New infinite canvas GUI via A1111 API, towards an "AI Photoshop" in browser

[BlinkDL]

Hi guys. I am building a new GUI via A1111 API, towards an "AI Photoshop" in browser.

https://www.painthua.com

It directly connects to your local A1111. Feedback are welcome :)

[mekrod]

wow this is crazy, thanks a lot.

[ghost]

Interesting, will try tomorrow

[CodeExplode]

This is fantastic and is something I've been meaning to get started on myself, pretty much exactly. :)

One feature which might help is a box masking method, where the mask can be snapped to the sizes or aspect ratios of SD images, so that you're able to verify that you're inpainting at the correct aspect ratio if inpainting at full resolution. You could draw out the box and then erase parts you don't want inpainted, but ensure the mask always touches the edges somewhere, or print out the current pixel size or aspect ratio of the masked area, to help adding a bit more masking until you hit an aspect ratio which SD accepts.

[BlinkDL]

Thank you :)

Current workflow: draw mask --> draw a box around masked regions --> send [box contents + mask] for inpainting

So the img2img already always happen within the box

[ghost]

So it uses my local version to render? Thats absolutely awesome! Thank you!

I tried the Krita once but quickly closed it - I didnt have time to learn yet another complex program but yours seems very easy and straightforward

[BlinkDL]

Yes :) So all your A1111 setup works.

[cibernicola]

Can't run --api
To create a public link, set share=Trueinlaunch(). Traceback (most recent call last): File "launch.py", line 249, in <module> start_webui() File "launch.py", line 244, in start_webui webui.webui() File "L:\StableDiffussionWebUI\webui.py", line 147, in webui create_api(app) File "L:\StableDiffussionWebUI\webui.py", line 97, in create_api from modules.api.api import Api File "L:\StableDiffussionWebUI\modules\api\api.py", line 8, in <module> from modules.api.models import * File "L:\StableDiffussionWebUI\modules\api\models.py", line 111, in <module> class TextToImageResponse(BaseModel): File "L:\StableDiffussionWebUI\modules\api\models.py", line 112, in TextToImageResponse images: list[str] = Field(default=None, title="Image", description="The generated image in base64 format.") TypeError: 'type' object is not subscriptable

Am I the only one who experiences this?

[BlinkDL]

Python 3.8 issue, but fixable: #4206

[cibernicola]

Thanks!
Working perfectly, really nice work!

[guumaster]

This is awesome! Are you considering sharing this code also as open source to accept community contributions?

[cibernicola]

I guess this?
https://github.com/BlinkDL/Hua

[BlinkDL]

The frontend will be open-source after I finish all major functions and refactor the code :)

Now it's full of hacks (100% vanilla JS) and very easy to bring in bugs if one don't fully understand it.

Meanwhile I will also write a clean & efficient backend for it and that will be open-source from the beginning.

[smoke2007]

there used to be a lot of security risks with listen and share, any idea on what risk we are putting ourselves in starting with --api in listening mode ? and having it connect to our running instance?

Can we run this code also locally ourselves?

[cibernicola]

As far as I know, the risks came, above all, when making the connection public, that is, generating a link outside the local network. This does not mean that opening the API has no risks, but I think the risks come, above all, from opening it to 100% external access to the local network.

[BlinkDL]

If the A1111 API only processes clean requests (by only allowing sensible parameters) then there is little risk

All requests should definitely be aggressively cleaned.

[cibernicola]

any piece of software on a PC that is turned on and connected to the Internet is at risk. :/
I understand what you are saying and I totally agree, the balance between development and production release (minimizing security related risks as much as possible) is an extremely sensitive and important point.

[Chryseus]

I'm not really comfortable unless it's running locally, the analytics running on it also makes me extra suspicious, I wish I could try it but I'm not going to until the source is available.

[BlinkDL]

Use adblock if you want to be extra safe :)

It's recording generation errors (explicitly: "/#draw/txt2img/512/512" and "/#draw/done/512/512" and "/#draw/err/TypeError:+Failed+to+fetch" - no prompt, no other details) so I can write about them in the guide, and fix bugs if I pushed some problematic code.

On the other hand, there is no way for one to limit information collection in online generation services.

[ghost]

Running offline was one of the first things i tested, and it was working without internet connection, so, just cut your connection and try it out. I did outpaint 4k landscape yesterday, pretty cool stuff. Though, i would prefer to have it run locally as well.

[digitalblossom]

Am I wrong in the impression that the js is heavily obfuscated, including string encryption? Is there a reason for that? Don't take this the wrong way, I'm not accusing you of anything, but I've never seen obfuscation to this extend outside explicitly closed-source paid software and malware.

[BlinkDL]

I am working on https://github.com/BlinkDL/RWKV-LM too, whose training is sponsored by Stability. So please be assured that it is safe :)

The frontend will be open-source after I finish all major functions and refactor the code.

Now it's full of hacks (100% vanilla JS) and very easy to bring in bugs if one don't fully understand it.

Meanwhile I will also write a clean & efficient backend for it and that will be open-source from the beginning.

[Chryseus]

I am working on https://github.com/BlinkDL/RWKV-LM too, whose training is sponsored by Stability. So please be assured that it is safe :)

The frontend will be open-source after I finish all major functions and refactor the code.

Now it's full of hacks (100% vanilla JS) and very easy to bring in bugs if one don't fully understand it.

Meanwhile I will also write a clean & efficient backend for it and that will be open-source from the beginning.

I find it hard to trust that it's safe when the code is heavily obfuscated, it actively tries to prevent it from being run locally, I was able to remove this and can at least say nothing appears malicious at the current time, still I would advise people not to use this until the code is opened up, this all seems highly unnecessary if you intend to open source it.

[HappyCatKitten]

This looks awesome.. but what am I doing wrong? running Linux

[alperinugur]

Hi There BlinkDL,

this is a great API.

Best part is you don't need a connection once you connect at the first. So internet cuts are not affecting!

great job

[NoCrypt]

I hope it supports zoom in/zoom out soon,
Currently zooming out will leave empty space in the canvas, zooming in will shows scrollbars, unless you refresh the page ofc.
The term "infinite" did not fit for current implementation since it's technically limited to your browser viewport, I think it's much more fitting if it is "Figma" kind of canvas. (also hand tool will be very helpful!)

Anyways, good job on the website!