Feature request: BYOK and custom enterprise model gateway support

[kinpoe-ray]

Summary

Please consider adding first-class BYOK (Bring Your Own Key) and custom model gateway support for the AI features in Diffusion Studio.

Use case

Many enterprise users already have:

• Internal image, video, speech, transcription, or multimodal models
• An OpenAI-compatible company gateway
• Separate gateway endpoints for different capabilities or business units
• Company-managed API keys and authentication policies

We would like to use those services directly from Diffusion Studio instead of routing AI generation and media understanding through the hosted Diffusion Studio services.

Proposed configuration

A provider configuration could include:

{
  "id": "company-gateway",
  "baseUrl": "https://ai-gateway.example.com/v1",
  "apiKeyRef": "COMPANY_AI_API_KEY",
  "models": {
    "image": ["company-image-model"],
    "video": ["company-video-model"],
    "audio": ["company-audio-model"],
    "speech": ["company-tts-model"],
    "transcription": ["company-asr-model"]
  }
}

The API key should remain in an environment variable, OS credential store, or local secret manager and should never be written to project files or committed to source control.

Desired capabilities

1. Configure custom providers and endpoints in the UI or a local config file.
2. Use custom providers for generate.image, generate.video, generate.voice, and generate.audio.
3. Use custom providers for transcription and media understanding where compatible.
4. Expose model capabilities such as supported aspect ratios, durations, audio, and input references.
5. Support OpenAI-compatible gateways first, with an adapter interface for non-compatible APIs.
6. Allow users to choose between Diffusion Studio hosted models and their own providers.
7. Preserve the existing local editing, dapi, agent workflow, caching, and export behavior.
8. Provide clear indication of which provider is being used and whether credits are consumed.

Why this matters

The editor and agent workflow are very useful in enterprise environments, but company users may be unable to upload media or prompts to an external hosted service. BYOK/custom gateway support would enable private deployments, internal model evaluation, cost control, compliance, and integration with existing AI infrastructure.

Is BYOK or self-hosted provider support on the roadmap? If not, would you be open to an extension point or documented adapter API?