Optional image and video contribution

First complete the platform installation and contributor setup. This guide covers the bundled Qwen image and Wan video profiles, not arbitrary ComfyUI workflows or custom nodes.

1. Check memory and select workloads

Images require at least 32 GiB after the contribution cap; videos require 64 GiB after the cap, on every OS. A 32 GiB machine is therefore not image eligible at the maximum 80% cap. Review the memory table in the setup guide.

Run opengpu install to select the eligible workloads. Lowering the cap later can invalidate readiness and require verification again.

2. Choose the media runtime

Windows/macOS native setup is implemented, but release/build tests do not prove that every bundled model works on every GPU. Only successful local generation verification establishes readiness. MLX serves LLMs; Mac ComfyUI uses MPS.

Setup may detect existing ComfyUI and offer it. To supply an endpoint explicitly:

opengpu install --comfyui-url http://127.0.0.1:8188

Use its actual port. MundusX does not take ownership of an external ComfyUI process or stop it. For external ComfyUI verification, pause contribution first and let LLM jobs finish. Before resuming LLM work, ensure the external server has released model memory. Managed media drains LLM work around generation.

3. Review downloads, install, and verify

The wizard can offer these steps interactively. If you deferred them, run:

opengpu media plan
opengpu media setup --yes
opengpu media verify
opengpu media status

For video:

opengpu media --video plan
opengpu media --video setup --yes
opengpu media --video verify
opengpu media --video status

plan displays the profile, download sizes, and budget. Read it before accepting downloads with --yes. Video also requires ffmpeg and ffprobe on PATH. Required free disk space includes missing weights and runtime headroom. Cached weights are verified and reused when valid.

Interactive downloads show bytes/progress bars; redirected output can show JSON events. Verification runs an actual generation and reports elapsed time. A saved configuration or reachable ComfyUI endpoint is not a passing generation test.

4. Optional local generation

opengpu media generate --prompt "A red ceramic teapot on a wooden table"
opengpu media --video --seconds 2 generate --prompt "A red ceramic teapot slowly rotates"

Release 0.2.20 supports every integer from 1 through 10 seconds with the 832×480, 16 FPS profile. Local generation reports its output artifact path. It does not automatically create a public user download URL.

5. Start network contribution

opengpu start --background
opengpu status
opengpu capabilities --probe

The node agent starts the media queue worker for selected image/video workloads. The worker claims only verified, memory-eligible profiles. Network media also depends on control-plane admission; the current admission path requires an admitted LLM contributor. Media-only local generation is not proof of admission. Custom deployments must configure their matching media service; normal public nodes use https://chat.mundusx.ai.

For queued image jobs, the worker uploads the generated image to the hosted media service and completes the job after successful upload. The requesting user receives the hosted result link through the client service. After a verified image upload, MundusX permanently removes its local upload copy and, for managed ComfyUI, the managed output copy. It cannot remove an external ComfyUI server's original through this integration. Video cleanup behavior is different; do not assume image deletion applies to video files. Queued service availability still depends on the corresponding control-plane deployment.

Managed ComfyUI runs on demand. Do not expect it to remain warm and listening on port 8188 while idle. To stop an interrupted MundusX-owned media runtime:

opengpu media stop

This does not stop an externally managed ComfyUI server.