVAblx6vHGJiM%2Fuploads%2Fgit-blob-72ddb9d98abd3ee0cf0170112817d8984de7c6c9%2Fedit-image-model-picker-01.png?alt=media) ## What it does Diffusion models can edit images they generated, given a textual instruction and the original image as the spatial truth. The pencil tool exposes that flow: describe what should change, the model applies the edit and returns a new version of the image with that change applied. The rest of the image is preserved as much as the model can manage. ## How to use it ![The whole edit: opening the workspace on a render, drawing two marks, typing a note on each, writing the prompt, and generating the result](https://1179478100-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FSWrjIqlgVAblx6vHGJiM%2Fuploads%2Fgit-blob-3fcef5959641c118a8c51882e4d7f083c5cd71f7%2Fmarkup-workspace-02.gif?alt=media) 1. **Pick the rendered image** in the gallery to the right of the viewport. The image opens in a larger view. 2. **Click the pencil icon.** The image opens in an editing workspace rather than a dialog: the picture with a drawing layer over it, and a controls column beside it. 3. **Mark what you mean.** Draw on the image to point at the thing you are talking about. You can leave several marks, each with its own note. A note starts from *"Add a note..."*, and the field for what that mark should change asks *"What should change at this moment?"* The marks you have made are listed under **Annotations**, with a count, and each row can be deleted from there. 4. **Say what should change.** On an image the prompt field reads *"Describe the changes you want"*. Plain language. *"Make the car lime green."* *"Remove the trash can from the sidewalk."* On a video the same field asks you to ask for anything, and names a new ending, a look, an effect or a camera move as examples. 5. **Add references if the words are not enough.** The **Elements** section takes reference images, for when it is easier to show the model than describe it. 6. **Pick the model and Output Quality**, then generate. The result lands as a new media item in the gallery. 7. **Click Done** when you are finished with the workspace. Marking beats describing when the image has more than one of something. *"Make the car green"* is ambiguous in a street full of cars. A mark on the one you mean is not. ## The edit is kept, not just the result The workspace is not a one-shot dialog. It stays with the media: your marks, your notes, the prompt, the references, the model and the quality all persist, so you can come back and adjust rather than starting again. When you generate, the new image carries a frozen copy of all of that. **What produced a given version stays recoverable**, which matters three weeks later when a client asks why a shot looks the way it does and nobody remembers the prompt. If you remember editing as a dialog, with a before-and-after preview and an Edit Instructions field, this workspace is what replaced it. ![The Edit image workspace on a rendered photo: two marks sit on the car, the drawing toolbar runs under the image with a Done control above it, and the panel beside it carries Elements with an Add button, Annotations listing two notes, and the prompt field](https://1179478100-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FSWrjIqlgVAblx6vHGJiM%2Fuploads%2Fgit-blob-f986e55b4e819b2a76ea5bb99ea0cbeff48148ee%2Fmarkup-workspace-01.png?alt=media) ## What works well * **Color swaps.** *"Make the car blue."* The model is good at recoloring specific objects. * **Object removal.** *"Remove the trash can."* The model fills the space behind with plausible content. * **Object addition.** *"Add a flag on the building."* The model places the object plausibly. * **Surface treatment changes.** *"Make the floor wet."* The model adds reflections and atmosphere. * **Lighting adjustments.** *"Add warm light from the right."* Limited but workable. ## What works less well * **Replacing a subject with a specific other subject.** *"Change the car to a Lamborghini Revuelto"* often produces a generic Lamborghini. For specific brand replacements, regenerate from the 3D scene with the right asset and image reference. * **Major composition changes.** *"Move the character to the left of the frame"* doesn't work cleanly. Composition lives upstream in the 3D scene. * **Multi-object orchestration.** *"Make the cars red and the trees green and the sky stormy"* tends to confuse the model. Do edits one at a time. {% hint style="info" %} The image-edit flow is fast iteration on small problems. For large problems, regenerate from the 3D scene. The dividing line: if the change you want is something a colorist would do in post, edit-in-place; if a director or DP would do it, fix upstream. {% endhint %} ## Specific reference language Help the model find the thing you want to change: * *"Remove the trash can from the sidewalk"* beats *"remove the trash can"* if there are multiple trash cans in the image. * *"Make the closest car red"* beats *"make the car red"* when there are multiple cars. * *"Make the dashboard glow blue"* beats *"add blue light"* for specific surface lighting. The model honors specific reference language better than vague language. If you'd describe the change to a colorist, describe it to the model the same way. ## Limits and known issues * **Edits are not undoable into the original.** Each edit produces a new version in the gallery. The original is preserved alongside; you can return to it whenever. * **Edit cost is per attempt.** A failed or unusable edit still costs credits. * **Edit support is per-model.** All current image models support in-place editing. The pencil icon dims if the active model doesn't, which is the cue to switch models before retrying. * **Quality drops on heavy iteration.** Editing an edited edited image five generations deep produces compound artifacts. After a few iterations, return to the source and start fresh. ## Related * [Generate image](/visualize/generate-image.md) * [First and last frame](/visualize/first-and-last-frame.md): for "fix in Photoshop, regenerate as start frame"-style edits. * [Custom styles](/visualize/custom-styles.md) * [Managing outputs](/visualize/managing-outputs.md) --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://help.intangible.ai/visualize/edit-images.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `build a script that syncs our docs to a CMS` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.