Panel media generation

Generate images and video, synthesize speech, transcribe recordings, and produce music or sound effects with locally installed, currently registered media executors. The Engine must be running and must report both the model and its requested task. A model file alone is not an execution capability.

Open Panel Instances to create and start a media instance, then Panel Builds and Engines to configure its engine. Open the Images, Video, Speech, Transcribe, or Audio links on the native media pages. Media guided tasks opens /rust/media. The classic controls remain usable on an instance configured for an external image or audio executor.

Guided task: Generate an image with a local model

1. Open /rust/media/guides/image as an administrator. Confirm the Running instance ID is the intended live instance and choose a Bound model ID from the live inventory.
2. Enter Image description and optional Width, Height, Steps, and Seed (optional) for this one request. The guide sends TextToImage to the native submission route; the interface refuses a model without a registered image executor.
3. Choose Submit guided job. A queued receipt includes the job ID; queued does not mean that an image exists.
4. Use Refresh job status to observe authoritative progress, or Cancel this job (undo) for a queued or running job. Tick Confirm job cancellation before selecting Request cancellation. The page rejects cancellation without this check. Cancellation requests do not certify immediate termination.
5. Choose Open output gallery and receipts, then inspect the image, inference receipt, and model-license notice in Images. A completed engine job only becomes a retained output after the artifact has passed the Panel's checks.

This guide is also registered as media.image-local. Its checks include the live model inventory and the returned task identity. Its undo calls the registered cancellation route using the recorded job ID. The native Images form is also available directly at /rust/lexi-engine/images with Model, Task, Prompt / speech text, and task-specific settings.

Guided task: Transcribe audio and speak a reply

1. Open /rust/media/guides/speech as an administrator. Identify the live instance and a transcription-capable model.
2. Put a regular WAV file into that instance's media-inputs directory; enter its absolute path as Input WAV absolute path. Choose Submit guided job with the Transcribe task.
3. Choose Refresh job status, then inspect the retained transcript on the native Transcribe page. Review its content before composing a reply.
4. In Speak a reviewed reply, enter Running instance ID, Bound speech model ID, and Reply text. Optionally set Voice ID (optional) and Language (optional); select Submit spoken reply.
5. Refresh the speech job and choose Open spoken reply gallery to play the resulting WAV. Cancel a pending job through Cancel this job (undo) after ticking Confirm job cancellation when necessary.

This guide is also registered as media.transcribe-and-reply. A transcription result is not silently turned into spoken text: the operator reviews the transcript and explicitly supplies the reply. Both steps require independent live model-task support.

Work directly on the five native pages

Every form is bound to an explicit local instance. Model and Task are selected only from live capabilities; an unavailable executor is never substituted with a different one. Files used as image, mask, or audio inputs must have approved types and absolute paths within the selected instance's media-inputs directory. Audio inputs are WAV, inpainting masks are PNG. An empty image or audio library does not imply that conversion is available.

For image or video jobs, Width (64-4096) and Height (64-4096) are subject to stricter executor limits, Steps (1-250) selects iterations, Seed (blank for random) gives reproducibility, and Guidance supports values from zero to 100. Edit strength (0-1) is for image editing. Video also has Frames (1-512), FPS (1-120) and Video output (y4m or apng). The native video path refuses durations over 120 seconds, frame dimensions above 2048 or not divisible by 32, and outputs exceeding its 32 MiB retention allowance. Recorded still images and audio also have file-size limits.

The Model license and Engine inference receipt (not qualification) appear beside retained jobs. Unreported commercial rights are not permission to use weights commercially. Refresh the page for progress; Cancel job only requests cancellation and does not remove stored results. Use Lexify this workload (measured optimization) to inspect performance optimization separately from media generation.

Configure models and devices

Open /rust/lexi-engine and set Models directory (absolute path), Model IDs, separated by commas, Multimodal VRAM budget (MiB), Multimodal device IDs (comma-separated), and Precision (auto, f16, bf16, f32). These fields map to --multimodal-models, --multimodal-model, --multimodal-vram-budget-mib, --multimodal-devices, and --multimodal-precision respectively. The interface disables unsupported engine flags. Save multimodal settings changes the profile, not the running process: restart through the normal instance controls to apply. Use existing model directory imports only a previously configured model directory path; it does not move files or transfer usage rights.

The classic image generation and audio task panels use their existing instance-bound API paths, separate from the native media profile. Pick the intended engine and inspect its available tasks; a result from one executor must not be treated as another executor's output.

Routes used by the media screens and guides

The guided API accepts string form values for instance, model, task, prompt, input path and optional settings. A submission responds with id, task and status=queued only when the existing adapter records real engine admission. The status response includes status, progress, error, artifacts and the original receipt when available. Unknown job IDs and unsupported tasks are refused; unauthorized users cannot run privileged guide actions.

Control reference

Each help marker (?) shows the following exact registry tooltip. Classic-only controls receive their help through their classic screen selector.

See Panel Models for model file management, Panel Instances for starting and restarting instances, and Panel Diagnostics for execution failures. All media results depend on the running executor and the actual model's license and resource requirements.