P: Inpainting for text to image

[Lavanya232490872d40]

I would really love a feature where we can specify the part of the AI-generated art that we want to change. Instead of having the AI generate new pictures each time, with completely random changes, I want to be able to specify which parts I have a problem with. For example, if the hands are not realistic, I want to be able to select them and generate images with just the hands changed (everything else the same). Or, if I really like the picture, but want to the see it in a purple hue, i want to be able to do that. I understand it might be really difficult to implement though. 

[sergio.v]

Muchas veces la IA realiza algunas partes de la imagen de forma extraña y otras de forma correcta, si se le pudiera decir que modifique solo una parte, con una sencilla seleccion (como en Photoshop) 

[Aries1111]

Perhaps adding a feature like "redefine". Maybe this could be a simple slider, that adds / subtracts detail almost in an HDR type of fashion.

[Eric29053374mh6i]

Would you use a feature to select components or image area to keep or replace before next iteration?

[Tracy.H]

Thank you for your feedback. For now, you can use the "Use as reference image" feature to iterate on a generated image. In the future, we have inpainting and outpainting.

[Tracy.H]

Thank you for the suggestion. Have you tried the "Use as reference image" feature? Would like to hear your feedback on this and how your suggestion could be integrated.

[Tracy.H]

Have you checked out this video of features we're exploring?

[marliton]

Hola. Según se ve en la web de Adobe, lo que dices será implementado en futuras versiones.

[Chesthair]

Hi!

Kinda like how one might use reference image for targetting a specific vision/image.

If we had a tool to select specific areas of the image we would like to see a change.

Like in the picture below, I would love to change my "humming cats" tail, remove some freckles and tidy the surroundings.

If I could maintain the general image, but have the ability to target an area or object - By having the option to set the degree of expected difference with each "refresh"/iteration.

Or maybe even add additional prompts for targeted object/area: "This tail is very furry, and curled" / "Tall grass".

This would really help increase the efficiency in reaching a high quality result.

 Thanks.

[Aries1111]

Thank you @Tracy.H . I did try that feature and do love how it visually displays the image reference and has a simple slider. I do though wish it were a little clearer to the user as to what information it is "Using as reference" from the image. I wish that the little pop-up card that shows the selected reference image had perhaps a few more sliders like, sharpen more, or maybe something like "target" and the user could perhaps zoom in or target a specific area of the reference image if let's say I wanted to use that area to focus on. Maybe it could be like a small word prompt field in that small card and when the user focuses on the area to be used as a refernce I can then type in something like "use these colors, + sharpen 50% more + add more detail to the wings only" or soemthing like that. 

[Tracy.H]

This is a wonderful suggestion. Have you checked out this video which shows inpainting? (Merging all requests for this feature here so we can track the demand for it)

[Tracy.H]

Thank you for the additional ideas @Aries1111 

[Gary Ferster]

Edits will be a big part of generative AI art and maybe calculating masks for prominent objects, like the technology already does with the object tool in PS, could be a great method for doing it. There will always be work left for retouchers, unless of course the next beta drop has a "edit the  _______ "  prompt that only edits the part of the artwork you reference! 

[Chesthair]

Ooohoo, this is exactly it!

+ + Alot more cool stuff in the video!

Will they be usable on the beta soon? Seems to be locked for now.

Thanks

[Tracy.H]

Thank you for the suggestion. Inpainting is on our map of future features however we haven't yet determined how it will be integrated in our beta tool.

[Tracy.H]

Thank you for your feedback! Have you tried the "Use as reference image" feature under the "..." menu when hovering over the image? You can change the prompt to adjust your image. You're other suggestion regarding replacing a specific part of the image may be addressed with Inpainting. It's not released yet, but you can see it in

this video

.

[Guest]

Would be great if users can have masking brush tool for selecting parts of the image with distorted details. AI task would be to improve details on that part or remove strange objects which often appears on image. 

[Tracy.H]

Thank you for the feedback! Inpainting is a feature on our roadmap.

[Lavanya232490872d40]

I hadn't seen the video. It's almost exactly what I was thinking of. I hope to see this feature soon!

[Podsville]

I was about submit a suggestion to allow an uploaded image to be used as part of the prompt in Text to Image, but the "Use as reference image" feature you mention implies it already exists. Alas, I cannot find the "..." menu anywhere. Can you point me to it?

[Podsville]

OK, I see the dots now. So you can't use one of your own images as a reference, only one which Firefly previously created. I really hope the option of using an uploaded image as reference will be added.

[Byron Smith]

As I jump down the rabbit hole of Holy Adobe power I realize one thing that would definitely increase the generation of content is by being able to retain certain aspects of a piece while enabling other variations. It would be great if you could select the things you want to retain before generating iterations.

[The Scott Valentine]

The general request in this thread is to enable "hinting," or the ability to mark a region to give the engine some direction and constraints. A great way to experience this would be to create a painted or enclosed region, then apply a secondary prompt to describe the changes for that area. Something like "smooth the skin" or "change this grass to mud" or everyone's favorite "give this person the appropriate number of limbs/digits" 😄

[Byron Smith]

This is exactly what I think would make the most sense.

Sent from Outlook for iOS<>

[biagiodistefano]

In "text to image" it would be convenient to be able to select, after sending the prompt and one of the results proposed by Firefly, a specific area of the image on which a modification or addition is requested with a dedicated prompt. This is because many times the requested results are great, but there are some small areas that you don't like or where you would like to try to put other elements.

[Tracy.H]

@biagiodistefano Merging you into this thread so we can track the demand for certain features