Expose a webservice

[OleJ1964]

Would it be possible to expose the SD as a webservice?

[outlookhazy]

I spent some time reverse engineering the gradio api that the UI is using - it appears to just be RPC over websockets. Biggest issue with that was that it appears likely auto-generated (referring to functions by number) so it wouldn't be stable as the source changes.

[OleJ1964]

Thanks for reply :-) I understand - but still better than spredning implementations all-over the place - id really like to be able to implement SD in various applications/projects (languages/platforms) - and a webservice just seem right (at the moment anyway - who knows where we are in 1 month)

[outlookhazy]

I agree entirely.
I would just build my own server that calls the python scripts as an external process, but it seems like a good portion of time is spent loading models, so I would guess that doing it that way is going to be much slower than having a python server managing what's loaded at any given time. Unfortunately, I don't have any background in Python, so I would have to start from the ground up.

[endomorphosis]

I added a pull request, which only modifies 3 lines, which exposes an API through gradio.

[outlookhazy]

Thanks for this! It seems the API is not well documented at all..
I'm having trouble finding any substantial information.
I found a post that suggests adding that api_name tag adds api endpoints in the form <host>:<port>/api/<api_name>, but I haven't been able to find exactly what that endpoint is (http? websocket?) what it's expecting (json? content?) or what return data would look like.. lol

Edit: haven't played with it yet, but the gradio tests seem to indicate the api endpoint is http. It's expecting a POST containing a json object in the format {"data":["param1","param2"]} and returns in the same format

[endomorphosis]

my pull request has been accepted, and you are correct, that is the format that it expects, which is the same format as is being sent over websockets

[OleJ1964]

Exactly - that was my initial thought. but the time increase I feel would not merrit it. If I knew anything about python I'd look into it myself. But i plan looking into python soon just bc. SD. (I really don't like the language or environment) Den fre. 9. sep. 2022 kl. 18.13 skrev Charlie ***@***.***>:

…

I agree entirely. I *would* just build my own server that calls the python scripts as an external process, *but* it seems like a good portion of time is spent loading models, so I would guess that doing it that way is going to be much slower than having a python server managing what's loaded at any given time. Unfortunately, I don't have any background in Python, so I would have to start from the ground up. — Reply to this email directly, view it on GitHub <#875 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A24WW6EBT7YY3ECRWLWAWFTV5NO27ANCNFSM6AAAAAAQIL4IQI> . You are receiving this because you authored the thread.Message ID: <sd-webui/stable-diffusion-webui/repo-discussions/875/comments/3609916@ github.com>

[OleJ1964]

Sounds absolutely fantastic.

…

On Mon, 12 Sep 2022, 12:18 endomorphosis, ***@***.***> wrote: I added a pull request, which only modifies 3 lines, which exposes an API through gradio. — Reply to this email directly, view it on GitHub <#875 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A24WW6GQFLJNA3ZHONPT2XLV537RFANCNFSM6AAAAAAQIL4IQI> . You are receiving this because you authored the thread.Message ID: <sd-webui/stable-diffusion-webui/repo-discussions/875/comments/3624561@ github.com>

[Arukaito]

@endomorphosis can you please share an example ?
Cant get it to work, response is always {"error":"'NoneType' object is not callable"}
im following the view api page of Gradio

[endomorphosis]

i think that the package maintainers have broken it, it changed so frequently and we go so annoyed with the moving target, that me and my partner just reimplemented our own api solution.

a previous pull request that I had, does in fact work appropriately, but there are multiple stages of calling the api to add a job, then checking on that queue until the response is ready, then finally reading the response.