# How to edit AI visualizations? See how to fix a render with artificial intelligence > Selections, references, camera angle and the 2x2 grid. Learn how to fix finished AI visualizations so every edit lands exactly where you want it. Source: https://www.raya.design/en/articles/how-to-edit-ai-visualizations-see-how-to-fix-a-render-with-artificial-intelligence Published: 2026-10-13 Updated: 2026-10-13 Author: Piotr Kurpiewski, Architekt, Grafik Category: Tutorial Language: en ![AI kitchen visualization created in the "Edit" tab in Raya](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x2miniatura.jpg) --- The best way to edit an AI visualization is in the "Edit" tab, in small steps: you select part of the image, type a short instruction next to that selection, and when you're swapping a specific piece of furniture, you add a clean photo of it as a reference object. Leave the general field for changes to the whole frame, such as time of day or lighting, and fix anything you can measure in centimeters in SketchUp. In the rest of this article I'll show you how to arrange your instructions so the AI does what you ask, how the camera angle and the 2x2 grid work, and where editing reaches its limits. **Key takeaways** - **You fix a finished visualization in the "Edit" tab** — The "Generate" tab creates a visualization from a 3D view or a sketch. "Edit" keeps the style of the finished image and changes only what you point to, for 2 credits in 2K. - **Fix the look in editing and the dimensions in the model** — Wall color, a new piece of furniture or the time of day are jobs for the "Edit" tab. Moving the sofa by 10 cm or rotating tiles by 90 degrees is faster and cheaper to fix in SketchUp. - **Mark a specific spot with a generous frame** — Keep the "What do you want to change?" field for changes to the whole frame, such as lighting or time of day. The description, frame and reference for a single change sit in one row of the panel. - **A reference shows only the object or only the material** — A cropped packshot instead of a store screenshot, and a photo of a sample instead of a tile name from a catalog. Raya doesn't know every product on the market. - **Two or three changes per round work best** — With ten changes, some won't go in or will land in the wrong place. When an edit doesn't work out, generate it again instead of adding more prohibitions. ## What to fix when editing an AI visualization, and what to fix in SketchUp? Before you select anything, ask yourself one question: does the change concern the look or the dimensions? The answer decides where you make it. AI models are probabilistic. The same operation repeated ten times will give ten slightly different results, and roughly one generation in ten will misread the guidelines. That's why we've put so much work into the **"Edit" tab**. Instead of promising you a flawless first generation, we gave you tools for quick fixes. A different wall color, a new lamp, fewer decorations on the table or golden hour instead of a cloudy afternoon are changes you'll make there in under a minute, without going back to the model. Precision, however, belongs to CAD software. Commands like "move the sofa 10 cm", "add 60x90 tiles and rotate them 90 degrees" or "widen the window by half a meter" are jobs for SketchUp or another program whose strength is accuracy. You can try making them in Raya, but don't count on a result that's precise to the centimeter. **A fix in the model takes you a few minutes and costs zero credits**, and when you generate again, the AI gets unambiguous information. The table below will help you quickly decide where to make a change: | Type of change | Where to make it? | Why? | | --- | --- | --- | | Dimensions, moving furniture, tile layout and size | SketchUp or another 3D or CAD program | Precision to the centimeter | | Finishes for the whole interior, a dozen or so materials at once | SketchUp, then the "Generate" tab | A material assigned to a surface in the 3D model will be faithfully reproduced by Raya | | More views of the same interior in one mood | Raya, the "Generate" tab with the "Match style" feature | This feature only works in the "Generate" tab | | Wall color, swapping furniture, removing or adding an object | Raya, the "Edit" tab with an area selection | The change covers only the indicated spot | | Time of day, lighting, weather | Raya, the "Edit" tab, the "What do you want to change?" field | The change applies to the whole frame | | A new shot, a close-up, a composition for social media | Raya, the "Edit" tab, the "Camera angle" or "2x2 Grid" feature | Raya creates new frames based on the finished visualization | If you've read the guide on [how to prepare a SketchUp model for AI visualizations](https://www.raya.design/en/articles/how-to-prepare-a-model-in-sketchup-for-ai-visualizations-framing-edges-shadows-and-export), you already know the rule: whatever you can fix in the model in a few minutes, fix in the model. Editing adds a second one: **fix the look in Raya and the dimensions in the model**. ![Original AI kitchen visualization](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x201.jpg) ![Correctly generated visualization with changed kitchen fronts](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x202.jpg) *Good example — Changing the color of the kitchen finish is a typical change you can make* ![Original AI kitchen visualization](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x201.jpg) ![The change failed because the instruction about dimensions is hard to understand](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x203.jpg) *Bad example — The command "widen the large kitchen