Adobe’s Generative Fill and Expand tools were first released in 2023 and have received several updates since. While generative AI images can be a controversial topic, there’s no doubt users can gain benefits by utilizing AI. When it works. False positives in content guideline warnings and inconsistent results often make simple tasks, like removing an object or extending an image, incredibly frustrating.

Users on Adobe’s support forums and Reddit have also been questioning whether the generative results have been getting worse instead of better. Adobe’s standard response to questions about guideline violations is that their goal is to provide a safe and enjoyable experience for all users. They don’t offer solutions, and instead dismissively point frustrated users towards the Report tool.

It can be very difficult to figure out how to effectively use these tools when Adobe doesn’t provide any useful information on how to use them. This guide provides actionable advice for artists, designers, and photographers that Adobe does not. Rather than offer generic advice, these tips can specifically help you make the most of these tools until Adobe finds a better way to implement them.

Making Adobe Generative Fill and Expand work for you

Context matters

Photoshop’s AI analyzes the area around your selection as well as the entire image. You may get bad results if the context is unclear, or perhaps a guideline violation warning if it’s something Adobe decided is problematic. It’s important to know that Adobe’s AI is heavily biased against women in images.

Even if you are trying to do something like add a hat to a man’s head, you might get a warning if there is a woman standing next to them. In either case, adjusting the context can help you work around these issues. Always duplicate your original image, hide it as a backup, and work in new layers for the temporary edits. Click on the top-most layer in the Layers panel before using generative fill.

Modifying the background

Generative Expand can sometimes provide enough additional context for generative fill in another area to work better. Cropping problematic or unwanted elements out of the image can remove potential guideline violation triggers.

1. Adobe’s Remove Background feature in the Contextual Taskbar can help if you are editing something in the subject area.
2. Then you can click Generate background or Import background.

Blurring the background is another method to work around this.

1. Use Select subject in the Contextual Taskbar.
2. Reverse the selection using Select > Inverse in the main menu.
3. Go to Filter > Camera RAW filter.
4. In the right settings menu go to Lens Blur. Adjust the amount of blur and click OK to save and return to the image.

Paint over problem areas

Create a new layer and black out areas you don’t want Generative Fill to see by using a Rectangle tool or Brush tool.

Using the Clone Stamp tool to roughly cover potential problem areas can sometimes work better than blacking them out. This doesn’t have to be precise. For example, you could clone and paint a woman’s shirt to appear longer if there is any stomach area showing.

Copy and paste something from another area of the image or from a stock image to cover a problem area. For example, in an outdoor photo, cover a subject or part of a subject with a tall shrub from a stock image.

Experiment with rough edits

1. Copy and paste an element you want in the image from a stock image. Make rough edits to blend it into your image and use a prompt that slightly changes the element. For example, generate a red pillow on a couch by pasting in a brown pillow and prompting for a red pillow. This can also be extremely helpful to add context elsewhere in the image.
2. Copy and paste something you want to adjust the location of. For example, if you want to make an arm outstretched instead of bent at the elbow, copy the arm from the elbow down and paste it in the position you want. Fill the old area using the previously mentioned methods.
3. Use the Clone Stamp to make some modifications. For example, change a short sleeve shirt into a long sleeve shirt by cloning some of the shirt and painting over the subject's arms.

Selection area

Don’t select the area with tight precision. Despite this being common advice, in many applications it’s generally better to also select a little over the edge of where you want to use Generative Fill. Sometimes, roughly circling the general area that you want to fill works better than tracing the outline of something.

You can choose Expand selection in the Contextual Taskbar if you use Select subject or the Quick Selection tool to make your initial selection.

When you’ve made a jagged or sloppy selection using the Lasso tool, you can choose Smooth selection in the Contextual Taskbar to straighten out the edges.

Choose Select and Mask in the Contextual Taskbar for more precision.

There are multiple selection tools settings, and you can see changes to your selection in a live preview.

