AI can make things look amazing, but often will make different changes than what you want. There is no perfect solution to this, but here is a guide on some things you can do to increase the chances of success:
Does it understand my image?
The AI model doesn't actually understand the picture or what you want. It tries to find patterns and concepts within the image (such as a concrete wall, or a shadow edge), and tries to make changes to these concepts based on the concepts you request in the prompt.
Using the Prompt Templates
You can create a prompt from scratch, but to start learning, it may be easiest to begin with one of our predefined prompt templates, which you can then modify as needed.
When adding a template, make sure the intention somewhat matches your original image. If you ask for an active construction site, but provide a grassy field, AI may struggle to determine what concepts to change, and instead provide you with an entirely different image.
We have 2 styles of templates available:
- Single Paragraph - For templates that change the environment of the scene (such as night or snow), the prompt usually works best with a paragraph.
- Sections with Headers - for templates that aim to populate or improve visual quality of the scene, we have the prompt split into sections. The purpose of this is for you to be able to remove or modify sections to better match your scene. You can also add your own sections.
Example using Sections with Headers
- Completed Construction is chosen, but because of the scene and the angle, the following sections should be removed from the textbox before pressing Generate:
- Sky Background (because sky is not visible in the image)
- Grass (because there was no grass on the ground, but there was some on a neighboring roof, it was adding grass to the new building's roof)
- Add Trees & Foliage (new trees were not needed, and with nowhere to place them, it added them in the street)
- Additionally, in this example, we may want to modify the Roadways and Cars section to provide specific instructions on the road on the left, where we used a grey zone to hide some of the satellite image.
Removed Sections
Updated Section - insert one row of text
Result
Setting up your Scene
Doing these things may improve the quality/control of how AI updates your image.
- Turn building edges off (if you are trying to get realistic textures)
- Add a resource for context, so that AI knows what types of things it should add, and what size/location they should be in
- such as a person, construction vehicle, car, tree, etc.
- You may want to temporary-hide labels and callouts
- Low-res models/textures are generally easier for the AI to improve upon, since there is more wiggle-room in their definition
Tips, Tricks, and Troubleshooting
Note that AI image models can only replicate concepts that it is familiar with from its training. Images for some stages of construction are not that commonly available, and may not be possible for AI to replicate.
The current AI models we use:
- Are good at adding and removing things, so try and describe things in that style.
- Work well with specifics, like saying "remove trees around the building".
- Work well with specifics and detailed descriptions.
It is not doing anything
If no changes are happening on the image output, this usually means that either the prompt is too vague in requesting what is needed, or the AI cannot interpret the concepts of the scene and match them to the concepts being requested in the prompt. Things you can try:
- Identify the specific objects in the scene that you want to change (so it can better map concepts to the image).
- Provide a more descriptive explanation of how you want specific things in the scene you want to change (you can also send the screenshot to an AI, and ask it to provide a description of what the output should look like).
- If the scene's objects are detailed or textured, try swapping objects in the scene to a simpler version, to allow for more artistic flexibility in what can be changed.
It is not changing a specific thing I want it to
AI wont always be good at changing every object into something else. It may not be able to classify the concepts of the object easily, or understand what should be done. Or perhaps it is getting distracted with other parts of a prompt. Things you can try:
- The suggestions from "It is not doing anything" can also apply here.
- If you are using a large prompt, try cutting everything else out of the prompt, to see if you can get it to just focus on that change (AI may be distracted from other concepts in the full prompt).
- The requests from another part of the prompt may be causing AI to try and do something in the opposite direction, conflicting with this.
It keeps changing something I don't want it to
If it keeps changing something you are not asking for, it is usually because it relates that output to other concepts it is receiving (either from the image or the prompt). You will probably want to isolate/remove the competing concept. Things you can try:
- If you are using a large prompt, try cutting everything unrelated to the affected area, out of the prompt, to see if the change stops.
- Try changing camera angle of the original screenshot, or even hiding/moving certain parts out of view, to see if it is one of the other objects/areas in the scene that is influencing it's bias.
- Try changing the original affected thing, such that it is perceived a little differently.
The added objects don't have the style I want
It can be challenging to force a specific style when the AI defaults to a different style. Things you can try:
- Adding something to the scene that has the style you are looking for
- If you are aiming to make things realistic, it may be easier to get it to add a fresh object instead of replacing a low-quality one (depending on the object and situation - so not always the case)
- Use specific descriptive words for the scene or object (words like photorealistic, rendering-quality, etc.). You can always ask AI for suggestions on types of words to try out.
The added objects are the right size
Usually when objects come in at the incorrect size, it is because something in the scene is distorting the AIs perception of what the real size/scale is. Things you can try:
- Adding reference objects in the scene to provide a sense of scale.
- Remove drawings or other abstract items that may be affecting perception.
- Add/remove/update textures of large objects, as the texture itself may be setting the perception of scale.
Comments
Please sign in to leave a comment.