front by 30 cm" isn't understandable to the AI and in the end changed nothing in the scene* [Try editing for free](https://www.raya.design/en/edit) ## How is the "Edit" tab different from "Generate"? Now that you know which fixes stay in Raya, make sure you're making them in the right tab. It sounds trivial, but this mistake comes up regularly. The **"Generate"** tab is for creating visualizations from a sketch, a view from a 3D model or a photo. It reads the geometry of the input and adds photorealism to it, meaning materials, light and surroundings. The **"Edit"** tab works on a finished visualization, keeps its style and changes only what you point to. An edit is a full generation, so it costs the same: **2 credits in 2K or 4 credits in 4K**. What happens when you upload a finished visualization to the "Generate" tab and add "change the sofa to pink"? The sofa may change while the rest of the image stays the same. But the AI may just as well read the lighting or decorations differently along the way, because this tab wasn't built for fixes - the result is hard to predict. The mistake also happens the other way round. A raw SketchUp view uploaded to the "Edit" tab will stay in viewport style, because editing keeps the style of the input. Heads up! The button at the bottom of the "Edit" tab is also called "Generate Visualization". What happens to the image is decided by the tab selected at the top of the screen. ![AI visualization before the change](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x205.jpg) ![Sofa color change made in the "Edit" tab](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x206.jpg) *The "Edit" tab is for making changes to finished AI visualizations.* ![Working view from SketchUp](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x204.jpg) ![Finished visualization generated from a SketchUp view in the "Generate" tab](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x205.jpg) *The "Generate" tab turns raw 3D views into photorealistic visualizations* ## When to type a change in the general field, and when to select an area? You're already in the right tab, so look at the panel on the right. You can type an instruction in three places, and each one is for something different: - **the "What do you want to change?" field** \- an instruction for the whole frame, - **the "Select area" feature** \- a red frame on the image and a description of the change in that one spot, - **the "Reference objects" feature** \- a photo of the object, its frame on the image and a short description. The general field handles changes that affect the whole image, such as time of day, lighting or weather. A simple command like "remove the chair" will also work there, as long as there's only one chair in the frame. When there are six chairs and three visible walls, the AI has to guess which one you mean. **You'll get the best result when you precisely select the spot to change and write the instruction right next to the selection.** The field accepts up to 500 characters, but short, unambiguous instructions work best, in practice up to 150-200 characters: "change the time of day from morning to golden hour", "paint the wall brick red", "remove the floor lamp". ![Editing an AI visualization: a simple instruction and a red frame selection](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x210.jpg) ![AI visualization with one changed wall](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x208.jpg) *Good example — Instruction next to the red frame: only the selected wall changes* ![Editing an AI visualization: an instruction without a selection](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x207.jpg) ![AI visualization with all walls changed](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x209.jpg) *Bad example — Instruction in the general field: the AI changes all the walls* ### How big an area should you select? Draw the frame with some margin. The model needs to see a bit of the surroundings in the selection to understand what it's actually changing: what the armchair stands on, where the sofa ends and which side the light falls from. A frame cropped tight to the object's outline cuts off that context. Treat the size of the frame itself as a suggestion. An inserted person or piece of furniture often comes out smaller or larger than the rectangle, because Raya matches the object's scale to the perspective of the scene on its own. Look at the example below. The orange frame reaches almost to the window, but the armchair from the reference was inserted at a natural scale. ![Armchair selection on an AI visualization with a margin covering the floor and part of the wall](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x211.jpg) *Frame with a margin: the model sees what the armchair stands on and how the light falls* *The image on the left produced the image on the right* ![AI visualization with a changed armchair](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x212.jpg) *The inserted armchair matches the scale of the image, not the size of the frame* ## How to arrange instructions so the AI doesn't mix them up? Three places for instructions are convenient, but they're also a trap. The most common mistake I see in edits is scattering one change across the whole panel: the description in one place, the selection in another and a mention of the reference in a third. Imagine handing a printout of the visualization to a retoucher. You circle the armchair with a red marker and write "green velvet" next to it. You paper-clip a catalog photo to the lamp and label it "this lamp instead of the current one". At the top of the page you jot down one sentence