Image generation keeps failing, and admin UI for it is unclear/confusing

[hybridherbst]

Which version of assistant are you using?

2.11.0

Which version of Nextcloud are you using?

32.0.2

Which browser are you using? In case you are using the phone App, specify the Android or iOS version and device please.

143.0.7499.41

Describe the Bug

Image generation via LocalAI with an external provider keeps failing, and I can't figure out why. I believe one issue might be that the options in the Assistant admin settings are confusing, and I'm not even sure if image generation is now on or off – or if there is a bug in the UI (which I suspect).

"Unified task processing" lists Generate Image as enabled:

The "Image Generation" section pretends there's no provider (despite LocalAI being installed and configured):

The "OpenAI and LocalAI integration" lists image generation as enabled and configured:

So which one is true, and why is "Image Generation" shown as not supported?

Text generation / chat works fine, so the model provider (OpenRouter, with its OpenAI-compatible API) is set up correctly.

Expected Behavior

Either the image generation is not configured correctly – then the Assistant should not show "Image Generation" at all.

Or it is configured correctly – then there should be clear notes on whether things are set up correctly or what the actual error is. The Admin UI should then properly show me if it's configured correctly.

It would also be great if there was a testing section in the setup – right now it's very easy to configure wrong models (e.g. set a text model for image generation) and the way to testing is over the Assistant UI and not in the admin settings.

To Reproduce

• Set up LocalAI with OpenAI
• Install Assistant
• Make sure to choose both a text and image model (that have these capabilities)
• Note that the Admin UI will still show "Image Generation: no provider"
• Note that image generation won't work even when model is configured correctly

[hybridherbst]

Potentially related: #175
Neither of the options – "Use authentication for image retrieval request" on/off – gives an image.
The logs contain the following:

A TaskProcessing core:text2image task with id 19 failed with the following message: 
OpenAI/LocalAI's text to image generation failed with: 
API request error: Client error: 
`POST https://openrouter.ai/api/v1/images/generations` resulted in a `405 Method Not Allowed` response

[janepie]

Thanks for raising this!

The Image generation part telling you there is no provider is from an old API which was deprecated in NC30, we should maybe check if we can get rid of that in a nice way or at least hide it, good point.

Unfortunately there is no good way to check which model is an image-providing model and which is a text-providing model afaik as that is not part of the OpenAI API. The only way to check would be to run a prompt and see if it produces an image, which is a bit costly.

[janepie]

Ah, I just saw that your tasks are also failing. Can you provide your Nextcloud and browser logs?

[hybridherbst]

Sure, thanks –
from Nextcloud perspective it looks like this:

Browser console logs:

Uncaught (in promise) Error: User cancellation
    at HTMLDivElement.<anonymous> (assistant.js:112:11)
    at Proxy.onCancel (AssistantTextProcessingModal.vue:47:20)
    at Hn (runtime-core.esm-bundler.js:199:19)
    at Lt (runtime-core.esm-bundler.js:206:17)
    at El (runtime-core.esm-bundler.js:4378:5)
    at h (vue-select.es.js:268:204)
    at Hn (runtime-core.esm-bundler.js:199:19)
    at Lt (runtime-core.esm-bundler.js:206:17)
    at HTMLButtonElement.e (runtime-dom.esm-bundler.js:730:5)Understand this error
runtime-dom.esm-bundler.js:621 Error: <linearGradient> attribute gradientTransform: Expected '(', "rotateX(285)".
bo @ runtime-dom.esm-bundler.js:621Understand this error
8globals.js:36 jQuery is deprecated: The global jQuery is deprecated. It will be removed in a later versions without another warning. Please ship your own.
Rn @ globals.js:36Understand this warning
assistant.js:143 [assistant] Task failed 
{
    "id": 23,
    "type": "core:text2image",
    "lastUpdated": 1765806922,
    "status": "STATUS_FAILED",
    "userId": "admin",
    "appId": "assistant",
    "input": {
        "numberOfImages": 1,
        "model": "google/gemini-2.5-flash-image",
        "input": "landscape with trees"
    },
    "output": null,
    "customId": "",
    "completionExpectedAt": 1765807002,
    "progress": 0,
    "scheduledAt": 1765806921,
    "startedAt": 1765806922,
    "endedAt": 1765806922,
    "allowCleanup": true
}

Nextcloud logs:

---

EDIT: I found one discrepancy – looks like OpenRouter uses the OpenAI-compatible image generation endpoint for chat: https://openrouter.ai/docs/guides/overview/multimodal/image-generation
And not the (older) one dedicated to image generation (https://platform.openai.com/docs/api-reference/images/create).

Since most interfaces are moving to this kind of multimodal generation as the "better way" (asking about and generation of images right in the chat), will Nextcloud Assistant do the same? Or at least allow choosing the chat completions endpoint also for image generation tasks?