The Mistral provider currently doesn't support image generation. Mistral offers image generation (powered by Black Forest Labs FLUX1.1 [pro] Ultra) through their Agents & Conversations API using the built-in image_generation tool.
AI module defines a TextToImageInterface for providers to implement
Implement TextToImageInterface on MistralProvider.
Mistral's image generation works differently from OpenAI's dedicated /v1/images/generations endpoint — it uses the Agents + Conversations API:
1. Create an Agent with {"type": "image_generation"} in the tools array
2. Start a Conversation with that agent, passing the prompt as input
3. The response contains a tool_file content chunk with a file_id
4. Download the generated image via GET /v1/files/{file_id}/content
The partitech/php-mistral library already provides all the required building blocks (MistralAgentClient, MistralConversationClient, downloadFile()). However, Response::updateFromArray() currently drops tool_file content entries from conversation responses — a small upstream fix is needed to expose the file_id.
So changes:
- Add TextToImageInterface to MistralProvider
- Add 'text_to_image' to getSupportedOperationTypes()
- Add model listing for text_to_image in getConfiguredModels()
- Implement textToImage() using the Agent + Conversation + file download flow
- Return TextToImageOutput with ImageFile containing the downloaded image binary
- Upstream fix to partitech/php-mistral: handle tool_file content type in Response::updateFromArray() (or bump minimum version once fixed)
- Implement TextToImageInterface in MistralProvider
- Determine which Mistral models support the image_generation tool and filter in getConfiguredModels()
- Decide on agent lifecycle strategy (create per request vs. cache agent ID)
- Testing
None. The existing AI API Explorer and any modules using text_to_image operation type will automatically pick up Mistral as an available provider.
MistralProvider will implement TextToImageInterface, adding the textToImage() method.
None
Show commandsStart within a Git clone of the project using the version control instructions.
Or, if you do not have SSH keys set up on git.drupalcode.org:
changes
Comments
Comment #2
petar_basic commentedCreated https://github.com/partITech/php-mistral/issues/82 and provided a fix.
Comment #4
petar_basic commentedAdded TextToImageInterface implementation. This uses the Agents API to create a temporary agent with the image_generation tool, starts a conversation with the prompt, then downloads the generated file from the response references.
Note: this depends on a fix in php-mistral (https://github.com/partITech/php-mistral/pull/83) to handle tool_file content in conversation responses. That PR needs to be merged and released before this can land. There's also an upstream bug where MistralConversation properties aren't initialized with defaults — I'll open a separate issue/PR for that.
Also left to do: add some test(s) for the new capability.
Comment #5
petar_basic commentedCreated: https://github.com/partITech/php-mistral/issues/84
Comment #6
fagoI reviewed the upstream issue and the MR.
Generally, it looks good, but the mistral-provider class is becoming too big. Could we structure this better, and have a separate helper class/service/trait to call out for?
Comment #7
petar_basic commentedextracted the image generation logic into a MistralTextToImageTrait, keeping textToImage() on the provider as a thin wrapper. Also merged with latest 1.1.x and aligned error handling to use handleApiException() for proper typed exceptions.
Comment #8
petar_basic commentedUpstream php-mistral fixes (tool_file parsing + uninitialized properties) are merged and released in 0.1.27. Bumped the requirement accordingly.
Also added one test for a successful generation flow
So this is ready for review