about the whole image: "warmer, evening light". The retoucher knows right away what to change and where. The editing panel in Raya works the same way. The "What do you want to change?" field is the note at the top of the page, a selection is a marker circle with a note, and a reference object is a photo clipped to a specific spot. Now the same retoucher gets a page where the circle around the armchair is labeled "here", the lamp photo lies loose with no indication of where it goes, and at the top someone wrote "put the lamp from the second photo above the table and change the armchair to green". It can be deciphered, but it's very easy to get it wrong. The AI has exactly the same problem. That's why you should stick to a hierarchy: **whole frame, then area, then object**. You type each instruction once and at its own level. The lamp photo, its orange frame and the description "pendant lamp above the table" sit in one row of the "Reference objects" feature, and in the general field you don't write a single word about the lamp. Also watch out for instructions that contradict each other. If you write "white walls" in the general field and "navy wall" next to a selection of one of the walls, the model gets two conflicting commands and has to decide on its own which one matters more. Repetition does similar damage, because the same change described in two places makes it harder for the model to understand the task. ![Raya editing panel: a chair reference with an orange frame, cabinet removals with a red frame and a general description](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x213.jpg) *Reference object, area selection and change description in a single editing round* *The image on the left produced the image on the right* ![Visualization generated from complex instructions](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x214.jpg) *The result of advanced changes made to an AI visualization* ### Why won't ten changes in one round work? Let's go back to the retoucher for a moment. This time they get a page with ten circles and ten notes: a new sofa, a different lamp, a green wall, an oak floor, plants in the corner, pictures above the chest of drawers, linen curtains and a few more small things. Something is bound to slip past them, and in a situation like that the AI won't ask what you meant. Some changes will go in, some will get lost along the way, and some will land in the wrong place, even if you describe each one correctly. That's why in one editing round, what works best is **two or three changes**. That's how many fit in the example above: the chair from the reference, the removed cabinets and the green wall. When a client sends a long list of comments, sort it by importance. Start with the changes that most affect how the interior is perceived, check the result, and only in the next round deal with the details. Each next round, however, is another processing of the same image. I explain below how many such rounds a visualization can take without losing quality. ### Should you add "leave the rest unchanged"? You don't have to. Many people end their instruction with a sentence like "don't change anything else", "leave the rest as it is" or "make no mistakes". Preserving the rest of the image is already set up on Raya's side, so a note like that adds nothing, and it actually gets in the way of the model reading the real change. If the AI changed something outside the selection, just generate the edit once more. Another try will give a better result than adding more prohibitions. ## What reference photo should you upload for an edit? Since the reference already has its place in the panel, the question is what should be in the photo itself. The "Reference objects" feature follows a simple logic: you show the AI what you want instead of describing it. For this to work, **the reference has to show only the object**. What works best is a sharp product photo on a light, plain background, cropped so the piece of furniture fills most of the frame. I've seen references in the form of an online store screenshot, with the price, description, several thumbnails from different angles and an "Add to cart" button. There are also styled room photos where two armchairs and a bookcase stand next to the chosen lamp. How is the AI supposed to know which of these elements you want to insert? Before you upload the photo, crop it to the object alone. It'll take you a few seconds. The reference object frame isn't mandatory, and the edit will work without it. With the spot marked, though, the result is more reliable, so it's worth adding it every time. ![Good reference for editing an AI visualization: a cropped photo of an armchair on a light background](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x215.jpg) *Good example — Just the armchair on a light background, cropped to the object* ![Bad reference for editing an AI visualization: an online store screenshot with the price, description and thumbnails](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x216.png) *Bad example — A store screenshot with the price, description and thumbnails is a puzzle for the AI* ### Why won't a product name from a catalog help? The instruction "put paradyz ravero wine glazed rectified matte porcelain tile 60x60 here" looks precise, but Raya doesn't know every product on the market and doesn't know what this tile looks like. It's better to write something general, for example "red large-format porcelain tile", and best of all to attach a good-quality photo that shows the material. You'll set the tile size more reliably in the model anyway. ![Editing an AI bathroom visualization with a tile inserted