Multimedia

Orb produces two things besides text: an image of the current scene, and spoken audio of a character's dialogue. Both are per reply, and both are off until you set up a backend.

Images

Orb can turn an assistant reply into a picture of that scene. You request every image on demand — nothing renders on its own.

Orb's Agent-lane LLM reads the conversation through the reply you picked and writes the diffusion prompt. That prompt is all the image model receives; the conversation, the character card, and the scene analysis stay behind.

Backend	What it needs	What it costs	Set it up
External ComfyUI	A ComfyUI server you run, a checkpoint, and an imported API-format workflow	Your own hardware and electricity	ComfyUI Setup
Cloud API	An API key from a supported provider	Money, per image, billed by the provider	Cloud Image Setup

Orb does not install ComfyUI or image models.

The choice is per style, not global: each style names the connection it renders on, so a local anime checkpoint and a commercial photorealistic API can sit one dropdown apart in the same conversation. A style keeps both backends' settings, so relinking it and relinking it back loses nothing.

Image Generation covers styles, the camera, reference images, variants, and manual prompt editing, and ends with a symptom table for renders that fail.

How a render happens

1. You select Visualize reply on an assistant reply.
2. Orb settles the camera (POV) for that reply — from the picker, or from the local classifier in Auto mode.
3. The Agent model writes the scene prompt, after an optional complex-scene analysis pass.
4. Orb assembles the final positive and negative prompts from the style, the visible character profile, and the composed scene.
5. The style's connection renders it: your ComfyUI server, or the provider's API.
6. The image attaches to the reply as a variant. Render details records the style, seed, prompts, and which lever chose the camera.

Steps 3 and 4 run again on Regenerate. Reroll replays the stored prompt with a new seed and skips them.

Speech

Orb speaks a character's dialogue, not the whole reply: a local extractor pulls the quoted lines out first and leaves narration and inner monologue unspoken. Select the speaker icon on a message to hear it, select a single quoted line to hear just that line, or turn on auto-speak for every new reply.

Microsoft Edge TTS ships in requirements.txt and needs no key, so speech works out of the box. Kokoro-82M and Fish Speech run locally; OpenAI-compatible endpoints and ElevenLabs are cloud services. Each character carries its own voice, backend, speed, and pitch in the character editor's Voice tab, and the global toggles live in Settings.

Text-to-Speech covers each backend, the extraction pipeline, and how to add a backend of your own.

Start here

• ComfyUI Setup

  You have a GPU. Install ComfyUI, grab a checkpoint, and import a working workflow — no cost per image.

• Cloud Image Setup

  You don't. Paste an API key and render — no GPU, no model download, billed per image.

• Image Generation

  Backend connected. Configure styles, the camera, reference images, and prompt editing, then make your first image.

• Reference Image Setup

  ComfyUI running. Add an edit workflow so a render can start from a picture and keep the character's face.

• Text-to-Speech

  Give a character a voice, pick a backend, and choose what gets spoken.

Warning

Cloud providers are third-party commercial APIs, and a ComfyUI server that is not on this machine is a remote server. Your scene prompts, your reference images if you turn them on, and the dialogue sent to a cloud voice all leave this machine. Orb asks you to confirm before sending images or prompts to a third party.