The prompt input is only as wide as the input image by default, for looks.

If the prompt goes beyond a certain length, the input should grow to show more of the input text.

Each time you inpaint the quality of the whole image drops
this is an example, before and after inpainting 4 times (to make the effect clear)

edit: i was only masking a small area on the right where the frame changed

As of last night, the Replicate API no longer returns data URLs from prediction input. This makes it a bit challenging to see the mask that was supplied with the request.

One alternative would be to upload the data URL to GCS/S3/something and then use that HTTPS url in the prediction request.

Simply put, all images are stretched/squished so that they are made to be square.

e.g.:

• https://unsplash.com/photos/sR7_ImYvt1Q

...looks like this:

Hi,

I was just trying to see what model you use - it looks like you have a link that points to stable diffusion, but I was a bit confused since stable diffusion generally does text-to-image without a mask. Is it a specific variant of sd/sdxl that does inpainting like this one?

Regards,
Surya

you can currently start the design process with a prompt, but it would be cool if you could also supply a starting image by dragging and dropping it onto the page.

We might want to do something fancy with the supplied image to make the user-supplied image fit the 512x512 constraint. 🤔

The UI should make it clear to the user when a result fails due to a NSFW response, or maybe even run the same query again until a successful response is returned.

The Stability AI team is slowing working on improving NSFW detection, but the current version still flags a bunch of false positives.

Sorry to ask this, but can someone tell me exactly where to place my API token?

Thanks

Error related to *.replicate.delivery image host pattern - the images are not returned to the user from the API, when a simple following of the instructions is undertaken. Upgrading to the next minor release of next solves it.

I also noticed an error in the log output that mentions an unknown property clasName - I'll submit a PR to fix all these!

I get an "Invalid token." error when I try and use the app after I deploy it online?

Hi,

Thank you for your good work, is it possible for me to use painting API on my own code, not on the web page?

Best regards

Observe:

It's defaulting to 0.8

Maybe start by trying something weaker like 0.5

cc @rossjillian

https://github.com/vinothpandian/react-sketch-canvas#methods

The current "eraser" draws a black blob on the prediction image:

It would be nice to make the erasing reveal the classic grey-and-white checkboard, à la Photoshop and DallE 2.

The current implementation embeds the prediction output image as a CSS background, making it difficult to download the image.

Let's add a download button.

so you can more easily see the diff when it inpaints