from a photo of a material sample](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x219.jpg) ![The result of a change using a photo of a material sample](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x218.jpg) *Good example — Sample photo and a short description: the tile matches the material* ![Editing an AI bathroom visualization with a tile described by its name from the manufacturer's catalog](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x217.jpg) ![The result of a change using a name from the manufacturer's catalog](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x220.jpg) *Bad example — Just the catalog name: Raya doesn't know what this tile looks like and guesses the result* ## How many changes fit in one edit? A well-prepared reference and a properly marked spot are half the battle. The other half is a reasonable number of changes, because this is where it's easiest to overestimate what AI can do. I'll put it plainly: **Raya is not a CAD program**. If you upload five references, draw four selections and add a general instruction on top, don't count on everything coming out right, even if you phrase each instruction as well as you possibly can. The more tasks in a single pass, the greater the chance that one of them gets skipped or interpreted in its own way. The same goes for edits where someone types a list of a dozen or so materials with exact assignments to the floor, countertops, fronts, walls, ceiling or curtains. Changes like that are made in a CAD program, because there you assign a material to a specific surface once and don't have to rely on luck. In Raya, smaller, targeted fixes work better than a global rebuild of the project. There's one more limitation we're open about: **generative compression**. Every further processing of the same image lowers its quality a little. With each new version of AI models this effect weakens, and today **three edits in a row on one image are safe**. With a longer series you have to accept the limits of the technology, so it's worth planning your fixes. Gather the client's comments and make them in one edit with three changes instead of three separate ones. You can do the upscale to 4K for 4 credits at the very end, once the client approves the version. ![AI dining room visualization before editing](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x221.jpg) *Visualization before editing* *The image on the left produced the image on the right* ![The same AI visualization after multiple edits](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x222.jpg) *Multiple edits inevitably cause image compression* ## What to do when an edit won't go in? Sometimes even a well-prepared edit doesn't work the first time. What then? Start with the simplest step and generate it again, and if necessary, two or three more times. AI sometimes makes mistakes, and another try with the same instruction is often enough. If that doesn't help, change your strategy: **remove the object and add it again from scratch**. The best example is a round table that's supposed to have three chairs around it. During training, AI models saw round tables with four chairs far more often, so the command "replace the four chairs with three chairs" can come back with four. The model simply doesn't know that only three can stand at a table like that. Fortunately, there's a way around it. First remove all the chairs, then add two, and in the next edit one more. Breaking the task into steps lets the model understand what you mean, and three edits fit within the safe limit. You can approach swapping a large piece of furniture for a small one the same way: first remove the sofa, and only then put an armchair in the freed-up space. Keep in mind, though, that this is more of a feel than a recipe. It's hard to describe exactly when the model needs this kind of breakdown and when it'll manage right away. A good signal is repetition. If several attempts in a row end with the same mistake, the model is most likely sticking to what it saw most often in training, and that's when breaking the task into steps makes the most sense. ![Round table without any chairs](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x223.jpg) ![Round table with three chairs on an AI visualization after removing the chairs and adding them in two steps](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x224.jpg) *Good example — Removing all the chairs and then adding three guarantees a better result* ![Round table with four chairs on an AI visualization](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x225.jpg) ![Round table on an AI visualization where four chairs are still standing after the command to leave three](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x226.jpg) *Bad example — The command "replace the four chairs with three around the table" doesn't give the desired result* ## How to change the camera angle on a finished visualization? So far you've been changing what's in the frame. The **"Camera angle"** feature changes the frame itself. You have six presets to choose from: "Low angle", "High angle", "From the left", "From the right", "Detail" and "Wide angle". Each of them creates a new shot based on the finished visualization and costs the same as a regular generation, that is 2 credits in 2K. ![Original AI hallway visualization](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x228.jpg) *Original AI hallway visualization* *The image on the left produced the image on the right* ![New shots created using the default camera settings](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x233.jpg) *You can control the camera position while keeping the