Generate images & video

MIRA can turn a text prompt into an image or a short video clip and show the result inline in your conversation. You don’t run a separate tool or open another app — you just ask.

Like voice, image and video generation are local-first: a small router picks a backend for each request. You can point it at a cloud API (OpenAI Images / Sora) or at a local generator running on your own machine — Stable Diffusion or ComfyUI — that needs no cloud key at all.

Choose a backend

The router resolves a backend per request: an explicit backend you name → the configured default backend → the first enabled one (local is preferred). So once any backend is enabled, image_generate and video_generate become available to the agent.

Configure this under Settings → Image & Video, or directly in the [image] / [video] config blocks (backend selection, endpoints, models — see Settings reference).

Image backends:

Backend	Where it runs	Needs a cloud key?	Notes
OpenAI Images	Cloud (OpenAI or compatible)	Yes	Uses your OpenAI provider key.
AUTOMATIC1111	Local (SD WebUI / Forge)	No	Talks to `/sdapi/v1/txt2img`; start the WebUI with `--api --listen`.
ComfyUI	Local	No	Built-in default SDXL workflow, or your own `workflow_json`.

Video backends:

Backend	Where it runs	Needs a cloud key?	Notes
OpenAI Sora	Cloud	Yes	Async render via the OpenAI Videos API.
ComfyUI	Local	No	Bring your own Wan / AnimateDiff / SVD workflow (`workflow_json`, required).

No cloud key needed for the local backends. Enable AUTOMATIC1111 or ComfyUI in Settings → Image & Video and MIRA generates entirely on your own hardware.

Generate an image

Just ask MIRA in chat:

“Generate an image of a red bicycle leaning against a stone wall at sunset.”

Behind the scenes MIRA calls image_generate, which routes your prompt to whichever image backend is active — OpenAI, a local Stable Diffusion WebUI, or ComfyUI. The image renders inline in the conversation, with download and copy controls so you can save or reuse it.

Describe what you want as specifically as you like — subject, style, lighting, composition. You can also ask MIRA to refine: “make it more minimal” or “same scene but in daylight”. Parameters like the model, size, and a negative prompt flow through to whichever backend is handling the request.

Generate a short video

Video works the same way — describe the clip:

“Make a 4-second clip of waves rolling onto a beach.”

MIRA calls video_generate, routed to OpenAI Sora or a local ComfyUI video workflow. Video rendering takes a while, so this runs asynchronously: MIRA enqueues the job, polls until it’s done, then returns the finished MP4 inline as a video player with playback controls and download.

You can steer a few parameters by asking for them — the model, the size (aspect ratio / resolution), and the length in seconds. Keep clips short; generation time and cost grow with length.

ComfyUI custom workflows

ComfyUI is driven by a workflow graph, so it’s the most flexible backend.

Image — MIRA ships a built-in default SDXL txt2img workflow, so ComfyUI works out of the box once the server is reachable. To use your own graph instead, paste an API-format workflow into image.comfyui.workflow_json.
Video — there’s no universal default (video graphs vary widely: Wan, AnimateDiff, SVD), so video.comfyui.workflow_json is required — the backend stays off until you supply one.

MIRA substitutes tokens into your workflow before submitting it, so the same graph adapts to each request. Put these tokens where the values belong:

{{prompt}} · {{negative}} · {{seed}} · {{width}} · {{height}} · {{steps}} · {{cfg}} · {{ckpt}} — and for video, {{frames}} (= seconds × fps) and {{fps}}.

Export the workflow from ComfyUI in API format, drop the tokens into the nodes you want MIRA to control, and paste it into the matching workflow_json setting.

Other generators

AUTOMATIC1111 and ComfyUI are first-class native backends — you no longer need an MCP workaround for Stable Diffusion or local models. Anything beyond these backends (a provider MIRA doesn’t speak natively) can still come in through an MCP server: MCP tools that return images, audio, or video also render inline in chat.