Iterative techniques

Large-scale edits can often cause Photoshop’s AI to return distorted outputs. Breaking edits into smaller, iterative steps can improve your final results. Think of using Generative Expand as if it were a brush. Rather than generating a lot of variations until you get one you want to use, choose a decent variation and fix problem areas with further Generative Fills. This can often lead to better results with far fewer generative variations.

When expanding larger areas, you might get distorted outputs as well, but you can also run into numerous violation warnings. Expanding in smaller increments can help reduce these problems. This is speculative, but I believe violation warnings can occur because a larger, expanded area gives more possibilities for content that could potentially violate guidelines.

For example, a photo of a woman standing that ends just below the belt line and clearly shows that she is wearing jeans can frequently cause warnings if you try to expand it to the knees. I think this is because Adobe’s AI could potentially generate a skirt or shorts that are too short for their strict guidelines. Smaller increments are more likely to generate denim in each iteration.

Working with large 4K or 8K images can be difficult. I suspect this may be for similar reasons, that Stable Diffusion XL (SDXL) works best in 1024 pixel aspect ratios. I’ve found that limiting the expand or fill areas to 1024 pixels improves results. There is a range for this, and you don’t need to measure pixels if you roughly know what a 1024x1024 block in your image looks like.

Think of a bounding box around your Generative Fill selection, and try to keep it inside that block. Keep the shortest side at or below 1024 when using Generative Expand. This isn’t a hard and fast rule, and if you want to take it a step further, I suggest you look up SDXL aspect ratios that include portrait and landscape.

Prompting tips

You probably know that Adobe has virtually no documentation that is actually helpful if you’ve tried to look up how to prompt well in Photoshop. Much of the information on how to prompt for Adobe Firefly doesn’t apply to Photoshop. This is especially frustrating because Adobe’s guideline violation warnings don’t tell you why you got a warning.

Guideline violations are still frequent when there is nothing in the image that seems to have the slightest possibility of being against the guidelines. In those cases, the violations could be triggered due to issues with the prompt. Although I still don’t know how to prompt well in Photoshop, I have picked up a few things over the last year that could be helpful.

1. Natural language descriptive prompts don’t work well. The results seem to get worse with longer prompts, but even one-sentence natural language prompts don’t seem to work well.
2. Don’t use direct instructional prompts. The AI doesn’t respond well to commands like remove, replace, modify, or add.
3. Use a blank prompt for the removal of objects. This is something Photoshop is good at if you just let it do its thing. Blank prompts can sometimes be useful to fill and expand, but it does leave a lot up to the AI.
4. One-word prompts can sometimes cause a warning. According to Adobe, this is because the AI might misinterpret a single word as an unsupported language. Try adding one or two descriptors when this occurs. For example, if your prompt is to add a hat to a subject’s head, instead of prompting only “hat,” you can prompt “black hat” or “black baseball hat” instead.
5. Try synonyms for words that you think might be causing a guideline violation warning.
6. Comma-separated prompts tend to work better than descriptive prompts. For example, “a black cat sits lazily, relaxing in the morning sunlight” could be changed to “black cat, sitting, sunlight.”
7. In general, short prompts formatted in a simple statement work best. Example, “Black cat sitting in sunlight.”

Conclusion

Adobe’s Generative Fill and Expand tools can be frustrating, but with the right techniques, they can also be very useful. Experimenting with selections, context, and prompts can play a big role in getting a quality result. Make sure to keep in mind the size of the area you are generating and consider working in iterative steps, instead of trying to get the perfect result from a single prompt.

Adobe could improve the user experience dramatically by simply including the reason a generation gets flagged as a guideline violation. They request we use their feedback system when this happens, but don’t give us any feedback in return. A basic tutorial on how to best prompt their AI model in Photoshop would also be very helpful, because prompting doesn’t work in the same way as using Adobe Firefly’s website. I’d rather see them make these simple improvements to the user experience before getting an update with a new feature added.

For those seeking more control, consider exploring tools like Stable Diffusion and ComfyUI. While they have a steeper learning curve and require a GPU with at least 6-8GB of VRAM, they can easily blow Photoshop out of the water. If you’re looking for something in-between, we know some great alternatives, and they’re even free, so you can save on Adobe’s steep subscription prices.