geometry of your scene* When editing, Raya works only on the image and can't see your furniture on the other side of the room. Everything that wasn't in the visualization it has to fill in on its own, and the more you rotate the camera, the more it fills in. A cabinet that stood just past the edge of the frame may come out different from your design in the "From the left" shot, and "Wide angle" will build out a part of the room the AI has never seen. Approach these shots with some leniency and generate them a few times. When you need a view from the other side of the room that matches the design, set up a new scene in SketchUp. ### What will "Detail" show? For the close-up, Raya picks the most interesting object in the frame, and that one is easy to predict: it's usually the element that stands at the compositional center and catches the eye immediately. If a coat rack stands in the middle of the frame, you'll get a close-up of the coat rack. If it's off to the side and the center shows mostly a wall with a light switch, don't be surprised when the detail shows the switch. So before you click "Detail", look at the visualization and judge what catches the eye. If it's not the object you're after, combine the preset with an instruction in the "What do you want to change?" field. Choose "Detail" and add, for example, "close-up on the pouf", and Raya will point the frame at the object you indicated. ![Visualization before the detail shot](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x228.jpg) *AI hallway visualization before the transformation* *The image on the left produced the image on the right* ![Detail from an AI visualization: close-up on the pouf](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x229.jpg) *Detail preset with a simple targeted instruction: "close-up on the pouf"* ## 2x2 Grid: how to make four frames from one visualization? "Camera angle" gives you one new shot. If you want several at once, reach for the **"2x2 Grid" feature**. It creates four different frames from your visualization and arranges them into a simple composition. The whole thing is one 2K image split into four parts, so it costs 2 credits. After turning on the switch, you choose the orientation: landscape 16:9 or portrait 9:16, which fits stories and reels right away. The AI picks frames that look good. Look at the example below. One living room visualization produced a top-down view, a shot of a woman in a yellow dress against a wall with a painting, a close-up of the lamp above the coffee table and a wider shot of the whole room. The woman ended up in two frames, because she's the strongest point of this scene. Doesn't the whole thing look like part of a photo shoot? ![Raw 3D view from SketchUp](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x232.jpg) ![Finished AI visualization with a woman in a yellow dress](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x230.jpg?v=1791913843758) *Original visualization made from a SketchUp view* *The image on the left produced the image on the right* ![Four visualizations created using the "2x2 Grid" feature](https://pub-55300e22613247cdaa083cb8965745af.r2.dev/articles/8858cea9-88ac-4898-a2e9-2c2b5ed1e05b/jak-edytowac-wizualizacje-ai-zaznaczenia-referencje-kamera-i-siatka-2x231.jpg?v=1791913750667) *Four frames created automatically with the "2x2 Grid" feature* Keep two things in mind with the grid. Turning the grid on grays out the other editing features, so fix the visualization first and make the grid at the very end. Each frame is also half the width and half the height of the whole image, which is why the grid works best as a ready-made Instagram composition or a presentation board. When you want one of the frames separately in full quality, try to get a similar shot with the "Camera angle" feature. ## Checklist before editing an AI visualization Editing sounds like a long list of rules, but in practice it comes down to a few habits. Before you click "Generate Visualization" in the "Edit" tab, check that: - the change concerns the look, and you've fixed the dimensions and layout in the model, - you're working in the "Edit" tab, not "Generate", - the "What do you want to change?" field describes only changes to the whole frame, - every spot to change has its own frame drawn with a margin, - the photo, frame and description of a single change sit in one row, - no instruction repeats or contradicts another, - you're not adding "leave the rest unchanged", - the reference shows only the object, sharp and cropped, - you describe the material by its features or with a photo instead of a catalog name, - there are a few changes, not a dozen, and the image hasn't already been through three edits in a row, - before clicking "Detail", you can see what draws attention in the frame, - you leave the 2x2 grid for the very end. I hope that after this guide, editing AI visualizations stops being a lottery for you and every fix lands where you want it. Start with one change on a visualization you know well: select an area with a margin, type a short instruction and compare the result with the original. And if you don't have a visualization to fix yet, start with your first generation. A new account gets **20 free credits valid for 14 days**, with no card required, which is about 10 visualizations in 2K or a few visualizations and edits to try out. Prepare the view following the [guide on exporting from SketchUp](https://www.raya.design/en/articles/how-to-prepare-a-model-in-sketchup-for-ai-visualizations-framing-edges-shadows-and-export), and polish the rest in the "Edit" tab. [Generate your first visualization for free](https://www.raya.design/en/generate)