🔌 OVERTLI OpenAI-Compatible Studio

Somewhere between "cloud by default" and "bring your own GPU" there's a middle path: the API you already pay for. GZ_OpenAICompatibleTextEnhancer is the Overtli Studio Suite's BYO-backend node. Point it at any provider that speaks the OpenAI protocol - Groq, Together AI, LocalAI, a vLLM box, your work's internal gateway - paste a base URL and key, and it drives text enhancement, image generation, video, TTS, STT, and music through that single endpoint. The pack's README lists exactly this roster of engines: text, image_gen, video_gen, text_to_speech_gen, speech_to_text_gen, text_to_music_gen.

How it works

active_engine picks the modality, model is a plain text field you fill with whatever your provider calls the model (default gpt-4.1-mini), and prompt is your input. That's the entire required surface. The node builds an OpenAI-style request against api_base_url with api_key and dispatches to the right engine - for media engines it mirrors the behavior of the pack's dedicated Pollinations nodes, but aimed at your endpoint instead. The IMAGE input feeds vision-capable models, and the AUDIO input feeds the speech-to-text engine. require_api_key defaults to true, which is a deliberate guard: this route genuinely needs one, unlike the free Pollinations tier.

The inputs that matter

• active_engine - which job. This is the field you'll flip most.
• api_base_url - your provider's OpenAI-compatible endpoint. persist_api_settings (default on) saves it plus the key so you only configure once; GZ_ProviderSettings can hold them too.
• model - your provider's model id. If text works but the video engine 404s, the README's caution applies: modality support depends on what the provider and model actually expose, and the node can only validate, not grant, capability.
• text_mode through ttaudio_mode - the same curated preset dropdowns as the rest of the pack, now applied to your backend.
• style_preset_1..3 and additional_styles - style layering, chainable from GZ_StyleStackNode.
• temperature, max_tokens, timeout_seconds - the usual dials.
• media_width/media_height, audio_voice, audio_speed, stt_response_format (srt/vtt included), stt_language - engine-specific knobs that only matter when that engine is active.

Outputs

Four sockets, mirroring the advanced router: text, image, video, audio. One run fills text plus whichever media output your engine produces; wire forward what you care about. A text engine leaves the media sockets empty, and that's fine.

Install and gotchas

Standard pack install - search "Overtli" in ComfyUI Manager, or clone and pip install -r requirements.txt, restart. The README's first troubleshooting entry is about this node: if the OpenAI-compatible route isn't in your node list at all, the pack didn't load cleanly - restart ComfyUI and check the console for import errors before assuming the node is broken.

Where people get burned: wrong base URL format. Some providers want https://api.groq.com/openai/v1 and others just the bare host - if calls 404 or throw auth errors at the path level, normalize the URL to include the /v1 your provider expects. And remember the content-policy reality of any hosted API: the filter belongs to the backend, not the node. There's no local patch for a refusal at your provider's end - if that matters for what you're making, the pack's GZ_LLMTextEnhancer route with your own LM Studio or Ollama is the sibling node designed for exactly that.